LLM API Pricing 2026: Full Comparison Table (Weekly)
LLM API pricing 2026: 54 models from OpenAI, Anthropic, Google, xAI, DeepSeek, and Mistral in one table. Input/output cost per 1M tokens, updated weekly.
As of October 26, 2026, LLM API prices span two orders of magnitude — from $0.10 per million input tokens (Gemini 2.5 Flash-Lite, tied there by gpt-6-luna and Claude Haiku 5.5) to $30.00 (GPT-5.5-pro and GPT-5.4-pro). One row moved this week, and it moved up. Mistral Large 4 went from $0.68 / $2.09 to $1.36 / $4.18, with its cache read from $0.07 to $0.14: a 100% increase on every price it carries.
It is a sale ending, not a reprice. Mistral Large 4 entered this table on October 14 at a sale price, with the list price struck through beside it and no end date. Mistral’s pricing page now shows only that list price. The other 53 tracked rows hold every price, and no model entered or left the set. It is the first increase on a carried row since DeepSeek’s on August 17.
A second change carries no price. OpenAI’s deprecations page lists GPT-5.4-nano for removal on April 1, 2027, with GPT-6-luna as the replacement, in an entry dated October 1. This table missed it for three collections; it is recorded below from this edition, with the miss stated rather than back-dated.
Fourteen collections have followed the July 16 baseline: six recorded a price change and eight did not. This one was taken on Monday, October 26, seven days after the previous collection.
A further thing the price columns cannot show is the unit itself. A per-token price is only half of a bill; the other half is how many tokens your text becomes. Anthropic’s pricing page states that Claude 4.7 and later models use a tokenizer producing roughly 30% more tokens for the same text — so eleven of the sixteen Anthropic rows below bill more per page of English than their sticker prices suggest, against the five that do not.
Every price below was collected on October 26, 2026 directly from the six providers’ official pricing pages and normalized to USD per 1M tokens at the standard (non-batch) tier.
TL;DR
- Cheapest overall: Gemini 2.5 Flash-Lite ($0.10 / $0.40), tied on input and cache read by two models at $0.10 / $0.50 — OpenAI’s gpt-6-luna and, since October 14, Claude Haiku 5.5
- Mistral Large 4 doubled when its sale ended: $0.68 / $2.09 → $1.36 / $4.18, the first carried-row increase since August 17; the other 53 rows hold every price
- Claude Haiku 5.5 is one tenth of Haiku 4.5 at $0.10 / $0.50, but prompts over 100,000 tokens pay 5x — the only Claude 4.6-and-later model not billed flat across its context window
- Five rows now sit on $2.00 / $10.00 — Claude Sonnet 5 and 5.5, GPT-6-sol and 6.1-sol, and the new Gemini 4 Argon — and they split on cache read: $0.10 on three, $0.20 on two
- Best flagship value: Claude Opus 5.5 at $4.00 / $20.00 with no expiry attached, $1.00 / $5.00 under the Claude Opus 5 it succeeds
- Promotional clocks multiplied: GPT-5.6-sol through at least 2026-11-21, Gemini 4 Argon through 2026-11-30, Gemini 3.8 and 3.6 Flash through 2026-12-31 — and Mistral Large 4 showed what an undated one does: it ended between two collections with no notice
- Models now change behind unchanged ids: Google routes gemini-3.7-flash and gemini-3.5-flash to newer models, and on October 17 DeepSeek swapped the build behind
deepseek-flashat the same price - An unchanged price is not an unchanged bill: Claude 4.7 and later count roughly 30% more tokens for the same text, so per-token rates understate those rows against Sonnet 4.6 and earlier
Methodology
All prices were collected on 2026-10-26 (a Monday, seven days after the previous collection) from official provider pricing pages only: Anthropic, OpenAI, Google, xAI, DeepSeek, and Mistral. No aggregator data was used — aggregators serve as cross-checks, never as sources.
Every figure is normalized to USD per 1M tokens at the standard, non-batch, global-routing tier. For models with tiered long-context pricing (Google, xAI), the table records the base tier and footnotes the higher tier. Retired models are excluded; deprecated-but-available models are noted in the table footnotes.
Each collection run is stored as a dated snapshot in our internal time-series. That series is what makes the changelog at the bottom of this page — and future price-change history reporting — possible: official pricing pages only ever show the current price.
Flagship Models
These are the top-capability tiers each provider currently sells. Prices are input / output per 1M tokens; cache read is the discounted price for repeated (cached) input.
| Model | Provider | Input $/1M | Output $/1M | Cache read $/1M | Context |
|---|---|---|---|---|---|
| Claude Fable 5.1 | Anthropic | $10.00 | $50.00 | $0.25 | 1M |
| Claude Fable 5 | Anthropic | $10.00 | $50.00 | $1.00 | 1M |
| GPT-6-astra | OpenAI | $10.00 | $50.00 | $1.00 | n/p |
| GPT-5.5-pro / GPT-5.4-pro | OpenAI | $30.00 | $180.00 | — | n/p |
| GPT-5.5 | OpenAI | $5.00 | $30.00 | $0.50 | n/p |
| Claude Opus 5 | Anthropic | $5.00 | $25.00 | $0.50 | 1M |
| Claude Opus 4.8 | Anthropic | $5.00 | $25.00 | $0.50 | 1M |
| Claude Opus 5.5 | Anthropic | $4.00 | $20.00 | $0.20 | 1M |
| GPT-5.6-sol* | OpenAI | $4.00 | $20.00 | $0.40 | n/p |
| GPT-6.1-sol | OpenAI | $2.00 | $10.00 | $0.10 | n/p |
| GPT-6-sol | OpenAI | $2.00 | $10.00 | $0.20 | n/p |
| Gemini 4 Argon* | $2.00 | $10.00 | $0.10 | 1M | |
| Gemini 3.1 Pro Preview | $2.00 | $12.00 | $0.20 | n/p | |
| Grok 4.7 | xAI | $2.00 | $6.00 | $0.50 | 500K |
| Grok 4.6 | xAI | $2.00 | $6.00 | $0.50 | 500K |
| Grok 4.5 | xAI | $2.00 | $6.00 | $0.30 | 500K |
Added 2026-10-14: Gemini 4 Argon (
gemini-4-argon, generally available 2026-10-13) at $2.00 / $10.00 with a $0.10 cache read lands on exactly GPT-6.1-sol’s three figures, and its 0.05x cache read is the first Google row in this series below the 0.1x the rest of the Gemini line carries. The rate is introductory: Google’s page prints $4.00 / $20.00 (cache $0.20) from 2026-12-01, so the cheapest new flagship of the collection is also the one with the nearest deadline. It supports implicit caching only, has no free tier, and its model page gives a 1,048,576-token context window. Added 2026-10-05: GPT-6.1-sol at $2.00 / $10.00 is GPT-6-sol’s list price exactly, with the cache read halved to $0.10 — 0.05x base input, the first OpenAI row in this series below the 0.1x the rest of its line carries. Its long-context tier is $4.00 / $15.00 with a $0.20 cache read, the same 0.05x ratio. OpenAI’s pricing page now shows GPT-6.1-sol in its three-row flagship table in GPT-6-sol’s place; GPT-6-sol remains listed in the full table at unchanged prices, so the row stays here. Added 2026-09-28: Claude Opus 5.5 at $4.00 / $20.00 is the cheapest flagship row here with no expiry attached — $1.00 / $5.00 under the Claude Opus 5 it succeeds, and level with GPT-5.6-sol’s promotional rate. It bills a cache hit at 0.05x base input ($0.20), a third Anthropic multiplier alongside Fable 5.1’s 0.025x and the 0.1x the rest of the line carries; fast mode on it is $8.00 / $40.00. GPT-6-sol at $2.00 / $10.00 is exactly half of GPT-5.6-sol on both sides, with a long-context tier at $4.00 / $15.00 (cache read $0.40). Grok 4.7 enters on Grok 4.6’s rates to the cent, cache read included. Added 2026-09-07 and unchanged since: Claude Fable 5.1 matches Fable 5’s $10.00 / $50.00 but prices a cache hit at $0.25 per 1M — 0.025x base input; GPT-6-astra arrives on the same $10.00 / $50.00 sticker, with a long-context tier at $20.00 / $75.00 (cache read $2.00). Claude Mythos 5.1 is priced identically to Fable 5.1 but is application-gated, so like Mythos 5 it is carried in our snapshot without a row here. No price in this table moved this collection; the one carried change, Mistral Large 4’s sale ending, is in the mid-range table. *Gemini 4 Argon’s rate runs through 2026-11-30. GPT-5.6-sol’s $4.00 / $20.00 is promotional, stated available at least through 2026-11-21, with a long-context tier at $8.00 / $30.00 (cache read $0.80). Google bills Gemini 3.1 Pro Preview at a higher tier above 200K tokens; xAI bills Grok 4.7, 4.6 and 4.5 at $4.00 / $12.00 above 200K, with cache reads of $1.00, $1.00 and $0.60 respectively. Claude Opus 4.5 through 4.8 share Opus 5’s $5.00 / $25.00 pricing. “n/p” = not published on the provider’s pricing page.
