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, Gemini 3.5 Flash 65,536, and Claude Opus 4.1 just 32,000. 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 near-universal except DeepSeek: every listed Anthropic, OpenAI, Google, and xAI model accepts image input. DeepSeek v4 is text-only.
- 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 (platform.claude.com), OpenAI (developers.openai.com), Google (ai.google.dev), xAI (docs.x.ai), DeepSeek (api-docs.deepseek.com), and Mistral (docs.mistral.ai). No aggregator, leaderboard, or third-party spec sheet was used as a source at any point.
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 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 | Anthropic | 1,000,000 | 128,000 | Yes | Jan 2026 |
| Claude Haiku 4.5 | Anthropic | 200,000 | 64,000 | Yes | Feb 2025 |
| Claude Opus 4.1 | Anthropic | 200,000 | 32,000 | Yes | Jan 2025 |
| 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 3.5 Flash | 1,048,576 | 65,536 | Yes, plus video, audio, PDF | 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 v4-flash | DeepSeek | 1,000,000 | 384,000 | No | Not published |
Two footnotes materially change how these rows read. Anthropic’s 128,000 output ceiling rises to 300,000 on the Message Batches API for Opus 5, Opus 4.8, Opus 4.7, Opus 4.6, Sonnet 5, and Sonnet 4.6 behind a beta header — so the synchronous number understates the asynchronous ceiling by more than double. And Anthropic’s newer models (Opus 4.7 and later, Sonnet 5, Fable 5) 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.
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.
| Model or family | Minimum cacheable prefix | Mechanism |
|---|---|---|
| Claude Opus 5, Claude Fable 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 3.5 Flash, Gemini 3.1 Pro Preview | 4,096 tokens | Implicit caching |
| 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. Gemini 2.5 Flash caches at 2,048; the newer Gemini 3.5 Flash 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 twelve-fold spread from Claude Opus 4.1’s 32,000 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 published ceiling that lets a single request return a document proportionate to a million-token input. Meanwhile Gemini 3.5 Flash pairs 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.
The same logic explains why DeepSeek’s text-only limitation is a sharper constraint than its price advantage suggests. Every other provider in this table accepts image input; Gemini additionally accepts video, audio, and PDF natively. A pipeline that touches screenshots, scanned documents, or charts cannot 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, not after it.
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 v4-flash, Gemini 2.5 Flash-Lite, Claude Haiku 4.5 | High-volume classification, extraction, routing |
| Mid | Claude Sonnet 5, Gemini 3.5 Flash, GPT-5.4-class | Most production workloads; the default tier |
| Flagship | Claude Opus 5, Claude Fable 5, GPT-5.6-sol, GPT-5.5-pro | Long-horizon agentic work, hardest reasoning |
Band assignments derive from the 2026-07-16 weekly pricing snapshot. They are deliberately coarse: a band does not move week to week, whereas a sticker price does, and one time-sensitive figure is worth carrying here — Claude Sonnet 5’s introductory rate expires August 31, 2026, after which it rises 50%. Budget against the September figure, not today’s.
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-opus-4-1, Aug 5, 2026 |
| OpenAI | 6 months (GA), 3 months (specialized), ~2 weeks (preview) | No | o3-mini and gpt-4-0613, Oct 23, 2026 |
| Not stated | No | imagen-4.0 family, Aug 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 is documented as safe until at least 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.
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 is out regardless of price. If it touches video, audio, or PDFs natively, Gemini 3.5 Flash is the only model in this table that handles 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, and only Anthropic’s batch API extension reaches 300,000.
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 publishes full specs for GPT-5.6-sol, GPT-5.5, and GPT-5.4 but not for the mini, nano, and pro variants, which are priced on the pricing page without corresponding spec pages. 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.
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.
- 2026-07-25 — 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, GPT-5.6-sol, and Gemini 3.5 Flash all qualify and DeepSeek does not. 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 family publishes 1,050,000 tokens, Google’s Gemini 3.5 Flash 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) to 4,096 (Claude Opus 4.6, Gemini 3.5 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 3.5 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.