Anthropic's autumn has had two parts. At the top end, Claude Fable 5.1 and the restricted Mythos 5.1 arrived on 1 September with a 1M-token context window and premium pricing of $10 / $50 per million tokens. Then, over 16 days, Anthropic refreshed the three models most people actually use: Opus 5.5 on 22 September, Sonnet 5.5 on 28 September and Haiku 5.5 on 7 October.
The common thread is price. Each release claims to be faster and cheaper than the model it replaces. Below, each claim is attributed to Anthropic and linked to the announcement, because benchmark and cost figures from developers deserve some scepticism until independent testers confirm them.
Where Fable and Mythos fit
It helps to understand the top tier first. Fable 5.1 is Anthropic's most capable broadly available model. Mythos 5.1 is offered only to participants in Anthropic's Project Glasswing. Anthropic is deliberately keeping its most powerful capabilities within a smaller group. That sets the tone for the 5.5 releases: they bring a lot of capability down to lower prices, but some of the top tier's safety controls come with them.
Claude Opus 5.5: the workhorse
Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, with cached input reads at $0.20. Anthropic makes two efficiency claims: that it costs about 40% less than Opus 5 on typical workloads, and that it generates output more than 30% faster. On Terminal-Bench 4.0, which tests how well a model works in a command-line environment, Anthropic reports 66.4%.
Three other details are more significant than the benchmark:
- External testing before launch. Anthropic says METR and Frontier Design evaluated Opus 5.5 before release. Independent pre-release testing is still not universal, so naming the testers matters.
- Released under "pacing" principles. Opus 5.5 is the first Anthropic release since Dario Amodei called for the industry to "pace the frontier". I look at that debate in Pacing the frontier.
- Top-tier safeguards. It ships with Fable-class cyber and biology safeguards and a Life Sciences Verification Program, through which genuine researchers can get the access they need. It also includes watermarking to support EU AI Act compliance.
Claude Sonnet 5.5: the middle option
Sonnet has long been the popular middle option: capable enough for serious work and cheap enough to use all day. Anthropic says Sonnet 5.5 is more than 30% faster than Sonnet 5 and up to 30% cheaper. Then, with the Haiku 5.5 launch on 7 October, Anthropic halved Sonnet 5.5's cache-read price to $0.10 per million tokens.
The cache-read cut deserves attention. Many real applications send the same long context, such as a policy manual, a codebase or a product catalogue, again and again with small changes. Caching lets you pay a reduced rate for the repeated part. Halving that rate can matter more to a business's bill than a headline price cut.
Claude Haiku 5.5: small and cheap
Haiku 5.5 is priced at $0.10 / $0.50 per million tokens for prompts up to 100,000 tokens. Anthropic says this makes it about 75% cheaper to run than Haiku 4.5. That price is the same as OpenAI's GPT-6 Luna, which launched two weeks earlier.
It is also the first Haiku with an effort setting, so you can tell the model to think harder on difficult inputs and answer quickly on easy ones. That kind of control used to be reserved for the bigger models.
The same announcement added computer-use and browser-use features to Anthropic's SDKs in beta, and introduced monthly API credits for Max and Team subscribers, with tiers of $100, $200 and up to $500. For people who build tools on Claude, that credit removes some of the friction between a chat subscription and API experiments.
How to choose
| Model | Price (in / out per M tokens) | Good for |
|---|---|---|
| Fable 5.1 | $10 / $50 | The hardest problems, very long documents |
| Opus 5.5 | $4 / $20 | Complex reasoning, long agent tasks, serious coding |
| Sonnet 5.5 | Up to 30% below Sonnet 5 (Anthropic) | Everyday professional work at volume |
| Haiku 5.5 | $0.10 / $0.50 (≤100k tokens) | High-volume tasks, sub-agents, classification |
Prices as published by Anthropic at launch. Check the official pricing pages before you commit to a budget.
A sensible approach for a small business is to start on Sonnet 5.5, send only the hardest tasks to Opus 5.5, and pass simple repetitive work to Haiku 5.5. Agent builders often go further and use Opus or Fable to plan and many Haiku sub-agents to do the steps.
The competitive picture
Anthropic's refresh did not happen in isolation. OpenAI's Sol and Luna arrived on 22 September, the same day as Opus 5.5, with their own 50% price cut. Google's Gemini 4 Argon became generally available on 13 October at an introductory $2 / $10. The pattern across all three companies is clear: capability that was expensive a year ago is now cheap, and companies are competing as much on price, speed and safety controls as on raw intelligence.
The bottom line
Claude 5.5 is an efficiency release with an important safety element. If you already use Claude, the main changes are lower bills and faster responses. If you are choosing a provider, the more interesting point is that Anthropic is shipping top-tier safeguards on its cheaper models, opening controlled access for legitimate researchers, and naming its external testers. Those are signs of how frontier models will be released from now on.
For a side-by-side look at every recent model, see the AI Models page.
Sources
- Claude Fable 5.1 and Mythos 5.1 — Anthropic (1 September 2026)
- Claude Opus 5.5 — Anthropic (22 September 2026)
- Claude Sonnet 5.5 — Anthropic (28 September 2026)
- Claude Haiku 5.5 — Anthropic (7 October 2026)
- Claude Platform release notes — Anthropic (accessed 22 October 2026)
- We Must Pace the Frontier — Dario Amodei (September 2026)
- Introducing GPT-6 Sol and Luna — OpenAI (22 September 2026)
- Introducing Gemini 4 Argon (general availability) — Google (13 October 2026)
Figures quoted from company announcements are the companies' own claims. Independent testing may give different results.