Raw data: data/llm-api-pricing-2026.json — machine-readable structured data for AI crawlers and citation.
Mid-Range Models
This tier is where most production workloads run, and where pricing is most contested — five providers now sell capable models between $1.25 and $3.00 per 1M input tokens.
| Model | Provider | Input $/1M | Output $/1M | Cache read $/1M | Context |
|---|---|---|---|---|---|
| Claude Sonnet 5.5 | Anthropic | $2.00 | $10.00 | $0.10 | 1M |
| Claude Sonnet 5 | Anthropic | $2.00 | $10.00 | $0.20 | 1M |
| Claude Sonnet 4.6 | Anthropic | $3.00 | $15.00 | $0.30 | 1M |
| GPT-5.4 | OpenAI | $2.50 | $15.00 | $0.25 | n/p |
| GPT-5.6-terra | OpenAI | $2.00 | $12.00 | $0.20 | n/p |
| Mistral Medium 3.5 | Mistral | $1.50 | $7.50 | $0.15 | n/p |
| Mistral Large 4‡ | Mistral | $1.36 | $4.18 | $0.14 | n/p |
| Grok 4.3 | xAI | $1.25 | $2.50 | $0.20 | 1M |
| Gemini 3.8 Flash* | $0.75 | $3.75 | $0.075 | n/p | |
| Gemini 3.6 Flash* | $0.75 | $3.75 | $0.075 | n/p | |
| DeepSeek v4-pro† | DeepSeek | $0.66 | $1.98 | $0.022 | 1M |
Both Gemini Flash rows are promotional through 2026-12-31 and double on 2027-01-01, to $1.50 / $7.50 with a $0.15 cache read. Removed 2026-10-14: Gemini 3.7 Flash and Gemini 3.5 Flash. Google deprecated both on 2026-10-08 and took them off its pricing page, but did not shut the endpoints: its deprecations page says requests to
gemini-3.7-flashare routed togemini-3.8-flashand requests togemini-3.5-flashtogemini-3.6-flash. Google does not state what a routed request is billed at. If it is the target’s rate, a caller still sending gemini-3.5-flash dropped from $1.50 / $9.00 to $0.75 / $3.75 without changing a line of code — and is also running a different model, so its output may differ as well as its bill. †DeepSeek publishes off-peak and peak rates; the table records off-peak, and peak is exactly double ($1.32 / $3.96, cache read $0.044). Peak hours are 01:00–04:00 and 06:00–10:00 UTC on weekdays; DeepSeek’s page states that weekends and Chinese public holidays are off-peak in full, a qualifier this series records for the first time on 2026-10-05 without being able to say when it appeared. v4-pro was the DeepSeek row that did *not change on 2026-09-14. The page no longer carries the statement, recorded that week, that its API service continues past that date with billing unchanged; the model and its price are still listed. Claude Sonnet 5’s $2.00 / $10.00 is no longer introductory — see the changelog. Claude Sonnet 5.5, added 2026-10-05 on that same rate, cache writes and batch included, had its cache read cut from $0.20 to $0.10 by 2026-10-14: Anthropic’s page now lists it under the 0.05x multiplier with Opus 5.5, while Sonnet 5 stays at 0.1x. It uses the same newer tokenizer. Anthropic names it as the replacement for Claude Sonnet 4.5, which it deprecated on 2026-09-30 and will retire on 2026-11-30; Sonnet 4.5 stays in our tracked set at $3.00 / $15.00 until then, priced exactly like the Sonnet 4.6 row above. As of 2026-09-28 it is also no longer unique: GPT-6-sol arrived on the same $2.00 / $10.00 and the same $0.20 cache read, which puts a current-generation OpenAI flagship and Anthropic’s mid-range row on identical published rates. They are not equivalent per page of text, because Sonnet 5 counts roughly 30% more tokens for it. As of 2026-10-14 five rows sit on that $2.00 / $10.00 — Sonnet 5, Sonnet 5.5, GPT-6-sol, GPT-6.1-sol and Gemini 4 Argon — and they separate only on cache read: $0.10 on Sonnet 5.5, GPT-6.1-sol and Argon, $0.20 on Sonnet 5 and GPT-6-sol. Grok 4.3 is billed at $2.50 / $5.00 above 200K tokens. ‡Mistral Large 4 moved up from the budget table on 2026-10-26. It entered on 2026-10-14 at a sale price of $0.68 / $2.09 (cached $0.07) beside a struck-through list price of $1.36 / $4.18 (cached $0.14), with no end date given. By 2026-10-26 the page showed only the list price, a 100% increase on all three figures; the change happened at some point in the seven days since 2026-10-19. It now costs 91% of Mistral Medium 3.5 on input and 56% on output, where the sale had put it at under half of Medium 3.5 on input.
