Post

LLM Cache Pricing 2026: The Real Cost of Cached Input

LLM cache pricing 2026 normalized across six providers: effective input cost at real hit rates, hidden write fees, and which vendors publish nothing at all.

Posted Updated
A shallow pool of still water in a carved dark stone basin, one shaft of light bending at the surface
A shallow pool of still water in a carved dark stone basin, one shaft of light bending at the surface
By Json House
45 min read
LLM Cache Pricing 2026: The Real Cost of Cached Input

Prompt caching is sold as a 90% discount, and it was true almost everywhere until September: Anthropic, OpenAI, and Google all priced a cache read at exactly one tenth of base input. Anthropic now runs three multipliers instead of one, OpenAI started running two on 2026-10-05, and Google followed on 2026-10-14 — none of the three original 90% vendors is flat any more. The number that decides your invoice is a different thing again. It is the effective input price — what you pay per million input tokens once your actual cache hit rate, the write surcharge, and the storage fee are all folded in.

On that measure the answer reversed on 2026-09-14, four weeks after it had reversed the other way, and it has held there since. DeepSeek’s August rise had pushed the budget-tier crossover out to 97.6%, past where any real workload sits, and sticker price took over the ranking. Retiring v4-flash for deepseek-flash at $0.15 base / $0.003 cache read brought that crossover back to 87.7% — inside the band where long-running agents operate. At a 95% hit rate DeepSeek is still the cheapest row on this page.

The flagship end flipped on 2026-09-28, when Claude Opus 5.5 paired a $4.00 base with a 0.05x cache read and undercut Grok 4.6’s $2.00 base above 87.0%. The 2026-10-05 change is quieter and arguably more useful. GPT-6.1-sol arrived at GPT-6-sol’s $2.00 list price with the cache read halved to $0.10, also 0.05x. It ties the older model at a 0% hit rate and beats it at every rate above that, so the saving is invisible on a price list and certain on an invoice.

The 2026-10-14 collection repeated that move on a row that already existed. Claude Sonnet 5.5’s cache read fell from $0.20 to $0.10, so it now sits on GPT-6.1-sol’s curve, and so does the new Gemini 4 Argon. Three vendors now sell a $2.00 model with a 95% cache discount. The same collection also filled a gap: Mistral’s per-model cache prices, absent from this page since it began, are now recorded, and they produce the lowest crossover on the page — Grok 4.3 × Mistral Medium 3.5 at 83.3%.

The 2026-10-19 collection changed no number on this page. Every base and cache-read price held, so all four effective columns, every discount figure and all eighteen crossovers stand as they were. What changed is the model behind the row that leads the 95% column: on 2026-10-17 DeepSeek replaced the build behind deepseek-flash with DeepSeek-V4.1-Flash-1017, and its change log states that pricing is unchanged. The $0.0104 figure is the same; the model earning it is not.

This page normalizes cache pricing for 24 models across six providers to a single unit, using prices collected on 2026-10-19, a Monday.

TL;DR

  • The lowest crossover is now 83.3%: Grok 4.3 × Mistral Medium 3.5. It appears because Mistral’s cache price was recorded for the first time on 2026-10-14, not because a price moved. Grok 4.6 × Claude Opus 5.5 (87.0%) and Flash-Lite × DeepSeek flash (87.7%) follow.
  • No price moved on 2026-10-19; the last change is still Claude Sonnet 5.5’s cache read, halved on an existing row on 2026-10-14. At $2.00 base / $0.10 it now discounts 95% and beats Claude Sonnet 5 at every hit rate above zero — $0.1950 per 1M effective input tokens at h = 95% against $0.2900.
  • Three vendors now share one $2.00 / 95% curve: Claude Sonnet 5.5, GPT-6.1-sol and the new Gemini 4 Argon. All three pass Grok 4.3 above 88.2% and Mistral Medium 3.5 above 90.9%.
  • Sticker price still loses the 95% column. DeepSeek flash costs $0.0104 at h = 95% against the $0.0145 that Flash-Lite, GPT-6-luna and Claude Haiku 5.5 share — though since 2026-10-17 a new DeepSeek build stands behind that figure.
  • No original 90% vendor is flat any more. 90% on most rows at Anthropic, OpenAI, Google and Mistral; 75–85% at xAI; 95% on Claude Opus 5.5, Sonnet 5.5, GPT-6.1-sol and Gemini 4 Argon; 97.5% on Fable 5.1 and Mythos 5.1; 96.7% and 98% at DeepSeek.
  • Caching still pays after two calls. Even at a 2x write multiplier, reusing a prefix three times beats not caching. Write fees are a rounding error for agents — and a real penalty for one-shot prompts.
  • “90% off” is not a 90% saving. The discount applies only to the cached share of your prompt. At a 50% hit rate it is a 45% cut.

Methodology

Base input and cache-read prices were collected on 2026-10-19 from the six providers’ official pricing pages, the same sources and normalization used in our weekly LLM API pricing table: Anthropic, OpenAI, Google, xAI, DeepSeek, Mistral. Caching mechanics — write cost, storage fee, TTL, trigger, minimum cacheable length — were last re-read from each provider’s prompt-caching documentation on 2026-08-03 and are unchanged: Anthropic, OpenAI, Google, xAI, and DeepSeek.

Effective input price is computed as base x (1 - h) + cache_read x h, where h is the cache hit rate: the share of input tokens served from cache. It deliberately excludes write and storage costs, which are handled separately below, because those are one-time or time-based rather than per-token and would otherwise make the columns non-comparable.

Every cell in the table below except the h = 0% column is computed by us from vendor rate cards; h = 0% is the vendor’s published base input price. All figures are USD per 1M tokens at the standard, non-batch, base long-context tier. Where a provider publishes nothing, the cell reads “not published” — we do not derive a figure and present it as a price. The one derivation we do make is labelled as such.

Effective Input Price by Cache Hit Rate

h = 0% is the sticker price. h = 95% is a realistic figure for an agent that re-sends a stable system prompt and tool definitions on every turn.

