Best LLM 2026: Capability and Limits Compared
Best LLM 2026 compared on the limits that actually block builds: context window, max output tokens, prompt cache minimums, knowledge cutoffs, and retirement dates.
The frontier models of 2026 have stopped differing in the ways benchmarks measure. Context windows have converged inside a 5% band across five providers. What has not converged is the operational envelope — max output tokens still spans 32,000 to 384,000, the minimum prompt length required for caching to work at all ranges from 512 to 4,096 tokens, and the notice a provider gives before deleting a model ranges from two weeks to six months. These are the numbers that decide whether a workload runs, what it costs, and how long your integration survives. They are published, but only one model at a time, on six different documentation sites. This page normalizes them into one comparison, collected on 2026-07-25 from official provider documentation only.
TL;DR
- Context window is settled: 1,000,000 to 1,050,000 tokens across Anthropic, OpenAI, Google, xAI, and DeepSeek — a spread under 5%. Stop choosing on it.
- Max output is not settled: DeepSeek v4 publishes 384,000 tokens, Anthropic and OpenAI 128,000, Google’s Gemini rows 65,536, and Claude Haiku 4.5 64,000 — a six-fold spread. This is now the binding constraint on generation workloads.
- Prompt caching has an undocumented floor: below the minimum cacheable prefix, caching silently does nothing — you pay the write premium and never get a read. Thresholds are non-monotonic even within one vendor.
- Vision is now universal except one row: every listed Anthropic, OpenAI and Google model accepts image input, and so does
deepseek-flashsince it replaced the text-only v4-flash on September 14, 2026. DeepSeek v4-pro remains text-only, and xAI has stopped stating image support per model on its models page. - Model lifetime is the unpriced risk: Anthropic is the only provider publishing retirement floors for models that are still active. OpenAI gives previews as little as two weeks.
Raw data: data/best-llm-2026.json — machine-readable structured data for AI crawlers and citation.
Methodology
Every figure on this page was collected on 2026-07-25 from first-party provider documentation only: Anthropic, OpenAI, Google, xAI, DeepSeek, and Mistral. No aggregator, leaderboard, or third-party spec sheet was used as a source at any point.
Rows collected later carry their own date in the table footnote. The two DeepSeek rows were re-verified on 2026-09-21 after DeepSeek retired v4-flash, and the deepseek-flash row again on 2026-10-19 after DeepSeek swapped the build behind it; the Anthropic and OpenAI rows were refreshed on 2026-09-28 and 2026-10-05; and the three Google rows and Claude Haiku 5.5 were collected on 2026-10-14, after Google withdrew the Gemini 3.5 Flash row this page had used since July.
Units are normalized as follows. Context window means the maximum input tokens accepted in a single request, reported as the provider states it — we do not round 1,048,576 to “1M”, because the exact figure is what your token budget must respect. Max output means the ceiling on tokens generated in one synchronous response, excluding batch-only or beta-header extensions, which are noted separately. Knowledge cutoff is reported per provider label; where a provider publishes two distinct dates we say which one we used.
Where a provider does not publish a figure, the cell reads “Not published” rather than being estimated, left blank, or filled from a secondary source. Those gaps are findings in their own right and are discussed under Limitations.
The Operational Envelope, Normalized
The table below covers the current flagship and mid-tier models from each provider. Every value is as published on the collection date.