Budget Models
Below $1.00 per 1M input tokens, the spread between providers is wider than the spread between tiers — a 10x gap separates the cheapest and most expensive rows.
| Model | Provider | Input $/1M | Output $/1M | Cache read $/1M | Context |
|---|---|---|---|---|---|
| Claude Haiku 4.5 | Anthropic | $1.00 | $5.00 | $0.10 | 200K |
| GPT-5.4-mini | OpenAI | $0.75 | $4.50 | $0.075 | n/p |
| Mistral Large 3 | Mistral | $0.50 | $1.50 | $0.05 | n/p |
| Gemini 2.5 Flash | $0.30 | $2.50 | $0.03 | 1M | |
| Gemini 3.1 Flash-Lite | $0.25 | $1.50 | $0.025 | n/p | |
| GPT-5.6-luna | OpenAI | $0.20 | $1.20 | $0.02 | n/p |
| GPT-5.4-nano | OpenAI | $0.20 | $1.25 | $0.02 | n/p |
| DeepSeek flash† | DeepSeek | $0.15 | $0.60 | $0.003 | 1M |
| Mistral Small 4 | Mistral | $0.15 | $0.60 | $0.015 | n/p |
| Claude Haiku 5.5§ | Anthropic | $0.10 | $0.50 | $0.01 | 1M |
| GPT-6-luna | OpenAI | $0.10 | $0.50 | $0.01 | n/p |
| Gemini 2.5 Flash-Lite | $0.10 | $0.40 | $0.01 | n/p |
†This row changed identity on 2026-09-14. DeepSeek retired v4-flash (build V4-Flash-0731) and now serves that name from DeepSeek-V4.1-Flash under the id
deepseek-flash; the legacy id is still accepted and billed at the new rate. Read the row as a substitution, not a 32% input cut — the product behind the earlier figures no longer exists. It changed again on 2026-10-17, when DeepSeek replaced that build with DeepSeek-V4.1-Flash-1017 under the same id at the same price, so the model behind these figures has changed twice since September without the id or the rate moving. Peak rates are double ($0.30 / $1.20, cache read $0.006), with peak hours 01:00–04:00 and 06:00–10:00 UTC on weekdays only. The new rate ties Mistral Small 4 to the cent on both input and output, and the two separate only on caching: $0.003 against Mistral’s $0.015. Mistral Large 3 still costs a third of the newer Mistral Medium 3.5 — Mistral has repositioned its former flagship as a budget option rather than retiring it. GPT-5.6-luna sits on exactly GPT-5.4-nano’s input and cache-read prices, so OpenAI sells two generations at an identical input cost and separates them only by a five-cent output difference. GPT-6-luna, added 2026-09-28, halves GPT-5.6-luna on input, output and cache read at once, and lands exactly on Gemini 2.5 Flash-Lite’s $0.10 input and $0.01 cache read — the floor of this series now has two occupants, separated only by output, where Flash-Lite is a dime cheaper. Its long-context tier is $0.20 / $0.75 (cache read $0.02), so OpenAI’s cheapest row is also the first budget row in this table to carry a long-context cliff. §Claude Haiku 5.5, added 2026-10-14, makes the floor a three-way tie: $0.10 / $0.50 with a $0.01 cache read, identical to GPT-6-luna, and one tenth of Claude Haiku 4.5 on input and output. Its rate holds only for prompts up to 100,000 tokens; above that it pays $0.50 / $2.50 (cache read $0.05), and Anthropic counts cache reads and writes toward the threshold. Haiku 4.5 remains listed at unchanged prices. GPT-5.4-nano is deprecated: OpenAI’s deprecations page schedules its removal for 2027-04-01 and names GPT-6-luna, which is half its input price, as the replacement. The entry is dated 2026-10-01 and this table first records it on 2026-10-26; the row stays at unchanged prices until removal. Mistral Large 4 left this table on 2026-10-26 for the mid-range one when its sale ended. Mistral’s page also carries a cached-input column at 0.1x base input, recorded here for the first time; earlier editions showed “n/p” because the series had not captured it, not because a price changed.
Coding and Specialist Models
This tier shrank sharply in the week to 2026-08-17. Mistral’s pricing page no longer lists Magistral Medium, Magistral Small, Devstral 2, or Devstral Small 2, and its models overview gives retirement dates for all four running from 2025-10-31 to 2026-07-31. Four of the seven rows this section carried the week before were models Mistral had already withdrawn — a reminder that a pricing page is a better record of what exists than of what stopped existing.
| Model | Provider | Input $/1M | Output $/1M | Context |
|---|---|---|---|---|
| Grok-build-0.1 (coding) | xAI | $1.00 | $2.00 | 256K |
| Codestral (coding) | Mistral | $0.30 | $0.90 | n/p |
| Ministral 3 (3B / 8B / 14B) | Mistral | $0.10–0.20 | $0.10–0.20 | n/p |
Grok-build-0.1 is billed at $2.00 / $4.00 above 200K tokens. Ministral 3 models are priced symmetrically (identical input and output rates), which is rare — output tokens typically cost 3–6x input. Ministral 3 sits outside the weekly-tracked snapshot set and is shown here for completeness of Mistral’s lineup, without week-over-week change tracking. Also on Mistral’s pricing page but outside the tracked set: Z.ai’s GLM 5.3 at $1.40 / $4.40, resold under Mistral’s API. Third-party resales are excluded from this series because the reseller does not set the model; the earlier GLM 5.2 resale no longer appears on the page as of 2026-10-14.
The Fine Print That Changes the Rankings
A per-token price table looks objective, but three structural details buried in provider documentation change what these numbers actually mean.
First, tokens are no longer a stable unit. Anthropic states that Claude 4.7 and later models, plus Mythos Preview, use a tokenizer producing roughly 30% more tokens for the same text, and that Sonnet 4.6 and earlier use the previous one. The split runs through Anthropic’s own table: eleven of the sixteen rows we track count one way, five the other.
The effect is an invisible price. Sonnet 5’s $2.00 input is closer to ~$2.60 per old-tokenizer-equivalent page, which erases most of its apparent advantage over GPT-5.4’s $2.50 — and makes the $3.00 Sonnet 4.6 the cheaper row for some English text, despite costing 50% more per token. That inversion appears nowhere in the price columns.
Among the five rows now priced at $2.00 / $10.00, this is one of only two things that separate them — the other is cache read. No other provider publishes tokenizer efficiency data at all, which makes cross-provider per-token comparison an approximation, not an equivalence. This is why cost-per-task benchmarking is replacing cost-per-token math in serious procurement.
Second, the price war moved from base rates to cache reads — but not everywhere, and not permanently. Anthropic, OpenAI, and Google all priced a cache read at 10% of base input, a flat 90% discount. All three have now gone past it on at least one model. Google was the last, on October 14.
On September 7 Anthropic broke its own flat rule: Claude Fable 5.1 and Mythos 5.1 bill cache hits at 0.025x base input, a 97.5% discount, while every older Claude model stayed at 0.1x. That is a discount of unusual size sitting on the most expensive models a vendor sells rather than the cheapest — a shape that only makes sense if the buyer being courted is one whose bill is mostly re-sent context and who is not shopping on sticker price.
On September 28 a third step appeared. Claude Opus 5.5 bills a cache hit at 0.05x base input — $0.20 against a $4.00 base, 95% off — so Anthropic now runs three multipliers simultaneously: 0.025x on Fable 5.1 and Mythos 5.1, 0.05x on Opus 5.5, 0.1x on everything else. What read as one exception for two halo models three weeks ago now reads as a schedule, and it slopes with price. The more a model costs per input token, the smaller the fraction Anthropic charges to re-read one.
By October 14 Anthropic had extended the 0.05x step down a tier. Claude Sonnet 5.5, which arrived on October 5 at the standard 0.1x, now bills a cache hit at $0.10 against a $2.00 base. That breaks the slope described above: the 0.05x multiplier now covers a $4.00 model and a $2.00 one, while the $5.00 Opus 5 and the $10.00 Fable 5 stay at 0.1x. The better predictor is no longer price but recency — the deeper discounts sit on the newest models of each line.
On October 5 OpenAI made the same move, and made it without touching a list price. GPT-6.1-sol arrived at GPT-6-sol’s $2.00 / $10.00 with the cache read halved to $0.10 — 0.05x base input, on the short-context and long-context tiers alike. Every OpenAI row before it had sat at 0.1x. The shape differs from Anthropic’s: OpenAI did not reserve the deeper discount for its most expensive model, since GPT-6-astra at $10.00 base stays at 0.1x. It put it on the successor to a mid-priced flagship, at the same sticker as the model it replaces.