ModelProviderh = 0% (vendor)h = 50% (jsonhouse derived)h = 80% (jsonhouse derived)h = 95% (jsonhouse derived)Discount (jsonhouse derived)
Claude Fable 5.1Anthropic$10.0000$5.1250$2.2000$0.737597.5%
Claude Opus 5Anthropic$5.0000$2.7500$1.4000$0.725090%
Claude Opus 5.5Anthropic$4.0000$2.1000$0.9600$0.390095%
GPT-5.6-solOpenAI$4.0000$2.2000$1.1200$0.580090%
Claude Sonnet 5Anthropic$2.0000$1.1000$0.5600$0.290090%
Claude Sonnet 5.5Anthropic$2.0000$1.0500$0.4800$0.195095%
GPT-5.6-terraOpenAI$2.0000$1.1000$0.5600$0.290090%
GPT-6.1-solOpenAI$2.0000$1.0500$0.4800$0.195095%
GPT-6-solOpenAI$2.0000$1.1000$0.5600$0.290090%
Gemini 4 ArgonGoogle$2.0000$1.0500$0.4800$0.195095%
Grok 4.6xAI$2.0000$1.2500$0.8000$0.575075%
Grok 4.5xAI$2.0000$1.1500$0.6400$0.385085%
Mistral Medium 3.5Mistral$1.5000$0.8250$0.4200$0.217590%
Grok 4.3xAI$1.2500$0.7250$0.4100$0.252584%
Claude Haiku 4.5Anthropic$1.0000$0.5500$0.2800$0.145090%
Gemini 3.8 FlashGoogle$0.7500$0.4125$0.2100$0.108890%
Gemini 3.6 FlashGoogle$0.7500$0.4125$0.2100$0.108890%
DeepSeek v4-proDeepSeek$0.6600$0.3410$0.1496$0.053996.7%
Gemini 2.5 FlashGoogle$0.3000$0.1650$0.0840$0.043590%
GPT-5.6-lunaOpenAI$0.2000$0.1100$0.0560$0.029090%
DeepSeek flashDeepSeek$0.1500$0.0765$0.0324$0.010498%
Claude Haiku 5.5Anthropic$0.1000$0.0550$0.0280$0.014590%
GPT-6-lunaOpenAI$0.1000$0.0550$0.0280$0.014590%
Gemini 2.5 Flash-LiteGoogle$0.1000$0.0550$0.0280$0.014590%

Raw data: data/llm-cache-pricing-2026.json — machine-readable structured data for AI crawlers and citation.

Changes on 2026-10-14. Claude Sonnet 5.5 now has a row: its cache read fell from $0.20 to $0.10 (0.05x) between 2026-10-05 and 2026-10-14, which moved it off Claude Sonnet 5’s curve and onto GPT-6.1-sol’s. Gemini 4 Argon arrived on that same curve, $2.00 base / $0.10 cache read, Google’s first row below 0.1x; it supports implicit caching only, and its rate is introductory through 2026-11-30, doubling to $4.00 / $0.20 on 2026-12-01.

Claude Haiku 5.5 arrived on Gemini 2.5 Flash-Lite’s figures, $0.10 / $0.01, but only for prompts up to 100,000 tokens; above that it bills $0.50 / $0.05, and Anthropic counts cache reads and writes toward the threshold, so a long cached context can push a whole request onto the higher rate. Gemini 3.7 Flash was removed: Google deprecated it on 2026-10-08 and routes its id to Gemini 3.8 Flash, which carries the same figures.

One row arrived on 2026-10-05. GPT-6.1-sol carries GPT-6-sol’s $2.00 base with a $0.10 cache read — 0.05x, the first OpenAI row off the 0.1x multiplier — so its curve starts at the same point and falls faster. Its 1.25x cache write ($2.50) is unchanged from the rest of the GPT-5.6-and-later line. Claude Sonnet 5.5 also arrived, on Claude Sonnet 5’s base and cache read to the cent, and like Grok 4.7 is carried in our snapshot without a row here.

Three rows arrived on 2026-09-28. Claude Opus 5.5 is the first Anthropic model on a 0.05x cache-read multiplier, between Fable 5.1’s 0.025x and the 0.1x the rest of the line carries, and it undercuts Claude Opus 5 on base price and cache read at once. GPT-6-sol lands on GPT-5.6-terra’s figures exactly, and GPT-6-luna on Gemini 2.5 Flash-Lite’s, so both add a row without adding a distinct effective-price curve. Grok 4.7 also arrived, on Grok 4.6’s base and cache read to the cent; like Claude Mythos 5.1 it is carried in our snapshot without a row here, because a duplicate curve tells a reader nothing the row above it does not.

Both Gemini 3.x Flash rows are promotional through 2026-12-31 and double on 2027-01-01, and Gemini 4 Argon doubles a month earlier, on 2026-12-01 — between them those dates will move every Google figure in this table except the two Gemini 2.5 rows.

DeepSeek’s rows are off-peak; peak hours (01:00–04:00 and 06:00–10:00 UTC, weekdays only — DeepSeek states weekends and Chinese public holidays are off-peak in full) bill exactly double on both the base and the cache read, so the effective columns scale but the discount does not. The DeepSeek flash row is new in identity as well as price: on 2026-09-14 DeepSeek retired v4-flash and now serves that name from DeepSeek-V4.1-Flash, so this row’s figures describe a different model than the same row did last week.

Mistral’s pricing page now lists a per-model cached-input price, and this page records it for the first time on 2026-10-14: $0.15 per 1M on Mistral Medium 3.5, exactly 0.1x base input. Until this collection the row was left unfilled rather than derived from Mistral’s headline “−90%”, and the published figure turns out to match that headline. This series did not capture the column before, so whether it is new on Mistral’s page could not be established. Mistral still publishes no write price, storage fee or TTL.

What the Sticker Price Hides

The per-token discount is only one of four cost dimensions, and the other three are disclosed unevenly.