| Model | Provider | Context (input tokens) | Max output | Image input | Knowledge cutoff |
|---|---|---|---|---|---|
| Claude Fable 5.1* | Anthropic | 1,000,000 | 128,000 | Yes | Jun 2026 |
| Claude Opus 5.5* | Anthropic | 1,000,000 | 128,000 | Yes | Jun 2026 |
| Claude Opus 5 | Anthropic | 1,000,000 | 128,000 | Yes | May 2026 |
| Claude Fable 5 | Anthropic | 1,000,000 | 128,000 | Yes | Jan 2026 |
| Claude Sonnet 5.5* | Anthropic | 1,000,000 | 128,000 | Yes | Jun 2026 |
| Claude Sonnet 5 | Anthropic | 1,000,000 | 128,000 | Yes | Jan 2026 |
| Claude Haiku 5.5‡ | Anthropic | 1,000,000 | 128,000 | Yes | Jun 2026 |
| Claude Haiku 4.5 | Anthropic | 200,000 | 64,000 | Yes | Feb 2025 |
| GPT-6-astra* | OpenAI | 1,050,000 | 128,000 | Yes | Apr 30, 2026 |
| GPT-6.1-sol* | OpenAI | 1,050,000 | 128,000 | Yes | Apr 30, 2026 |
| GPT-6-sol* | OpenAI | 1,050,000 | 128,000 | Yes | Apr 20, 2026 |
| GPT-6-luna* | OpenAI | 1,050,000 | 128,000 | Yes | May 18, 2026 |
| GPT-5.6-sol† | OpenAI | 1,050,000 | 128,000 | Yes | Feb 16, 2026 |
| GPT-5.5† | OpenAI | 1,050,000 | 128,000 | Yes | Dec 1, 2025 |
| GPT-5.4† | OpenAI | 1,050,000 | 128,000 | Yes | Aug 31, 2025 |
| Gemini 4 Argon‡ | 1,048,576 (shared with output) | 65,536 | Yes, plus video, audio, PDF | Not published | |
| Gemini 3.8 Flash‡ | 1,048,576 | 65,536 | Yes, plus video, audio, PDF | Not published | |
| Gemini 3.6 Flash‡ | 1,048,576 | 65,536 | Yes, plus video, audio, PDF | Not published | |
| Grok 4.7* | xAI | 500,000 | Not published | Not published | May 2026 |
| Grok 4.3 | xAI | 1,000,000 | Not published | Yes | Not published |
| Grok 4.5 | xAI | 500,000 | Not published | Yes | Feb 1, 2026 |
| DeepSeek v4-pro | DeepSeek | 1,000,000 | 384,000 | No | Not published |
| deepseek-flash | DeepSeek | 1,000,000 | 384,000 | Yes | Not published |
*Rows marked with an asterisk were collected on 2026-09-28, except Claude Sonnet 5.5 and GPT-6.1-sol, which were collected on 2026-10-05. ‡Rows marked with a double dagger were collected on 2026-10-14. Every unmarked row except the two DeepSeek ones is still the 2026-07-25 collection. †The three GPT-5.x rows carry their July figures because OpenAI’s models page no longer publishes a spec page for them — the specs did not change, the documentation did, and we have kept the last figures we read rather than blanking rows a vendor stopped describing.
Gemini 3.5 Flash has been removed. Google deprecated it on 2026-10-08 and took it off its pricing page and AI Studio, but did not shut the endpoint: requests to gemini-3.5-flash are now routed to gemini-3.6-flash. The three Google rows that replace it were collected from Google’s model pages on 2026-10-14. Those pages give no knowledge cutoff for any of them — where the July page for Gemini 3.5 Flash gave May 2026 — so the column says so rather than carrying the old date across.
Claude Opus 4.1 has been removed: Anthropic retired it on 2026-08-05, and it was the row this page cited as the max-output floor at 32,000 tokens. That floor is now Claude Haiku 4.5 at 64,000, which halves the spread across this table from twelve-fold to six-fold.
Three footnotes materially change how the remaining rows read. Anthropic’s 128,000 output ceiling rises to 300,000 on the Message Batches API for Opus 5.5, Opus 5, Sonnet 5.5, Sonnet 5, Haiku 5.5, Opus 4.8, Opus 4.7, Opus 4.6, and Sonnet 4.6 behind a beta header — so the synchronous number understates the asynchronous ceiling by more than double. Google’s Gemini 4 Argon has its own extension: 65,536 output tokens by default, up to 1,048,576 with long decode.
Anthropic’s newer models (Opus 4.7 and later, Sonnet 5 and 5.5, Haiku 5.5, Fable 5) also use a tokenizer that emits roughly 30% more tokens for identical text, which means a 1,000,000-token window on those models holds meaningfully less prose than a 1,000,000-token window elsewhere. Context windows denominated in tokens are not directly comparable when the tokenizers differ.
Gemini 4 Argon adds a third reason. Google states its 1,048,576 figure as a unified input-and-output limit, where the Gemini Flash pages state an input limit and a separate output limit. On Argon a request that asks for a long output has correspondingly less room for input, so the same number in the same column describes a smaller input budget than it does one row down.