On September 14 DeepSeek took the title back at 98%, and it got there by a different route. Anthropic cut a multiplier. DeepSeek changed the model: deepseek-flash prices a cache hit at $0.003 against $0.15 base input, restoring roughly the 98% ratio v4-flash carried before August 17 and briefly lost.
The competitive fact is that four vendors are now bidding on the cache column specifically — DeepSeek near the bottom of the table, Anthropic at the top, OpenAI since October 5 and Google since October 14 in the middle. Gemini 4 Argon bills a cache read at $0.10 on a $2.00 base, 0.05x, so the last of the three original flat-90% vendors now has a row below 0.1x. xAI sits at 75–85%, and Mistral, whose cached-input column this series first recorded on October 14, sits at a flat 90%.
xAI shows the direction running both ways within one vendor. It cut Grok 4.5’s cache read from $0.50 to $0.30 between our July 16 and July 28 snapshots, deepening that discount from 75% to 85% while leaving the $2.00 base input untouched. Then on August 17 it launched Grok 4.6 at the same $2.00 base with a $0.50 cache read — reverting to the 75% discount for the newer model. If cache reads are where the competition is, xAI just made its newest model less competitive on exactly that axis.
For agentic workloads, where the same system prompt and tool definitions are re-sent on every call, cached input routinely dominates total token volume. In that regime, the cache-read column of the tables above predicts your bill better than the input column does. A provider with a higher base price and cheaper effective caching can win on real invoices.
The cache-read column alone does not settle it, though, because the write fee and the minimum cacheable prefix differ by vendor and neither appears in this table. We normalized those into an effective cost per hit rate in LLM Cache Pricing 2026, which is the companion to this page for anyone whose bill is mostly repeated input.
Third, pricing structure — flat versus tiered — is a strategic split, not an accounting detail. Anthropic explicitly bills its 1M-token context window at standard rates: a 900K-token request costs the same per token as a 9K one. Google and xAI took the opposite path, roughly doubling per-token rates above 200K tokens, and OpenAI crossed that line on August 24: GPT-5.6-sol now carries a long-context tier at $8.00 / $30.00.
That left Anthropic as the only one of the four billing long context flat — until October 14. Claude Haiku 5.5 is priced by prompt length: up to 100,000 tokens it is $0.10 / $0.50, and above that every token in the request costs 5x. The threshold sits at half the 200K line Google and xAI use, and the step is steeper than their roughly 2x.
Anthropic’s larger models still bill the full 1M window flat, so the split now runs inside one vendor’s line: flat pricing on the models sold for long agent sessions, a cliff on the one sold for high-volume classification and routing.
Flat pricing sells predictability to agent builders whose context sizes vary wildly at runtime; tiered pricing protects margins on the expensive long-context serving path. Which structure wins will shape how retrieval-augmented and long-context architectures are designed, because a 2x cliff at 200K tokens is an architectural forcing function.
The Bigger Picture
The market is not uniformly deflating, and the week to August 17 exposed that “deflation” has been hiding two different things.
Two of the three price moves came with a clock attached. Google’s Gemini Flash rows are $0.75 / $3.75 only through December 31, 2026, and the pricing page already prints the January 1 rate beside it. Anthropic used the same instrument in the opposite direction, converting Sonnet 5’s introductory rate into the standard one and cancelling an increase it had pre-announced. Both are promotional pricing being managed. Neither says anything about what it costs to serve those models.
DeepSeek is the counterexample, and five weeks later it is a more complicated one than it looked. On August 17 it raised prices outright — 57% on v4-flash input, 136% on output — with no promotional clock and no pre-announced figure. We recorded that as a structural reprice. On September 14 it walked most of it back, but only on one axis and only by replacing the model.
The asymmetry is the finding. Input returned to within a cent of the pre-increase rate; output stayed at more than double it. Across the three states DeepSeek’s output-to-input ratio went 2x, then 3x, then 4x — converging on the 3–6x every other provider in these tables charges, and away from the cheap-generation model it launched as. Those two axes bill different workloads. Extraction, classification and retrieval are input-dominated and are close to their July prices again; generation and long-form reasoning are output-dominated and are paying the August increase in full. A vendor that needed margin on generation and not on ingestion would produce exactly this shape, and it is invisible in any table that quotes one blended price per model.
It also shows what a version bump can carry. Nothing on the page reads as a price change — the name that appears in customer code still resolves, and the invoice per token falls. What changed underneath is which model answers. A buyer tracking cost per token would log a discount here; a buyer tracking cost per task has to re-measure, because the thing being measured was replaced.
Frontier prices were the part of this table that only ever moved up — Fable 5 at $10 / $50, OpenAI’s pro tier at $30 / $180 — on the theory that frontier capability is price-inelastic. September 28 is the first collection that argues the other way, and it does so without editing a single cell. Anthropic’s newest Opus is priced below the Opus it succeeds. OpenAI’s newest sol and luna are priced at half of theirs.
Nothing was cut. The cuts arrived wearing new model names, and that distinction decides who receives them. A buyer who pins a model id — which is what every production deployment does, and what DeepSeek’s September 14 substitution punished buyers for not doing — sees none of this on next month’s invoice. The saving is available only to whoever edits a config and re-qualifies a model. Price competition conducted this way rewards the teams with capacity to re-benchmark and bills everyone else at last generation’s rate, which is the opposite of how a published price cut behaves.
October 5 narrowed that further. GPT-6.1-sol carries exactly the list price of the GPT-6-sol it succeeds, so a buyer screening on input and output columns sees no reason to move at all. The whole saving sits in the cache read, and it is worth having only to a workload that re-sends context. A list-price comparison now misses a reduction, not merely delays it.
October 14 complicated the pattern without reversing it. Claude Sonnet 5.5’s cache cut is the first reduction since August 24 that a pinned model id received, but the id was nine days old when it happened; almost nothing in production had pinned it yet. The large saving of the collection still arrived the old way — Haiku 5.5 at one tenth of Haiku 4.5, under a new name.
And Google showed a third way to move a price: withdraw an id, route its traffic to a newer model, and leave unstated what the routed request costs. A caller who never changed a config may now be paying a different rate for a different model, which is the DeepSeek substitution of September 14 again, this time at Google.
October 19 recorded the quietest form of the same move. DeepSeek upgraded the model behind deepseek-flash on October 17 and left every price where it was, so nothing in this table registers it. The vendor’s own description is a performance-per-cost gain: higher benchmark scores, shorter runtimes and shorter outputs. Those are vendor claims we have not measured, but the direction they point is the one that matters for a pinned buyer.
A shorter output at the same per-token rate is a cheaper task, so a pinned buyer may have received a real reduction this week that no price column will ever show. That is the mirror image of Anthropic’s tokenizer change, where unchanged rates hide a larger bill. In both cases the per-token price held, and the cost per task moved anyway.
October 26 brought the first move in the opposite direction since August. Mistral Large 4’s price doubled when an undated sale ended, and the new rate is one the page had printed beside the sale price all along. Nothing about the model changed, and nothing on the page announced the change. A buyer who budgeted the sale price for a model with no expiry on it has just paid for that assumption.
It is the same lesson as the promotional clocks above, with the clock removed. GPT-5.6-sol, Gemini 4 Argon and the two Gemini Flash rows all state when their rates end. Large 4’s page never did, so the honest reading of its sale price was always a ceiling on how long the rate would last, not a floor. A sale price with no date should be budgeted at the list price beside it.