ProviderWrite costStorage feeCache TTLTriggerMin. cacheable
Anthropic1.25x (5 min), 2x (1 h)None5 min or 1 hcache_control512–4,096 by model
OpenAI1.25x on GPT-5.6+; none beforeNone≥ 30 min (GPT-5.6+)Automatic1,024
GoogleNot published$1.00–$4.50 per 1M tokens/hourNot publishedImplicit, on by default2,048–4,096
xAINot publishedNot publishedNot published (eviction-based)Automatic, keyedNot published
DeepSeekNot publishedNot published“Hours to days”, not guaranteedAutomaticNot published
MistralNot publishedNot publishedNot publishedNot publishedNot published

Anthropic’s minimum cacheable length varies more than any other provider’s: 512 tokens on Opus 5 and Fable 5, but 4,096 on Opus 4.5, Opus 4.6, and Haiku 4.5. A prompt that caches on one Claude model is silently not cached on another, and the API returns no error when this happens.

The Deep Analysis: Three Things This Table Says

The 90% discount is a coordinated price, not a cost. Anthropic, OpenAI, and Google independently arrived at exactly 0.1x base input for a cache read. Cache serving costs differ across three different inference stacks; the identical multiplier does not. This is what a converged commodity price looks like — the number stopped tracking cost and became a market expectation nobody can be the first to break.

The consensus held for six weeks and then broke from both ends within eight days. On 2026-09-07 Anthropic stopped applying one multiplier to its own line: Claude Fable 5.1 and Mythos 5.1 bill a cache read at 0.025x base input, 97.5% off, while every older Claude model stayed at 0.1x. On 2026-09-14 DeepSeek went to 98% — not by cutting a multiplier but by replacing the model, retiring v4-flash for deepseek-flash at $0.15 base and $0.003 cache read.

Three weeks later the split stopped looking like an exception. Claude Opus 5.5 arrived on 2026-09-28 at 0.05x, so Anthropic now runs three multipliers at once: 0.025x, 0.05x, 0.1x. Read against base prices, the schedule slopes — $10.00 models discount deepest, the $4.00 model next, everything from $5.00 down stays at 90%. A vendor that priced a cache read as one flat fraction of input has started pricing it as a function of how expensive the input is.

On 2026-10-05 OpenAI moved. GPT-6.1-sol bills a cache read at 0.05x, on both its short- and long-context tiers, while every other OpenAI row — GPT-6-astra at $10.00 included — stays at 0.1x. OpenAI’s version does not slope with price the way Anthropic’s does. It sits on a $2.00 model, the successor to one this series first recorded at the same price a week earlier.

On 2026-10-14 the last of the three followed, and Anthropic’s schedule lost its slope. Google launched Gemini 4 Argon at $2.00 base with a $0.10 cache read, 0.05x, while every other Gemini row stays at 0.1x. In the same collection Anthropic moved Claude Sonnet 5.5, a $2.00 model, to 0.05x — so its deeper multiplier now covers a $4.00 model and a $2.00 one while the $5.00 Opus 5 stays at 0.1x. Read across all three vendors, the 0.05x rows have one thing in common, and it is not price. Each is the newest model in its line.

The vendors past 90% now sit at both ends of the price list and, since October 5, crowd the middle of it: three of the 24 rows here are $2.00 models discounted 95%. A deep cache discount at the bottom of the market is a share play on volume. The same discount at $10.00 base is aimed at a buyer whose bill is mostly re-sent context and who is not shopping on sticker price at all — and as the crossover table below shows, it is steep enough to put a $10.00 model under a $5.00 one.

xAI is the counter-movement. Grok 4.5’s cache read fell from $0.50 to $0.30 between our 2026-07-16 and 2026-07-28 snapshots, taking it from 75% toward the herd, and then Grok 4.6 launched on 2026-08-17 at the same $2.00 base with a $0.50 cache read — back to 75%. The newer model is strictly worse on the axis the market says it competes on.

Write fees appeared exactly when caching became automatic. Anthropic has always charged to write, and has always required you to ask for caching. OpenAI did the reverse for years: automatic caching, free writes. With GPT-5.6 it added a 1.25x write charge while keeping caching automatic and on by default above 1,024 tokens.

Those two facts interact in a way the documentation does not resolve. OpenAI’s guide says an eligible prefix “may” be cached and that written tokens bill at the write rate; it does not say whether a write is charged on a prompt that is never read again. There is an opt-out — prompt_cache_options.mode set to explicit — and the existence of an opt-out is itself informative. But the cost of failing to use it is not published, and one-shot long-prompt workloads are exactly where it would bite.

Storage is Google’s version of the same charge, moved in time. Google takes no write fee and instead bills $1.00 per 1M cached tokens per hour, rising to $4.50 on Gemini 2.5 Pro and 3.1 Pro. For an agent that keeps a 200K-token context warm all day, that is a real line item that never appears in a per-token comparison. Anthropic front-loads the same economics into the write; Google meters it. Neither is hidden, but they are not comparable without doing the arithmetic yourself — which is the point of this page.

Where the Ranking Flips

Cache hit rate does not merely scale costs down uniformly. It reorders them, because providers discount by different amounts from different starting points.

On 2026-08-03 five pairs crossed inside the 0–100% range and three of them crossed in the same 83–85% band. DeepSeek’s August rise pushed every budget-tier crossover out past 97%, and this page reported that the flip had stopped happening where agents run. That lasted four weeks. Eighteen pairs now cross. From 2026-09-28 to 2026-10-05 the lowest was a flagship pair; since 2026-10-14 it is a pair that could not be computed before:

Cheaper below the crossoverCheaper above itCrossover hit rate (jsonhouse derived)Effective price at crossover (jsonhouse derived)
Grok 4.3Mistral Medium 3.583.3%$0.3750 / 1M
Grok 4.6Claude Opus 5.587.0%$0.6957 / 1M
Gemini 2.5 Flash-LiteDeepSeek flash87.7%$0.0211 / 1M
GPT-6-lunaDeepSeek flash87.7%$0.0211 / 1M
Claude Haiku 5.5DeepSeek flash87.7%$0.0211 / 1M
Grok 4.3GPT-6.1-sol88.2%$0.3235 / 1M
Grok 4.3Claude Sonnet 5.588.2%$0.3235 / 1M
Grok 4.3Gemini 4 Argon88.2%$0.3235 / 1M
Mistral Medium 3.5GPT-6.1-sol90.9%$0.2727 / 1M
Mistral Medium 3.5Claude Sonnet 5.590.9%$0.2727 / 1M
Mistral Medium 3.5Gemini 4 Argon90.9%$0.2727 / 1M
Claude Opus 5Claude Fable 5.195.2%$0.7143 / 1M
Grok 4.6GPT-5.6-sol95.2%$0.5714 / 1M
Grok 4.5Claude Opus 5.595.2%$0.3810 / 1M
Grok 4.6Claude Fable 5.197.0%$0.5455 / 1M
GPT-5.6-solClaude Fable 5.197.6%$0.4878 / 1M
Gemini 2.5 FlashDeepSeek v4-pro97.8%$0.0359 / 1M
Grok 4.5Claude Fable 5.199.4%$0.3106 / 1M

The lowest crossover on this page is new to the page, not necessarily new to the market. Grok 4.3 × Mistral Medium 3.5 crosses at 83.3%: Grok 4.3 is $0.25 cheaper on base, but Mistral discounts 90% to Grok 4.3’s 84%. Neither price moved this collection. The pair enters the table only because Mistral’s cache price was recorded for the first time, so it may have been there all along. Read it as a crossover this page could not see before, and do not date it to 2026-10-14.

The $2.00 / 95% curve now has three vendors on it, and each one adds the same two crossovers. Claude Sonnet 5.5, GPT-6.1-sol and Gemini 4 Argon share every figure, so each passes Grok 4.3 at 88.2% and Mistral Medium 3.5 at 90.9%. Six of the seven new pairs are that one curve counted three times. Claude Haiku 5.5 adds the seventh by landing on Flash-Lite’s figures and inheriting its 87.7% crossover with DeepSeek flash.

The budget crossover came back 10 points. Flash-Lite × DeepSeek moved from 97.6% to 87.7%, because DeepSeek cut base and cache read together: the sticker gap against Flash-Lite narrowed from $0.12 to $0.05 while the discount slope stayed steep. That is the same tool working in the direction it is designed for, after four weeks of working against itself. One crossover vanished in the process — GPT-5.6-luna × DeepSeek at 60.6% is gone, because deepseek-flash now beats luna on base price and on cache read simultaneously, and a model losing on both axes cannot win at any hit rate.

The lowest crossover on this page is now a flagship pair. Grok 4.6 × Claude Opus 5.5 crosses at 87.0%, below the 87.7% budget hinge, and every flagship crossover recorded before it needed 95% or more. The mechanism is the one xAI has been running against itself for two months: Opus 5.5 discounts a cache read 95%, Grok 4.6 discounts 75%, and a 20-point gap in the discount overcomes a 2x gap in sticker price once most of the prompt is cached. A shallow cache discount has stopped being a footnote on a cheap model and become the reason it loses.

Claude Opus 5 has no crossover with Claude Opus 5.5 at all, and that absence is the finding. Opus 5.5 is cheaper on base ($4.00 against $5.00) and on cache read ($0.20 against $0.50) simultaneously, so it wins at every hit rate from 0% to 100%. A crossover needs each model to be better somewhere; here one is better everywhere. The Fable 5.1 × Opus 5 inversion this page recorded three weeks ago still holds as arithmetic, but the buyer it was aimed at now has a row that beats both without needing a 95% hit rate to get there.

GPT-6-luna arrived as an exact copy of Gemini 2.5 Flash-Lite — $0.10 base, $0.01 cache read — so it inherits Flash-Lite’s 87.7% crossover against DeepSeek flash and never crosses Flash-Lite itself. That was the first exact cross-vendor tie this page recorded, and on 2026-10-14 Claude Haiku 5.5 made it three-way at the bottom of the table, while the $2.00 / 95% curve became a three-way tie in the middle.

GPT-6.1-sol adds one crossover and wins five rows outright. Against Grok 4.3 it crosses at 88.2%: Grok 4.3 is $0.75 cheaper on base, but GPT-6.1-sol discounts 95% to Grok 4.3’s 84%, and above that hit rate the $2.00 model is the cheaper one.

Against the five other $2.00-base rows — GPT-6-sol, GPT-5.6-terra, Claude Sonnet 5, Grok 4.6 and Grok 4.5 — there is no crossover to report, because it ties them at 0% and has the lower cache read, so it is cheaper at every hit rate above zero. That is the Opus 5 × Opus 5.5 pattern again, except that this time the list price did not move at all.

On 2026-10-14 the same pattern reached Anthropic’s mid-tier on a row that already existed. Claude Sonnet 5.5 now dominates Claude Sonnet 5 exactly as GPT-6.1-sol dominates GPT-6-sol: same base, half the cache read. For a buyer who pinned Sonnet 5.5 in its first week, this is a saving that arrived without changing anything. For one still on Sonnet 5, it is a reason to switch that no list price shows.

One correction. The 2026-08-24 entry below recorded that OpenAI’s cut to GPT-5.6-sol moved no crossover, on the reasoning that nothing in the flagship band was priced where sol could cross it. That was wrong. Grok 4.6 × GPT-5.6-sol sat at exactly 100% before the cut — mathematically present, unreachable in practice — and the cut brought it to 95.2%. It has been in range for three weeks and this page missed it. The crossover set is now computed exhaustively over all pairs each week rather than by inspecting the pairs we expect to move.

The practical consequence is that sticker price is no longer sufficient at either end of the table. A plain pricing table now ranks the wrong quantity for long-running agents in the budget tier and for heavy-context work in the flagship tier, and it lands on the right answer only in the middle. That was close to the reverse of what this page reported in early September, and the reversal took two vendor changes eight days apart.

Does the Write Fee Actually Matter?

