Every Claude model available in August 2026

The complete current Claude model lineup with API IDs, context windows, prices, knowledge cutoffs and what each one is actually good at - plus which older models are retired.

As of August 2026 there are four Claude models you can call: Claude Fable 5 ($10/$50 per million tokens), Claude Opus 5 ($5/$25), Claude Sonnet 5 ($2/$10) and Claude Haiku 4.5 ($1/$5). The first three carry a 1,000,000-token context window by default. Six older models have already been retired and now return an API error.

This page is the reference table for all of them: exact API identifiers, context and output limits, knowledge cutoffs, thinking behaviour, platform-specific IDs for AWS and Google Cloud, and the retirement dates for everything Anthropic has switched off. Everything below was checked against Anthropic's own documentation on 21 August 2026.

#Full Claude model specifications, August 2026

All four current models take text and image input and return text. None of them generate images, video or audio, and none accept video or audio input. Prices are per million tokens on the standard Claude API.

Two details in that table catch people out. The 1M-token window is the default and the maximum on Fable 5, Opus 5 and Sonnet 5 - there is no beta header to send and no long-context surcharge, so a 900,000-token request bills at the same per-token rate as a 9,000-token one. And the knowledge cutoffs are not in version order: Opus 5 knows the world through May 2026, five months further than the nominally more capable Fable 5.

#A short profile of each current model

#Claude Fable 5 - the long-horizon agent model

Anthropic describes Fable 5 as next-generation intelligence for long-running agents. It has always-on adaptive thinking that cannot be disabled, never returns a raw chain of thought, and generates at roughly 75 output tokens per second - slow by 2026 standards. At $10/$50 it is exactly twice the price of Opus 5, and its rate limits are materially tighter. It also has its own refusal path: declined requests come back as HTTP 200 with stop_reason: "refusal" rather than an error. The full picture, including the invitation-only Mythos 5 variant, is on the Claude Fable 5 page.

#Claude Opus 5 - the general flagship

Released 24 July 2026, Opus 5 is Anthropic's recommended starting point for complex agentic coding and enterprise work: multi-hour autonomous coding runs, large refactors, vision-heavy workflows and computer use. It carries the newest knowledge of any Claude model and half Fable 5's price. Thinking depth is set through an effort ladder rather than a token budget. Details, migration notes and the cases where it is overkill are on the Claude Opus 5 page.

#Claude Sonnet 5 - the one most people should use

Sonnet 5 is the balance point: frontier-class behaviour at $2/$10, which since 10 August 2026 is the permanent price rather than an introductory one - the increase to $3/$15 that had been scheduled for 1 September was cancelled. One caveat worth doing the arithmetic on: Sonnet 5 uses the newer tokenizer, which emits roughly 30% more tokens for the same text than Sonnet 4.6 did, so the real-world saving is smaller than the sticker price implies. See Claude Sonnet 5.

#Claude Haiku 4.5 - the cheap, fast one

Haiku 4.5 is the outlier in the current lineup: a 200k context rather than 1M, 64k max output, a July 2025 training cutoff, and the only current model that still uses manual extended thinking with a token budget instead of adaptive thinking. That makes it the right choice for classification, extraction, routing and sub-agent work where volume matters more than depth. See Claude Haiku 4.5.

#Claude model IDs on Bedrock, Vertex AI and Microsoft Foundry

The same model carries a different identifier on each cloud. From the 4.6 generation onward the IDs are dateless - claude-opus-5, not claude-opus-5-20260724 - but a dateless ID is still a pinned snapshot, not an evergreen pointer. Anthropic does not update the weights behind an existing ID; a new version ships under a new ID.

Older dated IDs keep their snapshot suffix on every platform: Bedrock appends -v1:0, Google Cloud separates the date with @. Opus 4.6 was the last Bedrock ID to carry a -v1 suffix at all. Not every API feature crosses over - the MCP connector beta, for example, is unavailable on Bedrock and Google Cloud, which matters if you are building on MCP-based agents.

#Which Claude model should you pick?

Work down this list and stop at the first line that matches. Most teams stop at step two.

1. High volume, shallow task? Classification, tagging, routing, extraction, sub-agent calls inside a larger loop - use Haiku 4.5 at $1/$5 and accept the 200k context.
2. Anything else, by default? Sonnet 5. It is the workhorse: 1M context, adaptive thinking, $2/$10. Do not move up until a real task fails, not because a benchmark chart suggests you should.
3. Complex agentic coding, large refactors, computer use, or you need knowledge past January 2026? Opus 5 at $5/$25.
4. A long-horizon agent where a single failure costs more than the entire token bill? Fable 5 at $10/$50 - and only for the step that needs it, with cheaper models doing the rest of the loop.

Two levers change the arithmetic more than the model choice does. Prompt caching drops repeat input to 0.1× base price, and the Batch API takes a flat 50% off everything if you can tolerate a 24-hour window. A cached, batched Opus 5 call can cost less than an uncached Sonnet 5 one. Both are worked through with numbers on the Claude API page, and the subscription side - where model access is bundled rather than metered - is on the pricing page. If you want the choice framed as cost-per-completed-task rather than per token, use the Haiku vs Sonnet vs Opus comparison.

#Legacy Claude models that still work

Six older models remain callable but are no longer the recommended target for new work. They keep their published retirement floors - Anthropic commits to a "not sooner than" date rather than a fixed switch-off.

Sonnet 4.5 is the oldest Claude model still alive, and its floor is the nearest - 29 September 2026. If you are still pinned to it, Sonnet 5 is both newer and cheaper, which is an unusually easy migration to justify.

#Which Claude models are retired?

Six models have been switched off since February 2026. A request naming any of them returns an error from the Claude API; some remain reachable through Bedrock or Google Cloud, which is why teams occasionally believe a retired model is still working.

#How Claude version numbering works

The naming has changed twice. Models before 4.6 use a dated identifier - claude-sonnet-4-5-20250929 - plus a short alias that resolves to the most recent snapshot. From 4.6 onward the date disappears: claude-sonnet-4-6, claude-opus-5. The 5-generation then broke the pattern again by introducing a name, Fable, that sits above Opus rather than continuing the Opus/Sonnet/Haiku triple.

The practical rule: weights are fixed per ID, but the serving stack around them - router, safety classifiers, sampling logic - can change, so identical requests to the same ID can behave slightly differently over time. Pin the ID, and treat behavioural drift as expected rather than as a bug in your prompt. The developer overview covers how to structure evaluations that survive this.

#Every Claude model page on this site

#Frequently asked questions