What is newly in question is the assumption underneath the rest of the table: that the commodity tier moves in one direction, and that a falling number means the same product got cheaper.
This matters beyond procurement because API pricing is becoming the cost structure of AI-mediated traffic itself. As we documented in our analysis of ChatGPT ads and AEO economics, AI answers are now a monetized distribution channel; the margin between what a model costs to run and what its answers earn determines which models power that channel. And token efficiency is not only a provider-side variable — vague prompts inflate token consumption and degrade output quality simultaneously, so prompt discipline compounds directly with the per-token rates in these tables.
How to Choose: Recommendations by Workload
High-volume extraction and classification. Gemini 2.5 Flash-Lite ($0.10 / $0.40) holds the price floor among maintained models, and now shares it with two others — GPT-6-luna and Claude Haiku 5.5 both match its $0.10 input and $0.01 cache read, and cost a dime more on output.
Haiku 5.5’s rate holds only below 100,000 prompt tokens, which most classification calls never approach; a batch job that packs many documents into one long prompt can cross it and pay 5x. The September 14 swap brought DeepSeek back to $0.15 / $0.60, a tie with Mistral Small 4 and 50% above Flash-Lite on input rather than the 120% the August increase had opened. For stateless pipelines Flash-Lite still wins on sticker price; DeepSeek’s 1M context and near-free cache reads make it the stronger pick as soon as documents repeat across requests.
Agentic workloads with heavy context reuse. Rank by cache-read price, not input price: DeepSeek flash ($0.003), Gemini 2.5 Flash-Lite, GPT-6-luna and Claude Haiku 5.5 ($0.01 each), Mistral Small 4 ($0.015), GPT-5.4-nano and GPT-5.6-luna ($0.02 each), then Claude Haiku 4.5, Claude Sonnet 5.5, GPT-6.1-sol and Gemini 4 Argon ($0.10 each). Three flagship-class rows now sit at that $0.10, half of the $0.20 that GPT-6-sol and Claude Sonnet 5 charge.
If your agent re-sends a 50K-token system prompt on every call, these numbers are your effective input price. DeepSeek’s September 14 swap restored a 3.3x lead over Flash-Lite here, up from the 1.4x it had been cut to in August — but the lead now belongs to a build DeepSeek shipped on October 17, not to the one that was benchmarked into these pipelines.
Long-context analysis. Claude Sonnet 5 ($2.00 / $10.00, flat 1M context) is the standout, and since August 17 it is no longer on a clock — flagship-adjacent quality, no long-context surcharge, and no scheduled increase. Claude Sonnet 5.5 carries the identical list rate and the same flat 1M window, and since October 14 half the cache read, so for long contexts that are re-read it is now the cheaper of the two. Avoid Haiku 5.5 here despite its price: above 100,000 prompt tokens it bills $0.50 / $2.50. Grok 4.3 ($1.25 / $2.50 base, 1M) undercuts it on sticker price, but only if your prompts stay under the 200K-token cliff where xAI’s rates double.
Frontier reasoning. This tier changed on September 28. Claude Opus 5.5 at $4.00 / $20.00 matches GPT-5.6-sol’s rate without the November deadline attached to it, which removes the reason to budget that $4.00 as temporary. It also undercuts the Opus 5 it succeeds by $1.00 / $5.00, so the pinned-revision option is now the more expensive one.
GPT-6.1-sol at $2.00 / $10.00 is cheaper still and is the row to test first: same list price as GPT-6-sol, half the cache read. Gemini 4 Argon matches it on all three figures, but only until December 1, when its page doubles it to $4.00 / $20.00 — budget Argon at the December rate unless a shorter evaluation is all you need.
Opus 5 ($5.00 / $25.00) and Opus 4.5 through 4.8 at the same rate remain the choice where a pinned older revision matters more than the dollar. Claude Fable 5 ($10.00 / $50.00) prices above all of them; the pro-grade OpenAI models ($30.00 / $180.00) only make sense where a single hard task justifies a 7x premium over Opus 5.5.
Limitations
- Standard tier only. Batch discounts (typically 50% off at Anthropic, OpenAI, Google, and Mistral) and regional/data-residency surcharges (10% at OpenAI, Anthropic and xAI for US-only routing) are not baked into the tables.
- Context windows marked “n/p” are not published on the provider’s official pricing page; we do not fill gaps from third-party sources.
- Mistral’s cache-read prices are new to this series. They were first recorded on 2026-10-14; earlier editions showed “n/p” for them, so Mistral rows cannot be compared on cache read before that date.
- Coverage is six providers (54 models tracked). The tables show the notable rows and footnote price-identical siblings, so the visible row count is lower than the tracked count. Open-source hosted inference (Groq, Together, DeepInfra) and Alibaba/Qwen are excluded from this edition; we would rather ship a fully verified narrow table than a broad one with stale rows.
- DeepSeek’s rows are off-peak rates and are not strictly comparable to the rest of the table. Every other provider publishes one rate that applies at all hours. DeepSeek publishes two, and a workload that runs during the 01:00–04:00 and 06:00–10:00 UTC peak windows on a weekday pays exactly double the figures shown; weekends and Chinese public holidays are off-peak in full. We record the lower tier for consistency with how tiered pricing is handled elsewhere, not because it is the rate most users will pay.
- Prices are list prices. Enterprise volume discounts are negotiated and unobservable.
Update Cadence and Changelog
This table is refreshed weekly from the six official pricing pages, and every change lands here as a dated entry. Official pricing pages only show current prices — the change history below, and the snapshot series behind it, is what this page accumulates that they do not.