Less than it looks, if you reuse anything at all. For a prefix written once and read N−1 times, caching beats not caching at:

Write multiplierWhoBreak-even
NoneOpenAI pre-5.6, Google, and providers publishing no write feeN ≥ 2 calls
1.25xAnthropic 5-minute, OpenAI GPT-5.6+N ≥ 2 calls
2xAnthropic 1-hourN ≥ 3 calls

A worked example makes the scale concrete. A 50,000-token system prompt sent on 1,000 calls costs $100.00 of input on Claude Sonnet 5 with no caching. With the 5-minute cache and its 1.25x write, it costs $10.12 — an 89.9% saving, within a tenth of a point of the headline 90% despite paying the write premium. The write is amortized over 999 reads and disappears.

The write fee is therefore not an agent problem. It is a one-shot problem, and only on providers that cache automatically.

The Bigger Picture

Cache pricing is where the LLM market’s economics stopped being about model quality. A 90% discount on repeated input is a bet that the dominant workload is no longer a human typing a fresh question but an agent re-sending the same 50,000 tokens of instructions on a loop. Every provider has now priced for that world, which means every provider believes it.

That belief shows up in our other numbers too. The frontier is splitting between rising flagship prices and deflating everything else, and the models buyers actually pick are increasingly chosen on capability limits rather than raw benchmark scores. Caching sits underneath both: it is the mechanism that makes a large, static, expensive context economically survivable, and therefore the reason agent architectures grew context instead of trimming it.

It also changes who wins, and the last four weeks ran the mechanism in both directions on the same vendor. A provider with a mediocre sticker price and an aggressive cache discount can undercut a cheaper rival on real invoices — DeepSeek did that above an 84% hit rate until 2026-08-17, lost it when it raised its base, and recovered it on 2026-09-14 when the replacement model brought the base back down. The discount was never the asset on its own. It is leverage on the sticker price, and leverage moves with the thing it is levered on.

Claude Fable 5.1 was the same lesson from the other side. A 97.5% discount on a $10.00 base beats a 90% discount on a $5.00 base above 95.2% — enough leverage to invert two models in the same catalogue. That was a real inversion and also a narrow one, requiring a hit rate almost no workload sustains by accident.

Claude Opus 5.5 is what happens when the same instrument is aimed at a normal workload instead. A 95% discount on a $4.00 base crosses Grok 4.6’s $2.00 at 87.0%, which plenty of agents clear, and it dominates Anthropic’s own Opus 5 outright. Three weeks ago the deep-discount row was a halo product priced for a buyer who would not notice; now it is the mid-flagship row, and the thing it beats is a cheaper model from another vendor. That is the difference between a pricing experiment and a pricing strategy.

GPT-6.1-sol, a week later, shows the strategy spreading and taking a form that is harder to see. Anthropic paired its deeper multiplier with a lower sticker, so a price list at least showed something had changed. OpenAI kept the sticker identical and changed only the cache read, so a price list shows two identical rows. When the reduction lives entirely in the cache column, a buyer who never computes an effective price never finds it.

October 14 made that the norm rather than the exception. Anthropic halved Sonnet 5.5’s cache read without touching its list price, and Google launched its newest model straight onto the same 0.05x. Three vendors now sell an identical $2.00 / $10.00 sticker with a 95% cache discount behind it, beside two rows at the same sticker that discount 90%. On a price list those five rows are indistinguishable; at a 95% hit rate the first three cost a third less.

How to Choose: Recommendations by Workload

Stateless, high-volume extraction (hit rate near 0%). Rank on sticker price and ignore this page’s other columns. Gemini 2.5 Flash-Lite, GPT-6-luna and Claude Haiku 5.5 tie at $0.10, then DeepSeek flash ($0.15) and GPT-5.6-luna ($0.20). Haiku 5.5’s figure holds only below 100,000 prompt tokens.

Long-running agents with a stable system prompt (hit rate above 85%). The order inverts again, and the hinge is 87.7%. At h = 95% it is DeepSeek flash ($0.0104), then Gemini 2.5 Flash-Lite, GPT-6-luna and Claude Haiku 5.5 tied ($0.0145 each), GPT-5.6-luna ($0.0290), Gemini 2.5 Flash ($0.0435), DeepSeek v4-pro ($0.0539). Measure your hit rate before choosing: the two leaders swap across 87.7%, and this crossover has now moved three times in six weeks. Note also that the DeepSeek row has changed model twice under one id — from v4-flash to DeepSeek-V4.1-Flash on 2026-09-14, then to the V4.1-Flash-1017 build on 2026-10-17 — so a benchmark you ran before either date no longer describes what you would be buying.

Heavy-context work on Claude (hit rate above 85%). Start at Claude Opus 5.5, not Claude Opus 5. It is cheaper on base and on cache read at once, so it wins at every hit rate, and at h = 95% it costs $0.3900 per 1M effective input tokens against Opus 5’s $0.7250 — a 46% difference on a row that also carries the lower sticker price. The only reason left to run Opus 5 is a pinned revision you have already qualified.

The same now holds one tier down: Claude Sonnet 5.5 matches Sonnet 5’s base and halves its cache read, so at h = 95% it costs $0.1950 against $0.2900. Claude Haiku 5.5 is the cheapest Claude row here at $0.0145, but only while the whole prompt — cached tokens included — stays under 100,000 tokens; a heavy-context agent above that pays $0.50 base and $0.05 cache read.

Agents on a $2.00 tier. Claude Sonnet 5.5, GPT-6.1-sol and Gemini 4 Argon now share one curve, so the vendor choice at this price is a capability question, not a cost one — with the caveat that Argon’s figures double on 2026-12-01. On OpenAI, move from GPT-6-sol to GPT-6.1-sol if your workload caches at all. The list price is identical, so nothing on the invoice gets worse, and at h = 95% the effective input price falls from $0.2900 to $0.1950 — a third lower. As with any successor id, re-qualify output quality before switching; this page measures cost, not behaviour.

Flagship work above an 87% hit rate, whoever your vendor is. Compare on the effective column rather than the base price, because this is where the ranking now inverts. Claude Opus 5.5 passes Grok 4.6 at 87.0% and Grok 4.5 at 95.2% despite costing twice their base rate. If your prompts are mostly re-sent context, a $2.00 model on a 75% cache discount is more expensive than a $4.00 model on a 95% one.