One cell in the new rows is deliberately empty. xAI’s models page states Grok 4.7’s context window and a May 2026 knowledge cutoff but does not state image support per model, where the same page did for Grok 4.5 and 4.3 in July. Reading the older rows’ “Yes” across to the newer model would be an inference, so the cell says what the page says. A vendor removing a capability statement is not the same as a vendor removing a capability, and this page cannot tell you which happened.
The Cache Cliff Nobody Publishes Side by Side
Every major provider advertises a prompt caching discount of roughly 90%. What none of them advertise prominently is the minimum cacheable prefix — the length below which a cache marker is accepted and then never produces a single read. Anthropic states the failure mode explicitly: requests to cache fewer than the minimum “will be processed without caching, and no error is returned.” There is no warning, no exception, and no log line. The only signal is in the response itself. What that silent failure costs you in money, once write fees and hit rates are folded in, is worked out in LLM Cache Pricing 2026.
| Model or family | Minimum cacheable prefix | Mechanism |
|---|---|---|
| Claude Fable 5.1, Opus 5.5, Opus 5, Sonnet 5.5, Fable 5, Haiku 5.5 | 512 tokens | Explicit breakpoint |
| Claude Opus 4.8, Sonnet 5, Sonnet 4.6, Sonnet 4.5 | 1,024 tokens | Explicit breakpoint |
| Claude Opus 4.7 | 2,048 tokens | Explicit breakpoint |
| Claude Opus 4.6, Opus 4.5, Haiku 4.5 | 4,096 tokens | Explicit breakpoint |
| Gemini 2.5 Flash, Gemini 2.5 Pro | 2,048 tokens | Implicit caching |
| Gemini 4 Argon, 3.8 Flash, 3.6 Flash, 3.1 Pro Preview | 4,096 tokens | Implicit caching (Argon: implicit only) |
| OpenAI GPT-5.x family | Not published | Supported |
| xAI Grok family | Not published | Supported, priced per cached token |
| DeepSeek v4 | Not published | Supported, cache-hit pricing |
There is exactly one way to find out whether caching engaged: read the usage fields on the response. If both the cache-creation and cache-read token counts come back as zero across repeated requests with an identical prefix, the prompt is below the minimum. Check this once per model you deploy, because the threshold is not a property of your code — it is a property of the model string you passed.
The striking property here is that the threshold is not monotonic within a single vendor. Claude Opus 5 caches at 512 tokens; the older Opus 4.6 requires eight times as much. Claude Haiku 5.5 caches at 512 where Haiku 4.5 needs 4,096 — the same eight-fold gap, inside the budget line. Gemini runs the other way: 2.5 Flash caches at 2,048, and every Gemini 3.x and 4 model on Google’s list requires 4,096. A team that tuned a 3,000-token system prompt against Opus 5 and then pinned an older model for cost reasons would lose caching entirely, silently, and see their bill rise rather than fall — the exact opposite of the intended change.
This is the single most consequential number on this page that has no cross-provider reference anywhere. It is published, per model, buried in caching documentation, and it moves between generations.
Why the Comparison Moved to the Envelope
For three years, “which model is best” was answered with benchmark scores, and that answer was defensible because capability differences were large and measurable. In 2026 it is not. The convergence in the first table is not a coincidence of this quarter — it is what happens when a capability becomes table stakes. Once every serious provider ships a million-token window, the window stops being a reason to choose anyone.
What replaced it is less visible because it is asymmetric. Providers converge on the specs they market and diverge on the specs they merely document. Context window is marketed, so it converged. Max output is documented, so it did not: the same table shows a six-fold spread from Claude Haiku 4.5’s 64,000 to DeepSeek v4’s 384,000. Cache minimums are documented in a subsection of a subsection, so they diverge not just between vendors but between generations of the same vendor.
There is a second, quieter reason the envelope matters more now. As context windows grew, the workloads people run against them changed shape. A 1M-token window invites you to feed in an entire codebase or a book — and the natural next request is to get a substantial artifact back. That is an output problem. DeepSeek’s 384,000-token ceiling is not a vanity number; it is the only default synchronous ceiling that lets a single request return a document proportionate to a million-token input. Meanwhile Google’s Gemini Flash rows pair a 1,048,576-token input with a 65,536-token output — a 16:1 ratio that is fine for extraction and summarization and structurally wrong for generation.