| Date | Change |
|---|---|
| 2026-10-26 | One carried row repriced, upward: Mistral Large 4’s sale ended. $0.68 / $2.09 → $1.36 / $4.18, cached input $0.07 → $0.14, +100% on all three. The new rate is the struck-through list price the page showed when the row entered on 2026-10-14; no end date was ever given, so the change happened at some point between 2026-10-19 and 2026-10-26. It is the first increase on a carried row since DeepSeek’s on 2026-08-17, and the row moves from the budget table to the mid-range one. The other 53 rows hold every price. No model entered or left the set; the tracked count stays at 54. Correction: OpenAI’s deprecations page lists GPT-5.4-nano for removal on 2027-04-01, replacement GPT-6-luna, in an entry dated 2026-10-01. The 2026-10-05, 2026-10-14 and 2026-10-19 editions did not record it; it is recorded from this edition, and the earlier entries are left as written. Newly recorded outside the tracked set: OpenAI’s 2026-10-14 notice removing GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini on 2027-01-14; Z.ai’s GLM 5.3 Flash resold on Mistral’s API at $0.15 / $0.50, excluded with the other resales. OpenAI’s 2026-10-23 shutdown of o3-mini, gpt-4-0613 and other older models has passed; none were tracked rows. Anthropic’s Claude Haiku 4.5 is still Active with no deprecation notice, eleven days past its 2026-10-15 floor. Dated commitments still standing: DeepSeek tool-call billing from 2026-11-01, GPT-5.6-sol through at least 2026-11-21, Claude Sonnet 4.5 retirement on 2026-11-30, Gemini 4 Argon through 2026-11-30, and Gemini 3.8 and 3.6 Flash through 2026-12-31. |
| 2026-10-19 | Nothing repriced, and a model changed behind an existing id. All 54 tracked rows hold the input, output and cache-read prices recorded on 2026-10-14, across all six providers, and no model entered or left the set — an eighth consecutive collection with no input or output change on the carried set. Collected Monday 2026-10-19, five days after the Wednesday 2026-10-14 file, which returns the series to its Monday cadence. One substitution with no price effect: on 2026-10-17 DeepSeek upgraded the model behind deepseek-flash from DeepSeek-V4.1-Flash to DeepSeek-V4.1-Flash-1017, stating that the calling method and pricing are unchanged; the legacy ids deepseek-v4-flash and deepseek-v4-flash-vision-exp are now served by the new build. Newly recorded, none of them row prices: DeepSeek’s standalone Search and Fetch APIs at $0.75 per 1,000 successful requests each, with server-side tool calls through DeepSeek Harness, Claude Code or Codex billed from 2026-11-01 (Beijing time); xAI’s 1.1x rate for its US regional endpoint, currently on Grok 4.7 and Grok 4.6 only; and Anthropic’s 1.1x multiplier for US-only inference on Claude 4.6 and later, which this series had not recorded before and which was not established as new. Anthropic’s tentative retirement floor for Claude Haiku 4.5 (no sooner than 2026-10-15) passed with no deprecation notice; the row stays at unchanged prices. Mistral Large 4 remains at its sale price with no end date. Dated commitments still standing: DeepSeek tool-call billing from 2026-11-01, GPT-5.6-sol through at least 2026-11-21, Claude Sonnet 4.5 retirement on 2026-11-30, Gemini 4 Argon through 2026-11-30, and Gemini 3.8 and 3.6 Flash through 2026-12-31. |
| 2026-10-14 | The first carried-row price change since 2026-08-24, and it is a cut; three models arrived and two left. Collected Wednesday 2026-10-14: the Monday 2026-10-12 slot was missed and is left as a gap, so this entry covers nine days and the date of each change within them is unknown. Of the 51 rows carried from 2026-10-05, 50 hold every price. Claude Sonnet 5.5’s cache read fell from $0.20 to $0.10 (−50%): Anthropic’s page now lists it under the 0.05x multiplier with Claude Opus 5.5; input, output and cache writes are unchanged, and Claude Sonnet 5 stays at $0.20. Input and output held on all 51 carried rows. Tracked set 53 → 54. Added: Claude Haiku 5.5 at $0.10 / $0.50 (cache read $0.01, 1M context) for prompts up to 100,000 tokens and $0.50 / $2.50 above — one tenth of Claude Haiku 4.5, a three-way tie at the floor with GPT-6-luna and Gemini 2.5 Flash-Lite on input and cache read, and the only Claude 4.6-and-later model priced by prompt length rather than flat across its window; it uses the newer tokenizer, so eleven of the sixteen Anthropic rows now count roughly 30% more tokens. Added: Gemini 4 Argon at $2.00 / $10.00 (cache read $0.10), introductory through 2026-11-30 and $4.00 / $20.00 from 2026-12-01. Added: Mistral Large 4 at a sale price of $0.68 / $2.09 (cached $0.07) against a struck-through $1.36 / $4.18, with no end date stated. Removed: Gemini 3.7 Flash and Gemini 3.5 Flash, deprecated 2026-10-08 and gone from the pricing page; Google routes their ids to Gemini 3.8 Flash and 3.6 Flash respectively and does not state the billing for routed requests. Newly recorded, not a price change: Mistral’s cached-input column at 0.1x base input on all four carried Mistral rows. Newly observed outside the tracked set: an OpenAI Ultrafast tier at six times standard on GPT-6-astra and GPT-6.1-sol; OpenAI’s access-restricted life-sciences models gpt-rosalind-research and gpt-rosalind-discovery at $5.00 / $25.00; xAI’s Grok 4.7 Fast, sold only inside Cursor and Grok Build. Mistral no longer lists its GLM 5.2 resale or the free Leanstral 1.5 endpoint. Dated commitments now standing: GPT-5.6-sol through at least 2026-11-21, Gemini 4 Argon through 2026-11-30, and Gemini 3.8 and 3.6 Flash through 2026-12-31. |
| 2026-10-05 | Nothing repriced; two models arrived at their predecessors’ list price, and one of them halved the cache read. All 51 rows carried from 2026-09-28 hold their input, output and cache-read prices — a sixth consecutive collection with no price change on the carried set, across all six providers. Nothing was removed. Tracked set 51 → 53. Added: Claude Sonnet 5.5 at $2.00 / $10.00 (cache read $0.20, 1M context), identical to Claude Sonnet 5 on every line of Anthropic’s table including cache writes ($2.50 / $4.00) and batch ($1.00 / $5.00), and on the newer tokenizer — so ten of the fifteen tracked Anthropic rows now count roughly 30% more tokens for the same text. Added: GPT-6.1-sol at $2.00 / $10.00, GPT-6-sol’s list price, with the cache read at $0.10 against GPT-6-sol’s $0.20 — 0.05x base input, the first OpenAI row in this series below 0.1x; long-context tier $4.00 / $15.00, cache read $0.20. OpenAI moved GPT-6.1-sol into its three-row flagship table in GPT-6-sol’s place, and GPT-6-sol stays listed in the full table at unchanged prices. Neither predecessor was removed. One carried row changed status without changing price: Anthropic deprecated Claude Sonnet 4.5 on 2026-09-30 and scheduled its retirement for 2026-11-30, naming Claude Sonnet 5.5 as the replacement; it stays in the tracked set as deprecated-but-available, and its $3.00 / $15.00 is unchanged. Anthropic’s pricing page labels only retired models, so this status comes from its model deprecations page, read the same day. Two DeepSeek page changes with no price effect: the page states that peak hours apply Monday through Friday, with weekends and Chinese public holidays off-peak in full — recorded here for the first time, and whether it is new could not be established — and it no longer carries the statement that v4-pro service continues past 2026-09-14 with billing unchanged, though v4-pro is still listed at the same rate. Newly observed outside the tracked set: Mistral’s Leanstral 1.5, a Lean 4 proof-engineering model offered as a free endpoint for a limited period with no per-token price; Google’s Gemini 3.8 Flash TTS, 3.8 Flash-Lite TTS and 3.8 Live, under the existing TTS and live exclusions. Still outside the set and unchanged: Gemini Omni Flash, Gemma 4, Mistral’s two GLM resales, the Ministral 3 line, OpenAI’s cyber line, and OpenAI gpt-5.2 and earlier. Two dated commitments still stand: GPT-5.6-sol through at least 2026-11-21, and the three Gemini Flash rows through 2026-12-31. |