One-shot long prompts on OpenAI GPT-5.6+. Measure cache_write_tokens on a representative request before assuming caching is free. If the prompt is never reused, prompt_cache_options.mode set to explicit is the documented way to suppress the implicit breakpoint.

Anything on Claude. Check the minimum cacheable length for your specific model before budgeting a saving. At 4,096 tokens on Haiku 4.5 and Opus 4.5, a prompt that caches fine on Opus 5 will silently not cache, with no error returned.

Anything on Mistral. You can now model the per-token side: Mistral publishes a cached-input price at 0.1x base on each model, and Mistral Medium 3.5 passes Grok 4.3 above 83.3%. What you still cannot model is the rest of the cache — Mistral publishes no write price, storage fee, TTL or trigger.

Limitations

  • Effective price excludes write and storage costs. Those are one-time or hourly, not per-token, and are analyzed separately above rather than folded into a single number that would hide which component moved.
  • Cache hit rate is an input, not a measurement. We publish the formula and the price inputs; we have not measured hit rates on real workloads. That is a separate experiment and a future post.
  • Four of six providers publish no TTL, so time-based cache expiry cannot be modelled for them. Real hit rates on those providers may be materially lower than configured.
  • Mistral’s cache price enters this page on 2026-10-14. Before that the row was unfilled, so any Mistral crossover, including the 83.3% one, has no earlier history here. Its write, storage and TTL terms remain unpublished.
  • Prices are list prices at the standard non-batch tier. Batch discounts and negotiated enterprise rates are excluded.
  • The effective price is per token, and a token is not a constant amount of text. Anthropic states that Claude 4.7 and later models count roughly 30% more tokens for the same input than Sonnet 4.6 and earlier, so the Claude rows here are not all denominated in the same unit. The discount percentages are unaffected — they are ratios within a single model — but comparing an effective price across that line understates the newer models.
  • Coverage is 24 models across six providers, chosen to span every tier and every disclosure pattern rather than to be exhaustive.
  • Three Google rows, one Anthropic row and both DeepSeek rows are conditional prices. Gemini 3.6 and 3.8 Flash are promotional through 2026-12-31 and Gemini 4 Argon through 2026-11-30, each doubling after; Claude Haiku 5.5’s figures apply only to prompts up to 100,000 tokens; DeepSeek’s figures are off-peak and bill double during weekday peak UTC windows. Both effective columns and crossovers move with them.
  • The DeepSeek flash row changed model, not just price, on 2026-09-14, and changed model again at an unchanged price on 2026-10-17. Its figures are comparable to the previous v4-flash row as a line item on an invoice, but not as a measurement of the same product. Week-over-week deltas on that row should be read as a substitution, and a flat week on it does not mean the same model.

Update Cadence and Changelog

This page is refreshed weekly from the same snapshot series that backs our pricing table, and every change lands below as a dated entry. Caching mechanics — write fees, TTLs, minimum lengths — are re-read from provider documentation monthly, since they change far more slowly than prices.