Gemini 4 Argon, added 2026-10-14, is Google’s answer and shows where the output limit is moving. Its default is the same 65,536, but long decode raises it to 1,048,576 — the highest output figure any provider in this table publishes. The catch is the unified window: a million-token output leaves almost nothing for input, so Argon’s long decode suits generating from a short brief, not transforming a long document.
The same logic applied to DeepSeek’s text-only limitation until September 14, 2026, and the way it stopped applying is worth more than the limitation was. A pipeline touching screenshots, scanned documents or charts could not use DeepSeek at any price; the cheapest model that cannot do the task has an infinite effective cost, which is why an operational-envelope comparison belongs before a price comparison and not after it.
Then DeepSeek retired v4-flash and served the same name from a model that accepts images. No announcement of a capability launch, no new model id on the price list — the constraint simply stopped existing inside a row that had not changed its name. A reader who shortlisted on this table in July and did not re-check would still be excluding DeepSeek from image work today, on a limitation that now applies only to v4-pro. That is the failure mode an envelope comparison has that a price comparison does not: a price you quote stale is wrong by a percentage, and a capability you quote stale is wrong by a yes or no.
Cost Bands, and Where the Exact Numbers Live
Absolute per-token prices belong in one place and one place only, because two pages on the same site showing different prices for the same model is worse than not publishing prices twice. Exact figures, updated weekly against all six official pricing pages, live in LLM API Pricing 2026. What this page provides instead is a band, which is stable at monthly resolution and sufficient for shortlisting.
| Band | Representative models | What it buys |
|---|---|---|
| Budget | deepseek-flash, Gemini 2.5 Flash-Lite, GPT-6-luna, Claude Haiku 5.5, Gemini 3.8 Flash, Claude Haiku 4.5 | High-volume classification, extraction, routing |
| Mid | Claude Sonnet 5.5, Claude Sonnet 5, GPT-6.1-sol, GPT-6-sol, Gemini 4 Argon, GPT-5.4-class | Most production workloads; the default tier |
| Flagship | Claude Opus 5.5, Claude Fable 5.1, GPT-6-astra, GPT-5.6-sol, GPT-5.5-pro | Long-horizon agentic work, hardest reasoning |
Band assignments derive from the 2026-10-14 weekly pricing snapshot. They are deliberately coarse: a band does not move week to week, whereas a sticker price does.
One line this section used to carry has been withdrawn rather than updated. It warned that Claude Sonnet 5’s introductory rate would expire on August 31, 2026 and rise 50%, and told you to budget against the September figure. Anthropic cancelled that increase on 2026-08-17 and the $2 / $10 rate is now the standard price. The warning was correct when written and wrong for six weeks afterwards, which is the failure mode a monthly page inherits from weekly facts — and the reason the band table names bands instead of numbers.
The bands themselves now hide something worth naming. GPT-6-sol sits in Mid, not Flagship, because it is priced at $2.00 / $10.00 — a current-generation frontier model on the same rate as Claude Sonnet 5. A tier that used to mean “the cheap tier of good models” now contains a model from the top of its vendor’s lineup. Its successor GPT-6.1-sol, added 2026-10-05, holds the same band at the same list price; the weekly pricing table records the one figure that separates them, a cache read cut in half.
Gemini 4 Argon, which Google calls its most capable model, joined Mid on 2026-10-14 for the same reason: $2.00 / $10.00. It will not stay there. Google’s page already prints $4.00 / $20.00 from December 1, which puts it in Flagship at the next band review. A band derived from an introductory price is a band with an expiry date, and this is the first row in the table to carry one.
Model Lifetime: The Risk That Has No Number
An integration’s real dependency is not a model’s capability but its continued existence. This is the least standardized dimension in the entire comparison, and the differences are structural rather than cosmetic.