| 2026-09-28 | Nothing repriced; four models arrived, three of them below the generation they succeed. All 47 rows carried from 2026-09-21 hold their input, output and cache-read prices — a fifth consecutive collection with no price change on the carried set, across all six providers. Nothing was removed. Tracked set 47 → 51. Added: Claude Opus 5.5 at $4.00 / $20.00, which is $1.00 / $5.00 below the Claude Opus 5 it succeeds, with a cache read at 0.05x base input ($0.20) — a third Anthropic multiplier alongside Fable 5.1’s 0.025x and the 0.1x the rest of the line carries — and fast mode at $8.00 / $40.00. Added: GPT-6-sol at $2.00 / $10.00 (cache read $0.20), exactly half of GPT-5.6-sol on both sides and identical to Claude Sonnet 5’s published rate, with a long-context tier at $4.00 / $15.00. Added: GPT-6-luna at $0.10 / $0.50 (cache read $0.01), half of GPT-5.6-luna on all three figures, which ties Gemini 2.5 Flash-Lite’s input and cache read and gives this table’s price floor a second occupant; long-context tier $0.20 / $0.75. Added: Grok 4.7 at $2.00 / $6.00 (cache read $0.50), identical to Grok 4.6 to the cent. The week therefore reads flat on a carried-set delta and cheaper at the frontier at once, and the two are recorded separately here on purpose: no pinned model id got cheaper, so the saving reaches only a buyer who changes one. Newly on Google’s page and outside the tracked set: Gemini Omni Flash (gemini-omni-1.1-flash) and Omni Flash Preview at $1.50 input, output $9.00 for text and $17.50 for video, described as a video generation and editing model billed at 5,792 tokens per second of 720p video — excluded on the same grounds as the image, TTS, live and transcription lines. Gemma 4 appears with every paid-tier cell reading “Not available”, so there is no API rate to record. Still outside the set and unchanged: Mistral’s two GLM resales (5.3 and 5.2, both $1.40 / $4.40), the Ministral 3 line, OpenAI’s cyber line, and OpenAI gpt-5.2 and earlier. Two dated commitments still stand: GPT-5.6-sol through at least 2026-11-21, and the three Gemini Flash rows through 2026-12-31. |
| 2026-09-21 | Nothing moved. All 47 tracked rows hold the input, output and cache-read prices recorded on 2026-09-14, across all six providers, and no model entered or left the tracked set. This is the fourth consecutive collection without a price change on the carried set and the first of the four in which the set is also identical in composition — the three before it each carried an arrival, a removal or a substitution. Two entries are recorded that are not price changes. Anthropic’s tokenizer split (Claude 4.7 and later counting roughly 30% more tokens for the same text) is now named row by row in our snapshot series; it has been on the provider’s page and in this post’s Fine Print section since July, and is recorded now so that the snapshot files do not read as though the eight affected rows were directly comparable to the five that are not. And OpenAI’s pricing page lists a cyber line — gpt-5.6-cyber and gpt-5.5-cyber, both $12.50 / $75.00 with cached input $1.25 — that this series had never recorded; that gpt-5.5-cyber pairs with a generation already outside our July 16 scope suggests it predates the tracked set rather than arriving this week, but the page carries no dates, so that is inference and the rows stay out pending a scope decision. Mistral added GLM 5.3 at $1.40 / $4.40, a third-party Z.ai model resold on its API, joining the GLM 5.2 resale already excluded on the same grounds. Two dated commitments still stand: GPT-5.6-sol through at least 2026-11-21, and the three Gemini Flash rows through 2026-12-31. |
| 2026-09-14 | DeepSeek retired v4-flash; the row was replaced, not repriced. The pricing page now lists deepseek-flash (build DeepSeek-V4.1-Flash) at $0.15 / $0.60 off-peak, cache read $0.003, and states that the legacy ids deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted but that the models behind them have been retired, with those requests served and billed at the Flash rate. Against the rate v4-flash carried last week this is −31.8% on input, −9.1% on output and −57.1% on cache read; against the rate it carried before the 2026-08-17 increase it is +7% on input and +114% on output. The August increase was therefore reversed on input and kept on output, and DeepSeek’s output-to-input ratio went 2x (pre-August) to 3x to 4x across the three states. The 98% cache discount retakes the deepest-cache position from Claude Fable 5.1 eight days after Anthropic took it. No other row moved — a third consecutive collection with no price change on the carried set, across all six providers. Tracked set holds at 47: one out, one in. DeepSeek also states that v4-pro API service continues past 2026-09-14 with billing unchanged, cancelling a date that had been read as an end-of-life; v4-pro’s own prices are untouched. Two dated commitments still stand: GPT-5.6-sol through at least 2026-11-21, and the three Gemini Flash rows through 2026-12-31. Still outside the tracked set and unchanged: Mistral’s resale of Z.ai GLM 5.2 and the Ministral 3 line. |
| 2026-09-07 | No price change, four arrivals. All 43 rows carried over from 2026-09-01 hold their prices — a second consecutive collection with no move anywhere across the six providers. Tracked set 43 → 47. Added: Claude Fable 5.1 and Claude Mythos 5.1 at $10.00 / $50.00, matching Fable 5 and Mythos 5 on base rates but pricing cache hits at 0.025x base input ($0.25 / 1M) against the 0.1x every other Claude model uses — the first split in Anthropic’s cache-read multiplier recorded in this series, and at 97.5% the deepest cache discount in these tables. Added: GPT-6-astra at $10.00 / $50.00 (cache read $1.00) with a $20.00 / $75.00 long-context tier, the most expensive OpenAI row outside the $30.00 / $180.00 pro tier. Added: Gemini 3.8 Flash at $0.75 / $3.75 (cache read $0.075) on the same 2026-12-31 introductory expiry as Gemini 3.7 and 3.6 Flash, taking the rows scheduled to double on 2027-01-01 from two to three. Nothing was removed. Collected Monday 2026-09-07, six days after the previous file rather than seven, because 2026-09-01 was collected on a Tuesday. Two dated commitments still stand: GPT-5.6-sol through at least 2026-11-21, and the three Gemini Flash rows through 2026-12-31. Still outside the tracked set and unchanged: Mistral’s resale of Z.ai GLM 5.2, the Ministral 3 line, and DeepSeek’s deepseek-v4-flash-vision-exp. |
| 2026-09-01 | No change. All 43 tracked rows hold the prices recorded on 2026-08-24, across all six providers — the first collection in this series with no price move anywhere. No model entered or left the tracked set. Collected Tuesday 2026-09-01 rather than Monday; the 2026-08-31 slot was missed and is left as a gap rather than backfilled, so this entry covers eight days. Two dated commitments still stand on the pages: GPT-5.6-sol’s promotional rate through at least 2026-11-21, and the 2026-12-31 expiry on both Gemini Flash rows. Anthropic’s cancelled Sonnet 5 increase reached its 2026-09-01 date with the rate unmoved. Still outside the tracked set and unchanged: Mistral’s resale of Z.ai GLM 5.2 and the Ministral 3 line. |
| 2026-08-24 | OpenAI cut GPT-5.6-sol from $5.00 / $30.00 to $4.00 / $20.00 (cache read $0.50 → $0.40) — input −20%, output −33.3%, the largest single-model cut recorded here since the GPT-5.6-luna cut on 2026-08-03. The page marks the rate promotional and available at least through 2026-11-21, so it is recorded as a dated cut rather than a new standard price. The same model gained OpenAI’s first long-context tier at $8.00 / $30.00 (cache read $0.80); every other OpenAI row stays flat. GPT-5.6-terra and GPT-5.6-luna did not move, and no other provider changed a price. No models added or removed — tracked set holds at 43. Still outside the tracked set and unchanged from last week: Mistral’s resale of Z.ai GLM 5.2 ($1.40 / $4.40) and the Ministral 3 line. DeepSeek added deepseek-v4-flash-vision-exp at v4-flash rates, excluded here as a non-text model. |
