Anthropic Publishes a Claude Model Guide: How Haiku, Sonnet, Opus and Fable Divide the Work

What happened

According to the claude.dev blog, Claude currently has four general-purpose models: Claude Haiku 5.5, Claude Sonnet 5.5, Claude Opus 5.5 and Claude Fable 5.2. The post also says Claude Mythos 5.2 is its most capable model for cybersecurity and life-sciences research, available through verification programs.

claude.dev says most work should start on Opus 5.5 with the effort level left at the default, medium. The effort level is a setting that decides how hard the model thinks, how many tools it calls and how much it checks its own work. Opus 5.5 is also the default model in the coding tool Claude Code. A token is the unit of text a model processes and bills for, and Opus 5.5 costs half as much as Fable 5.2 per input and output token.

How the four models divide the work

The post's recommendations are as follows. Well-defined everyday work, such as fixing bugs or building features from a spec, can move down to Sonnet 5.5. High-volume, simple work such as classification, routing and data extraction can use Haiku 5.5, which also suits subagents, the small helpers a main model dispatches to handle side tasks. Move up to Fable 5.2 for the hardest coding, finance and science projects, when Opus 5.5 at high is still not enough, or for tasks that need to run unattended for a long time. claude.dev also cautions that migrations or performance optimization may still need Opus 5.5 even when the spec is clear.

Adjust effort level before changing models

claude.dev recommends adjusting the effort level before switching models. The post notes that raising effort increases latency, meaning a longer wait for a response. Raising effort also cannot give a model knowledge or skills it does not already have; in that case a larger model is still needed.

The post cites Anthropic's own benchmarks, which are standardized tests that compare models on fixed tasks. On Terminal-Bench 3.0, the previous-generation Fable 5.1 scored 58.0% at max; Opus 5.5 scored 58.9% at high, using half the tokens. On FrontierCode 2.0, Fable 5.2 scored 45.3% at high, on par with Opus 5.5's 45.5% at max, at roughly half the cost per task. These figures come from the vendor's own testing.

In another test on a SWE-bench Pro subset, claude.dev says Opus 5.5 ran all tasks at low first and re-ran the failures at high, resulting in about 97% of tasks passing at roughly $0.17 each. Running every task at high gave a 95.3% pass rate at $0.29 each.

Compatibility to check before switching models

• claude.dev says that on Sonnet 5.5, Opus 5.5 and Fable 5.2, setting tool_choice (the setting that forces the model to use a tool) to any or tool returns a 400 error.
• Opus 5.5 and Fable 5.2 always "think" first, meaning they reason internally before answering; requests that turn thinking off return an error.
• If the conversation before a Fable 5.2 thinking block is modified, the API rejects the request with a 400 error unless you choose to drop the affected thinking. Requests made through Claude Code, the Claude Agent SDK or Claude Managed Agents need no changes.
• In Claude Code, haiku is a shorthand alias for a model. On the Anthropic API it now points to Haiku 5.5, which requires Claude Code v2.1.293 or later; on Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS it still points to Haiku 4.5.
• Opus 5.5, Sonnet 5.5 and Fable 5.2 run safety classifiers that most often flag cybersecurity and biology content; in Claude Code, flagged requests may be handled by an older model. The post says Fable 5.2 keeps Fable 5.1's fallback models: cybersecurity requests fall back to Opus 4.8 and biology requests to Opus 5.

claude.dev also mentions prompt caching, a mechanism that stores previously sent conversation content so the repeated portion is billed at a lower price. A cache belongs only to the model that wrote it. As a result, the first turn after switching models requires paying again to write the cache for the whole conversation.

What it means for general readers and developers

For general readers, the guide shows that models from the same company are not simply "bigger is better"; price, speed and task difficulty have to be weighed against each other. claude.dev recommends that developers compare models with these steps:

1. Pull samples from your own real tasks and write a way to check the result of each one.
2. Run those tasks on each candidate model at two or three effort levels, then compare the results.
3. Before switching for real, "shadow" test on a portion of real traffic, letting the new setup run in the background without affecting the live service.

The post also suggests setting subagents to a smaller model, for example by writing model: haiku or model: sonnet in the subagent definition. Code editing and work that requires judgment should stay with the main model.

Browse the latest news in this topic

• AI news