| Provider | Minimum retirement notice | Publishes dates for still-active models | Nearest confirmed retirement |
|---|---|---|---|
| Anthropic | At least 60 days | Yes — a “not sooner than” date per active model | claude-sonnet-4-5-20250929, Nov 30, 2026 (deprecated Sep 30, 2026) |
| OpenAI | 6 months (GA), 3 months (specialized), ~2 weeks (preview) | No | tts-1, tts-1-hd and gpt-4o-mini-tts, Jan 6, 2027; nearest text models GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini, Jan 14, 2027 |
| Not stated | No | gemini-3.1-flash-tts-preview, gemini-3.1-flash-live-preview and three Gemini 2.5 audio and TTS previews, Nov 17, 2026 | |
| xAI | Not published | No | Not published |
| DeepSeek | Not published | No | Not published |
| Mistral | Not verified — deprecation page unreachable on 2026-07-25 | Unknown | Unknown |
Read the two middle columns together, because they pull in opposite directions. OpenAI’s headline commitment is the most generous on paper — six months beats sixty days by a factor of four. But that clock only starts when a deprecation is announced, and OpenAI does not publish anything for a model that has not been deprecated yet. Anthropic’s commitment is shorter and its disclosure is longer: every active Claude model carries a published floor, so claude-opus-5-5 is documented as safe until at least September 22, 2027 and claude-opus-5 until July 24, 2027, without any announcement having been made. One provider tells you how much warning you will get; the other tells you the earliest date you could possibly need it.
Anthropic’s table also shows what that disclosure is worth once a date arrives. claude-opus-4-1 moved from a published floor to Deprecated on 2026-06-05 to Retired on 2026-08-05, and the sixty days between those two states are the only window a team had to migrate.
It has now happened a second time, and the timing shows how a floor behaves. claude-sonnet-4-5-20250929 carried a floor of September 29, 2026. Anthropic deprecated it the next day, on September 30, and set retirement for November 30 — 61 days, one more than the minimum — with claude-sonnet-5-5 named as the replacement. The floor was not a forecast of when the model would go; it was the first day the countdown was allowed to start, and Anthropic started it at once.
The next floor on the list was claude-haiku-4-5-20251001, at October 15, 2026, followed by claude-opus-4-5-20251101 at November 24. Haiku 4.5’s floor has now passed, and the Sonnet 4.5 pattern did not repeat. Sonnet 4.5 was deprecated the day after its floor; on October 26, eleven days after Haiku 4.5’s, Anthropic’s deprecations page still lists it as Active with no notice. Under the sixty-day minimum it therefore cannot retire before December 25, 2026, and that date moves a day later for every day without a notice. That is our arithmetic from Anthropic’s two published commitments, not a date Anthropic has published.
Its likely successor arrived on October 14: claude-haiku-5-5, with a floor of October 7, 2027, though Anthropic has not named it as a replacement. The lesson for anyone reading floors is that a passed floor is permission, not a schedule. Haiku 4.5 can now be deprecated on any day, and the only safe planning assumption is sixty days from whichever day that turns out to be.
Google added a third pattern on 2026-10-08, and it does not fit either column. It deprecated gemini-3.7-flash and gemini-3.5-flash with no shutdown date at all: the deprecations page says the endpoints are not shutting down and automatically route requests to gemini-3.8-flash and gemini-3.6-flash. Nothing breaks, so nothing alerts. A team that pinned gemini-3.5-flash is now running a different model under the old name, which is the DeepSeek substitution of September 14 reached by a different route — and a reason to log the model version a response reports, not only the one you requested.
The preview tier is where this gets sharp. OpenAI states that preview models “may be retired with much shorter notice, such as 2 weeks.” Google’s gemini-3.1-flash-lite-preview was retired on May 25, 2026. gpt-5-chat-latest shut down on July 23, 2026 — two days before this data was collected. Anything with preview in its name is a different risk class from a stable model, and none of these providers price that difference.
This is also why we snapshot the table monthly rather than linking out. Retirement announcements are removed from documentation once the retirement completes; the record of what was promised and when does not survive at the source. The changelog at the bottom of this page is that record.
How to Choose, by Constraint
Work down the constraints in this order, because each one eliminates candidates the next cannot recover.
Does the model accept your input at all? If your pipeline touches images, DeepSeek v4-pro is out regardless of price, though deepseek-flash has accepted image input since September 14, 2026. On Grok 4.7, test rather than assume: xAI’s page no longer states image support per model. If it touches video, audio, or PDFs natively, the three Gemini rows are the only models in this table that handle all three without a preprocessing step.
Can it produce an output the size you need? Compare your longest expected artifact against the max output column, not the context window. Below roughly 64,000 tokens every model here qualifies. Above 128,000, only DeepSeek v4 qualifies synchronously by default. Anthropic’s batch API extension reaches 300,000, and Gemini 4 Argon’s long decode reaches 1,048,576 at the cost of input room.