DateChange
2026-10-19No change to any figure; one row changed model. Every base and cache-read price behind this table holds the value collected on 2026-10-14, so all four effective-price columns, every discount figure and the eighteen crossovers are unchanged because every input to the arithmetic is identical. The table stays at 24 models. On 2026-10-17 DeepSeek replaced the build behind deepseek-flash with DeepSeek-V4.1-Flash-1017, stating that pricing is unchanged, so the DeepSeek flash row keeps its $0.0104 at h = 95% and its 87.7% crossover against Flash-Lite, GPT-6-luna and Claude Haiku 5.5 while describing a different model. Collected Monday 2026-10-19, five days after the Wednesday 2026-10-14 collection. Caching mechanics were not re-read and are unchanged from 2026-08-03.
2026-10-14An existing row’s cache read was halved, Mistral became computable, and the crossover set went from eleven pairs to eighteen. Collected Wednesday 2026-10-14; the Monday 2026-10-12 slot was missed and is left as a gap. Claude Sonnet 5.5’s cache read fell from $0.20 to $0.10 (0.05x), giving it a row of its own at $1.0500 / $0.4800 / $0.1950 for h = 50% / 80% / 95% — the first base or cache-read change on an existing row here since 2026-08-24. Added: Gemini 4 Argon at $2.00 / $0.10, Google’s first row below 0.1x, introductory through 2026-11-30; Claude Haiku 5.5 at $0.10 / $0.01 for prompts up to 100,000 tokens ($0.50 / $0.05 above). Removed: Gemini 3.7 Flash, deprecated 2026-10-08 and routed by Google to Gemini 3.8 Flash. Mistral Medium 3.5 is filled for the first time at $0.15 cache read (90%), from a per-model column this series had not captured before; that is a newly recorded price, not a changed one. The table goes from 22 models to 24. New crossovers, recomputed exhaustively: Grok 4.3 × Mistral Medium 3.5 at 83.3%, now the lowest on the page; Grok 4.3 × Claude Sonnet 5.5 and × Gemini 4 Argon at 88.2%; Mistral Medium 3.5 × GPT-6.1-sol, × Claude Sonnet 5.5 and × Gemini 4 Argon at 90.9%; and Claude Haiku 5.5 × DeepSeek flash at 87.7%. Claude Sonnet 5.5 now dominates Claude Sonnet 5 with no crossover, as GPT-6.1-sol dominates GPT-6-sol. Caching mechanics were not re-read and are unchanged from 2026-08-03.
2026-10-05One row arrived at an unchanged list price, and OpenAI left the flat 90%. Every base and cache-read price already in this table holds the value collected on 2026-09-28, so no existing effective-price column, discount figure or crossover changed. The table goes from 21 models to 22. Added: GPT-6.1-sol at $2.00 base / $0.10 cache read — GPT-6-sol’s list price with the cache read halved, a 95% discount on a 0.05x multiplier and the first OpenAI row off 0.1x — with effective prices of $1.0500 at h = 50%, $0.4800 at h = 80% and $0.1950 at h = 95%. Claude Sonnet 5.5 also arrived, on Claude Sonnet 5’s base and cache read to the cent, and is carried in the snapshot without a row here. The crossover set goes from ten pairs to eleven, recomputed exhaustively. New: Grok 4.3 × GPT-6.1-sol at 88.2%. GPT-6.1-sol has no crossover with the five other $2.00-base rows (GPT-6-sol, GPT-5.6-terra, Claude Sonnet 5, Grok 4.6, Grok 4.5) because it ties them at 0% and is cheaper at every hit rate above it. DeepSeek’s page now states that its peak windows apply on weekdays only, with weekends and Chinese public holidays off-peak in full; recorded for the first time this week, with no effect on the off-peak figures used here. Caching mechanics were not re-read and are unchanged from 2026-08-03.
2026-09-28Three rows arrived, and the flagship band produced its first crossover inside the agent band. Every base and cache-read price already in this table holds the value collected on 2026-09-21, so no existing effective-price column and no existing discount figure changed. The table goes from 18 models to 21. Added: Claude Opus 5.5 at $4.00 base / $0.20 cache read — a 95% discount on a 0.05x multiplier, Anthropic’s third multiplier alongside Fable 5.1’s 0.025x and the 0.1x the rest of the line carries — with effective prices of $2.1000 at h = 50%, $0.9600 at h = 80% and $0.3900 at h = 95%. Added: GPT-6-sol at $2.00 / $0.20 and GPT-6-luna at $0.10 / $0.01, which land exactly on GPT-5.6-terra’s and Gemini 2.5 Flash-Lite’s figures, so each adds a row without adding a distinct curve. Grok 4.7 also arrived, on Grok 4.6’s base and cache read to the cent, and is carried in the snapshot without a row here. The crossover set goes from seven pairs to ten. New: Grok 4.6 × Claude Opus 5.5 at 87.0%, the lowest crossover on this page and the first flagship pair inside the range long-running agents operate in; Grok 4.5 × Claude Opus 5.5 at 95.2%; and GPT-6-luna × DeepSeek flash at 87.7%, inherited from Flash-Lite’s identical figures. Claude Opus 5.5 has no crossover with Claude Opus 5, because it is cheaper on base and on cache read simultaneously — the first row this page has recorded as dominated inside a single vendor’s catalogue. Caching mechanics were not re-read and are unchanged from 2026-08-03.
2026-09-21No change, and this time nothing arrived either. Every base and cache-read price behind this table holds the value collected on 2026-09-14, so all four effective-price columns, all eighteen discount figures and the full crossover set are unchanged — not re-derived to the same answer, but unchanged because every input to the arithmetic is identical. The table stays at 18 models. Flash-Lite × DeepSeek remains at 87.7% and Claude Opus 5 × Claude Fable 5.1 at 95.2%. One caveat was added to the Limitations section rather than the table: Anthropic’s tokenizer split means the Claude rows are not all denominated in the same unit of text, which leaves the discount percentages intact (they are ratios inside one model) but makes effective prices understate Claude 4.7-and-later rows against Sonnet 4.6 and earlier. Caching mechanics were not re-read and are unchanged from 2026-08-03.
2026-09-14Both ends of the table left the 90% consensus, and a crossover returned to the agent band. DeepSeek retired v4-flash and now serves that name from deepseek-flash (build DeepSeek-V4.1-Flash) at $0.15 base / $0.003 cache read, a 98% discount — the row is a substitution, not a repricing, and its figures describe a different model than last week’s. Effective prices on it fall to $0.0765 at h = 50%, $0.0324 at h = 80%, $0.0104 at h = 95%, which takes back the h = 95% lead from Gemini 2.5 Flash-Lite. Flash-Lite × DeepSeek moved from 97.6% to 87.7%, back inside the range where long-running agents operate, and GPT-5.6-luna × DeepSeek (60.6%) vanished because the new row beats luna on base and cache read at once. Added two rows missed in the 2026-09-07 collection, which this page skipped: Claude Fable 5.1 ($10.00 / $0.25, a 97.5% discount — the first Anthropic row not on the flat 0.1x multiplier) and Gemini 3.8 Flash ($0.75 / $0.075), taking the table from 16 models to 18. Fable 5.1 opens four crossovers at the top of the price list, including Claude Opus 5 × Claude Fable 5.1 at 95.2%, where a $10.00 model becomes cheaper per effective input token than a $5.00 one. Correction: the 2026-08-24 entry stated that no crossover moved; Grok 4.6 × GPT-5.6-sol in fact fell from exactly 100% to 95.2% on that cut and was missed for three weeks. Crossovers are now computed exhaustively across all pairs rather than by inspecting expected movers. Caching mechanics were not re-read and are unchanged from 2026-08-03.
2026-09-07Not collected. This page skipped its weekly refresh; the two models that arrived that week (Claude Fable 5.1, Gemini 3.8 Flash) were folded into the 2026-09-14 entry instead. No base or cache-read price moved at any of the six providers during that week, so no effective-price column in the 2026-09-01 table was wrong while the gap stood — but the table was two rows short of complete, and Fable 5.1’s 97.5% discount was the largest single change to the discount picture since this page began. The miss is recorded rather than backfilled.
2026-09-01No change. Every base and cache-read rate behind this table holds the value recorded on 2026-08-24, so all four effective-price columns and all sixteen discount figures are unchanged. This is the first collection in the series with no price move at any of the six providers. Collected Tuesday 2026-09-01; the 2026-08-31 Monday slot was missed and left as a gap rather than backfilled. The 2027-01-01 doubling on both Gemini Flash rows still stands and will move every Google Flash figure here when it lands.
2026-08-24GPT-5.6-sol cut to $4.00 base / $0.40 cache read (from $5.00 / $0.50), lowering every effective column on that row by 20% — $2.2000 at h = 50%, $1.1200 at h = 80%, $0.5800 at h = 95%. The discount stays 90%, because OpenAI cut base and cache read by the same proportion, so the row moves down without changing shape. No crossover moved: sol sits in the flagship band where nothing it could cross is priced, and the three surviving crossovers are all between budget models. The h = 95% leader is still Gemini 2.5 Flash-Lite. OpenAI marks the new rate promotional through at least 2026-11-21, so this row now carries an expiry the others do not. No other price changed; the table stays at 16 models. Caching mechanics were not re-read and are unchanged from 2026-08-03.
2026-08-17The 83–85% crossover band disappeared. DeepSeek raised v4-flash to $0.22 base / $0.007 cache read (from $0.14 / $0.0028) and v4-pro to $0.66 / $0.022 (from $0.435 / $0.003625), taking its discount from 98–99% to 96.7–96.8%. Google halved Gemini 3.6 Flash to $0.75 / $0.075. Of the five crossovers, three vanished — Grok 4.3 × Gemini 3.6 Flash (Gemini now wins on both axes) and both DeepSeek v4-pro pairs (pushed past a 100% hit rate) — and the surviving Flash-Lite × v4-flash moved from 84.7% to 97.6%. One is new: GPT-5.6-luna × DeepSeek v4-flash at 60.6%. The h = 95% leader changed hands from DeepSeek v4-flash to Gemini 2.5 Flash-Lite. Added Gemini 3.7 Flash and Grok 4.6; the table is now 16 models. Caching mechanics were not re-read and are unchanged from 2026-08-03.
2026-08-03OpenAI cut GPT-5.6-luna to $0.20 base / $0.02 cache read (−80%) and GPT-5.6-terra to $2.00 / $0.20 (−20%). Both keep a 90% discount, so every effective column scaled without changing the discount picture. One crossover changed hands: Grok 4.5 x GPT-5.6-terra (90.9%) is gone and GPT-5.6-luna x DeepSeek v4-pro (93.5%) is new. The three crossovers inside the 83–85% band are untouched. Also corrected the crossover section, which described three pairs as the total when three was the count inside the band.
2026-07-28Initial edition. 14 models, six providers. Records xAI’s Grok 4.5 cache read moving $0.50 to $0.30 and OpenAI’s 1.25x write charge on the GPT-5.6 family.