| 2026-08-17 | The largest week in the series. DeepSeek raised prices, the first increase recorded here from any provider: v4-flash $0.14 / $0.28 → $0.22 / $0.66 (cache read $0.0028 → $0.007) and v4-pro $0.435 / $0.87 → $0.66 / $1.98 (cache read $0.003625 → $0.022), all off-peak. This is the increase the 2026-08-10 entry flagged as expected but unsized. DeepSeek also shipped build V4-Pro-0813. Google halved Gemini 3.6 Flash to $0.75 / $3.75 (cache read $0.075) to match new arrival Gemini 3.7 Flash, launched 2026-08-13 at the same rate; both are promotional through 2026-12-31 and double on 2027-01-01. Anthropic cancelled the Claude Sonnet 5 increase — $2.00 / $10.00 is now the standard price and the 2026-09-01 move to $3.00 / $15.00 will not occur. Added: Gemini 3.7 Flash, Grok 4.6 ($2.00 / $6.00, cache read $0.50). Removed: Claude Opus 4.1, retired from the first-party API on 2026-08-05, and Magistral Medium, Magistral Small, Devstral 2, and Devstral Small 2, all absent from Mistral’s pricing page with retirement dates between 2025-10-31 and 2026-07-31. Those five had been carried in earlier snapshots after their retirement dates passed; the earlier files are left as collected rather than corrected. Tracked set: 46 → 43 models. |
| 2026-08-10 | No change. Every row in the tables above holds last week’s price, and no model entered or left the six providers’ pricing pages — the first flat week since the series began. Two things visible on the pages but not yet in the table: Anthropic’s Sonnet 5 introductory rate still expires 2026-08-31, and DeepSeek’s pricing page states that a significant increase is expected, without naming a date or a figure. Neither is recorded as a price until it takes effect. |
| 2026-08-03 | OpenAI cut two GPT-5.6 models: luna from $1.00 / $6.00 to $0.20 / $1.20 (−80%, cache read $0.10 → $0.02) and terra from $2.50 / $15.00 to $2.00 / $12.00 (−20%, cache read $0.25 → $0.20). Both were cross-verified against OpenAI’s models page before recording. No models added or removed; every other row across the six providers was unchanged. Gemini 3.5 Flash-Lite’s $0.03 cache read was recorded for the first time — a gap in the prior collection, not a price change. |
| 2026-07-28 | Added Claude Opus 5 and Gemini 3.6 Flash. Grok 4.5 cache read cut from $0.50 to $0.30 (−40%), the only base-table price change. DeepSeek’s legacy IDs retired on schedule. No snapshot was taken on 2026-07-20 or 2026-07-27; those two weeks are absent from the series and were not reconstructed. |
| 2026-07-16 | Initial 2026 edition: baseline of 39 models across 6 providers. Supersedes the March 2026 draft dataset. |
FAQ: LLM API Pricing 2026
What is the cheapest LLM API in 2026?
Gemini 2.5 Flash-Lite, at $0.10 input / $0.40 output per 1M tokens — though it shares the floor rather than owning it. GPT-6-luna arrived on September 28, 2026 and Claude Haiku 5.5 on October 14, both at $0.10 / $0.50, matching Flash-Lite’s input and its $0.01 cache read and separating only on output. Haiku 5.5’s rate applies only to prompts up to 100,000 tokens.
DeepSeek is next at $0.15 / $0.60, tied with Mistral Small 4, after retiring v4-flash on September 14, 2026 and serving the name from deepseek-flash. That swap undid the August 17 increase on input but not on output, which is why DeepSeek is close to its old input price and nowhere near its old output price. It is cheapest of all on cached input at $0.003 per 1M — a 98% discount — so a workload that reuses context should be priced on the cache-read column rather than this one.
Is Claude Sonnet 5’s price going up on September 1, 2026?
No. Anthropic cancelled the increase. The $2.00 / $10.00 rate announced at launch as introductory is now the standard price, and the pricing page states plainly that the scheduled move to $3.00 / $15.00 will not occur. This page carried the September deadline as its most time-sensitive number for six weeks; it is gone. If you budgeted a 50% cost increase for Sonnet 5 workloads from September, that reserve can be released.
Why is Claude Fable 5.1’s cache read cheaper than Claude Opus 5’s?
Anthropic prices a cache hit as a multiplier on base input, and since September 7, 2026 it has stopped using one multiplier for the whole line. There are now three. Claude Fable 5.1 and Claude Mythos 5.1 bill a cache hit at 0.025x base input ($0.25 per 1M tokens). Claude Opus 5.5, added September 28, bills at 0.05x ($0.20), and Claude Sonnet 5.5 joined it at 0.05x ($0.10) by October 14, a week after arriving at 0.1x. Every other Claude model, Opus 5 and the new Haiku 5.5 included, stays at 0.1x — which suggests the deeper multipliers are decided model by model rather than set as a new default.
Until October the schedule sloped with price. Sonnet 5.5’s move broke that: a $2.00 model now carries a deeper discount than the $5.00 Opus 5. What the deeper multipliers share now is recency — they sit on the newest model in each Anthropic line except Haiku.
Are per-token prices directly comparable across providers?
Only approximately. Anthropic’s newest models tokenize the same text into roughly 30% more tokens than its earlier models, and other providers publish no tokenizer efficiency data at all. Two models with the same $/1M price can produce meaningfully different invoices for the same documents — compare cost per task when the decision matters.
Do these prices include long-context surcharges?
No — tables record base tiers. Google and xAI roughly double per-token rates for prompts above 200K tokens (footnoted per table), and OpenAI joined them on August 24: GPT-5.6-sol is billed at $8.00 / $30.00 on long context against a $4.00 / $20.00 base, and its GPT-6 rows carry long-context tiers too. Anthropic bills its full 1M context at standard rates on every Claude 4.6-and-later model except Claude Haiku 5.5, which charges $0.50 / $2.50 above 100,000 prompt tokens against a $0.10 / $0.50 base.
If no price changed this week, can I assume my bill is flat?
Not on Anthropic models. The per-token rates in these tables are unchanged, but Anthropic’s page states that Claude 4.7 and later models count roughly 30% more tokens for the same text than Sonnet 4.6 and earlier. A price that holds and a token count that rises produce a bill that rises. Cost-per-token is the wrong unit for any comparison that crosses that line; measure cost per task instead.
Did LLM API prices change in the week to October 26, 2026?
One row did, and it went up. Mistral Large 4 moved from $0.68 / $2.09 to $1.36 / $4.18, and its cache read from $0.07 to $0.14, a 100% increase on all three. The sale it carried when it entered this table on October 14 ended, and the page gave no end date in advance. The other 53 tracked rows hold every price, and no model entered or left the set.
If you priced a Mistral Large 4 workload at the sale rate, double the token cost line. The broader lesson is about undated discounts: when a page prints a sale price beside a struck-through list price with no expiry, budget at the list price.
How often is this table updated?
Weekly, from official provider pages only, with every change recorded in the changelog above. Missed weeks are marked as gaps rather than back-filled: the 2026-07-20 and 2026-07-27 snapshots were not taken, and the 2026-10-12 collection was taken two days late and dated 2026-10-14, and reconstructing them after the fact would put unobserved prices into a series whose whole value rests on every row having been read off a live page. The series began July 16, 2026 and compounds into a price-change history that current-price-only official pages do not offer.
Related Resources
- ChatGPT Ads 2026: AEO Traffic and Reddit Citations — the revenue side of the same equation: what AI-mediated answers earn
- GIGO Prompt Engineering: Why Vague Prompts Fail — cutting token waste at the prompt level compounds with every rate in this table
Primary Sources
Data verified 2026-10-26. Machine-readable dataset: /data/llm-api-pricing-2026.json
- Anthropic API pricing — Anthropic
- OpenAI API pricing — OpenAI
- Gemini API pricing — Google
- xAI models and pricing — xAI
- DeepSeek API pricing — DeepSeek
- Mistral API pricing — Mistral AI