Will caching actually engage? Measure your stable prefix — system prompt plus tool definitions — in tokens, then check it against the cache minimum table. If it sits between 512 and 4,096 tokens you are in the zone where the model choice determines whether caching works at all. For subscription-tier equivalents of this same tradeoff, see Best LLM Subscription 2026.
How long do you need the integration to survive? For anything you will not revisit within a year, prefer a provider that publishes a forward-looking floor, and avoid preview-tier models entirely.
Limitations
Coverage is uneven by provider, and that unevenness is the finding rather than an omission. xAI publishes context windows and cached-token pricing but no max output figure for any Grok model. DeepSeek publishes no knowledge cutoff. OpenAI’s disclosure moved between our July and September collections: it published full specs for GPT-5.6-sol, GPT-5.5 and GPT-5.4 in July and now publishes them only for the GPT-6 family, leaving the 5.x rows priced on the pricing page with no spec page behind them.
We kept the July figures and marked them, because a withdrawn document is not a changed specification — but it does mean those three rows can no longer be re-verified at the source. Mistral’s model overview lists capabilities without token limits, and its deprecation page returned an error on the collection date, so Mistral is excluded from the lifecycle table rather than guessed at.
Knowledge cutoffs are not like-for-like. Anthropic publishes a “reliable knowledge cutoff” alongside a broader “training data cutoff” and the two differ — Claude Haiku 4.5 lists February 2025 and July 2025 respectively. This table uses the reliable figure for Anthropic, which is the conservative one; every other provider publishes a single unlabeled date whose definition is not stated. Anthropic therefore looks worse on this column than a naive comparison would suggest.
Google’s model pages no longer state a knowledge cutoff for the Gemini rows collected on 2026-10-14, so that column reads “Not published” for all three; the July Gemini 3.5 Flash page did state one.
Cache minimums for OpenAI, xAI, and DeepSeek are marked “Not published” because we could not locate a stated threshold in first-party documentation. All three support caching and price it; absence of a published floor is not evidence that no floor exists. If you rely on caching with those providers, measure the cached-token count directly rather than assuming.
Finally, the cache minimums listed here apply to each provider’s own API. Anthropic explicitly notes that Amazon Bedrock, as an AWS-operated platform, has its own per-model minimums, failure behavior, and usage-field names. If you consume Claude through Bedrock, Google Cloud, or Microsoft Foundry rather than the first-party API, verify against that platform’s documentation — the model is the same, the operational envelope may not be.
Update Cadence and Changelog
This table is refreshed on the first Tuesday of every month. Each refresh re-collects all figures from the six official documentation sites, updates data_updated and last_modified_at, and appends a dated line below describing what changed. Retirement dates that pass between refreshes are retained in the changelog rather than deleted, because the record of a retired model’s announced timeline is not preserved at the source.
- 2026-10-26 — Out-of-cycle lifecycle correction, prompted by the weekly pricing collection. The lifecycle table still showed two dates that had passed as the nearest confirmed retirements. OpenAI shut down
o3-mini,gpt-4-0613and the other models in its April 22 notice on 2026-10-23; its nearest scheduled removal is nowtts-1,tts-1-hdandgpt-4o-mini-ttson 2027-01-06, and among text models GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini on 2027-01-14, from a notice dated 2026-10-14. Google’sveo-3.1previews andgemini-omni-flash-previewreached their 2026-10-22 shutdown; its nearest is now five audio, TTS and live preview models on 2026-11-17.claude-haiku-4-5-20251001is still Active with no deprecation notice on 2026-10-26, so its earliest possible retirement moves to 2026-12-25. No capability row was re-collected. - 2026-10-19 — Out-of-cycle re-verification of one row and one lifecycle date, prompted by the weekly pricing collection. On 2026-10-17 DeepSeek replaced the model behind