FAQ: LLM Cache Pricing 2026

How much does prompt caching actually save in 2026?

The cached-read discount converged on 90% at OpenAI, Google, and most of Anthropic’s line, and that consensus has been coming apart since September. Anthropic now runs three multipliers: 97.5% on Claude Fable 5.1 and Mythos 5.1, 95% on Claude Opus 5.5 and, since October 14, Sonnet 5.5, and 90% on everything else in its line. OpenAI joined on October 5 with 95% on GPT-6.1-sol, and Google on October 14 with 95% on Gemini 4 Argon, both keeping 90% on their other rows. DeepSeek discounts 98% on deepseek-flash, Mistral 90% across its line, and xAI 75–85%.

That discount applies only to the cached share of a prompt, so your realized saving is roughly the hit rate times the discount. A 50% hit rate against a 90% discount is a 45% cut to input cost, not 90%.

Do you pay to write to the cache?

At Anthropic, always: 1.25x base input for the 5-minute cache, 2x for the 1-hour. OpenAI introduced the same 1.25x charge with the GPT-5.6 family, having charged nothing on earlier models. Google levies no write fee but bills storage at $1.00–$4.50 per 1M cached tokens per hour. xAI, DeepSeek, and Mistral publish no write price.

Which LLM API is cheapest for a cache-heavy agent?

It depends on your hit rate again, which it did not four weeks ago. Gemini 2.5 Flash-Lite is cheapest from 0% through 87.7% — tied there by GPT-6-luna since September 28 and by Claude Haiku 5.5 since October 14, all on identical base and cache-read prices — and above 87.7% DeepSeek flash takes over, reaching $0.0104 per 1M effective input tokens at a 95% hit rate against the $0.0145 those three share. That crossover has now moved three times in six weeks — 84.7%, then 97.6% after DeepSeek’s August 17 rise, then 87.7% once it retired v4-flash for a cheaper model on September 14 — so inherit no answer here without measuring your own hit rate first.

Can a more expensive model be cheaper for a cache-heavy agent?

Yes, and since September 28, 2026 it happens at a hit rate agents actually reach. Claude Opus 5.5 costs $4.00 per 1M input tokens against Grok 4.6’s $2.00, but discounts a cache read 95% where xAI discounts 75%. Above an 87.0% hit rate the $4.00 model is the cheaper one per effective input token.

Every earlier crossover of this kind sat at 95% or higher, which is a rate few workloads hold. This one does not require anything unusual, which is why the flagship comparison can no longer be done on base prices alone.

Is a newer model at the same list price ever cheaper for cached work?

Yes, and there are now two such pairs. GPT-6.1-sol, from October 5, 2026, carries GPT-6-sol’s $2.00 per 1M input tokens but bills a cache read at $0.10 instead of $0.20. Claude Sonnet 5.5 reached the same position against Claude Sonnet 5 on October 14, when its cache read was halved. At a 0% hit rate each pair costs exactly the same; at any higher hit rate the newer model is cheaper, reaching $0.1950 per 1M effective input tokens at 95% against $0.2900.

A list-price comparison shows no difference between the two rows at all. The saving appears only once you weight the cache-read column by your own hit rate, which is the calculation this page exists to do.

Is prompt caching automatic or do I have to enable it?

OpenAI, Google, DeepSeek, and xAI cache automatically; Anthropic requires an explicit cache_control field. That difference now costs money: because GPT-5.6 caches automatically above 1,024 tokens and bills writes at 1.25x, a long prompt you never reuse may be billed above base rate unless you set prompt_cache_options.mode to explicit.

Why can’t I compare cache prices straight from provider pricing pages?

No two providers publish the same fields. Anthropic gives multipliers, DeepSeek an absolute cache-hit price, xAI and Mistral one cached-input column each, and Google a per-hour storage rate on top of a cache-read price. Mistral’s per-model column entered this page only on October 14; before that, a headline −90% was all this series had recorded. Four of six publish no TTL. Normalizing those into one unit is the reason this page exists.

Primary Sources

Data verified 2026-10-19. Machine-readable dataset: /data/llm-cache-pricing-2026.json

ai-data-statistics
prompt-caching llm-pricing api-cost claude gpt-5 gemini deepseek 2026