deepseek-flashwith the DeepSeek-V4.1-Flash-1017 build under the same id and at the same price. Its pricing page, read 2026-10-19, still lists a 1,000,000-token context, a 384,000-token maximum output and image input, so the row’s figures stand and nothing in the capability table changed. Lifecycle:claude-haiku-4-5-20251001passed its 2026-10-15 floor without a deprecation notice and is still listed as Active on 2026-10-19, so the earliest possible retirement under the sixty-day minimum is now 2026-12-18. No other row was re-collected. - 2026-10-14 — Out-of-cycle refresh of the Google rows and one Anthropic addition, prompted by the weekly pricing collection. Row removed: Gemini 3.5 Flash, which Google deprecated on 2026-10-08 and now routes to
gemini-3.6-flash; this page had used it as Google’s representative row since July, in the capability table, the cache-minimum table, the cost bands, the FAQ and the shortlisting guidance, all of which have been updated. Rows added, all collected 2026-10-14: Gemini 4 Argon (1,048,576 tokens stated as a unified input-and-output limit, 65,536 output by default and up to 1,048,576 with long decode, text, image, video, audio and PDF input), Gemini 3.8 Flash and Gemini 3.6 Flash (1,048,576 input, 65,536 output, same input types); Google’s pages state no knowledge cutoff for any of the three. Claude Haiku 5.5 (1,000,000 context, 128,000 output, image input, June 2026 reliable knowledge cutoff, retirement floor October 7, 2027, 512-token cache minimum, eligible for the 300,000-token batch extension). Cache-minimum table updated from Anthropic’s and Google’s caching documentation, re-read 2026-10-14: the 512-token Claude row now also lists Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5, and the 4,096-token Gemini row lists Argon, 3.8 Flash and 3.6 Flash. Lifecycle table: Google’s nearest confirmed retirement had passed (imagen-4.0, 2026-08-17) and is now theveo-3.1previews andgemini-omni-flash-previewon 2026-10-22; Haiku 4.5 still has no deprecation notice on the eve of its 2026-10-15 floor. The scheduled monthly refresh of 2026-10-06 was missed and is recorded as a gap; this entry covers Google and the new Anthropic row only, and OpenAI, xAI and DeepSeek rows were not re-collected. - 2026-10-05 — Out-of-cycle addition of two rows, prompted by the weekly pricing collection; the scheduled monthly refresh is 2026-10-06. Rows added, both collected 2026-10-05: Claude Sonnet 5.5 (1,000,000 context, 128,000 output, image input, June 2026 reliable knowledge cutoff, retirement floor September 28, 2027) and GPT-6.1-sol (1,050,000 context, 128,000 output, image input, April 30, 2026 cutoff). The Message Batches 300,000-token extension now lists Sonnet 5.5. Lifecycle table corrected. It still gave
claude-sonnet-4-5’s September 29, 2026 floor as Anthropic’s nearest date after that date had passed. Anthropic deprecatedclaude-sonnet-4-5-20250929on 2026-09-30 and scheduled its retirement for 2026-11-30, namingclaude-sonnet-5-5as the replacement; that is now the nearest confirmed Anthropic retirement. The next floor isclaude-haiku-4-5-20251001at 2026-10-15, with no deprecation notice as of collection. No other row was re-collected; Google’s rows remain the 2026-07-25 figures. - 2026-09-28 — Out-of-cycle refresh of the Anthropic and OpenAI rows, prompted by the weekly pricing collection. Two errors corrected. Claude Opus 4.1 was retired on 2026-08-05 and has been removed from the capability table; this page had carried it for seven weeks as the max-output floor at 32,000 tokens, so that floor and the spread built on it were both wrong — the floor is Claude Haiku 4.5 at 64,000 and the spread is six-fold, not twelve. And the cost-band section still warned that Claude Sonnet 5’s introductory rate would expire on 2026-08-31 and rise 50%; Anthropic cancelled that increase on 2026-08-17, so the warning has been withdrawn rather than updated. Rows added, all collected 2026-09-28: Claude Opus 5.5 and Claude Fable 5.1 (1,000,000 context, 128,000 output, image input, June 2026 reliable knowledge cutoff), GPT-6-astra, GPT-6-sol and GPT-6-luna (1,050,000 context, 128,000 output, image input, cutoffs of April 30, April 20 and May 18, 2026), and Grok 4.7 (500,000 context, May 2026 cutoff, image support not stated). The Message Batches 300,000-token extension now lists Opus 5.5. Two documentation changes recorded rather than resolved. OpenAI has withdrawn the spec pages for GPT-5.6-sol, GPT-5.5 and GPT-5.4, whose rows keep their July figures and are marked; xAI has stopped stating image support per model, so Grok 4.7’s vision cell reads “Not published” while the older Grok rows keep the “Yes” collected in July. Claude Opus 5.5’s minimum cacheable prefix is not stated on the models overview and the caching documentation was not re-read this week, so it is absent from the cache-minimum table rather than guessed. Google’s rows were not re-collected and remain the 2026-07-25 figures, which means Gemini 3.6, 3.7 and 3.8 Flash are still missing from this page.
- 2026-09-21 — Out-of-cycle correction to the two DeepSeek rows, prompted by the weekly pricing collection rather than a scheduled refresh. DeepSeek retired v4-flash on 2026-09-14 and serves that name from
deepseek-flash, which accepts image input where its predecessor did not. The capability table, the vision bullet, the envelope analysis and the shortlisting guidance all said DeepSeek was text-only; that is now true only of v4-pro. Context window (1,000,000) and max output (384,000) are unchanged on both rows and were re-verified. No other provider’s rows were re-collected, so they remain the 2026-07-25 figures — see Methodology. The scheduled monthly refreshes of 2026-08-04 and 2026-09-01 were missed and are recorded as gaps rather than folded into this entry. - 2026-07-25 — Initial publication. 13 models across 6 providers. Baseline collected 2026-07-25.
FAQ: Best LLM 2026
What is the best LLM in 2026?
There is no single answer, because the models no longer differ mainly in capability — they differ in operating limits. If you generate long documents, DeepSeek v4 leads on max output at 384K tokens. If you need vision plus a 1M context, Claude Opus 5.5, Claude Sonnet 5.5, Claude Haiku 5.5, GPT-6.1-sol, Gemini 4 Argon and Gemini 3.8 Flash all qualify, and so does deepseek-flash since September 14, 2026 — only DeepSeek v4-pro is still text-only. Pick by the limit that blocks your workload, not by a leaderboard rank.
Which LLM has the largest context window in 2026?
They have converged. OpenAI’s GPT-5.x and GPT-6 families publish 1,050,000 tokens, Google’s current Gemini models 1,048,576, and Anthropic, xAI (Grok 4.3), and DeepSeek all publish 1,000,000. The spread across five providers is under 5%. Context window is no longer a differentiator; max output tokens, which ranges from 32K to 384K, still is.
Why does prompt caching sometimes not work even when enabled?
Every provider enforces a minimum cacheable prefix, and prompts below it silently fail to cache — no error, no warning. The thresholds range from 512 tokens (Claude Opus 5, Claude Haiku 5.5) to 4,096 (Claude Opus 4.6, Claude Haiku 4.5, Gemini 3.8 Flash). A 3,000-token system prompt caches on Opus 5 and Gemini 2.5 Flash and silently does not on Opus 4.6 or Gemini 3.8 Flash.
How much notice do providers give before retiring a model?
It varies by an order of magnitude. OpenAI commits to at least 6 months for generally available models, 3 months for specialized variants, and as little as 2 weeks for previews. Anthropic commits to at least 60 days but is the only provider publishing forward-looking “not sooner than” dates for models that are not yet deprecated. Google publishes no standard duration.
Is a knowledge cutoff date comparable across providers?
Not directly. Anthropic publishes two separate dates — a “reliable knowledge cutoff” and a broader “training data cutoff” — and they differ by months on some models. Every other provider publishes a single unlabeled date. Comparing one vendor’s conservative figure against another’s broad figure is not a like-for-like comparison.
Related Resources
- LLM API Pricing 2026: Full Comparison Table (Weekly) — exact per-token prices for all six providers, refreshed weekly.
- Best LLM Subscription 2026: What You Really Pay For — the same tradeoff at the consumer subscription tier.
- AI Crawler Ecosystem 2026: Who Scrapes the Web and Who Pays — how the models on this page acquire the data they run on.
Primary Sources
Data verified 2026-10-26. Machine-readable dataset: /data/best-llm-2026.json
- Models overview — Anthropic
- Model deprecations — Anthropic
- Prompt caching — Anthropic
- Models — OpenAI
- Deprecations — OpenAI
- Gemini 4 Argon model page — Google
- Context caching — Google
- Gemini API changelog — Google
- Models — xAI
- Pricing and model specifications — DeepSeek
- Models overview — Mistral AI