Claude Opus 5.5 Released: What Changed from Opus 5 and Fable 5.1, and Migration Pitfalls

On September 22, 2026, Anthropic released Claude Opus 5.5. The announcement describes it as the first model in the Claude 5.5 family, and it arrives just two months after Opus 5 (released July 24). In the model overview of the official documentation, it is now the recommended starting point for most use cases when you're unsure which model to pick, and Opus 5 has moved to Legacy (older, but still available).

The short version: the price went down compared with Opus 5 ($4 input / $20 output, versus $5 / $25 for Opus 5). But the API behavior changed a lot, and there are four changes that turn code that worked on Opus 5 into 400 errors. On top of that, the default effort you get when you don't specify one dropped from high to medium. If you switch over based on price alone, you can run into either errors or a quality shift you never notice. This article checks the official documentation and the announcement against each other in the original text and walks through what changed, how much it costs, and what breaks during migration, in the order you'll actually deal with them.

Information as of September 24, 2026: Written two days after release, after reading the original text of the Claude Platform documentation (models overview, the Opus 5.5 page, What's new, the migration guide and pricing), Anthropic's announcement, and the Claude Code documentation and CHANGELOG. Claude Sonnet 5.5 and Claude Haiku 5.5, which the announcement says are coming within weeks, had not been released at that point.

1. Opus 5.5 at a glance: performance, pricing, migration risks

In one sentence, Opus 5.5 is "a cheaper Opus that you handle the same way as Fable 5.1." For three of the four breaking changes, the official documentation itself notes that they also apply to Claude Fable 5.1. If your code has already been through a Fable 5.1 migration, there's much less left to do.

2. Core specs and availability

Let's start with the numbers. The comparison covers three models: Opus 5, which it replaces; Fable 5.1, the tier above; and Sonnet 5, the tier below.

Sources: Anthropic, "Models overview", "Claude Opus 5.5", "Claude Opus 5", "Migrating to Claude Opus 5.5" (checked September 24, 2026). Speed is a relative label within the current lineup and isn't listed for Opus 5, which is Legacy. Retirement dates are commitments for Anthropic-operated platforms; Amazon Bedrock and Google Cloud set their own.

What stands out in the table is that only two rows, "Thinking" and "Default effort," differ from Opus 5. Context and max output are identical, and the knowledge cutoff moved forward by just one month. On the spec sheet, what separates it from Opus 5 is how you handle it and what it costs; the performance gap is covered in the table in section 4. Also note that with the Message Batches API, adding the beta header output-300k-2026-03-24 raises the output limit to 300K tokens (same as Opus 5).

Platforms

3. Pricing: cheaper than Opus 5, but it depends on what you compare it to

The official What's new page gives the price as $4 / $20, below Opus 5's $5 / $25. This is the first change to Opus pricing after $5 / $25 held from Opus 4.5 through Opus 5 (confirmed in the table on the official pricing page). Here are the detailed rates.

Source: Anthropic, "Pricing" (checked September 24, 2026). The fast mode row only covers models listed in the pricing page's fast mode table. Sonnet 5's $2 / $10 was introduced at launch as introductory pricing through August 31 and then became the standard price (per a note on the same page).

"Cheaper" means something different against each model

Compared with Opus 5, it's cheaper on every line item. Against Fable 5.1 or Sonnet 5, however, the gap doesn't open evenly across line items. The chart below shows Opus 5.5's rate when the other model's rate is set to 100.

Three takeaways:

• Coming from Opus 5, the more cache-heavy the workload, the more you save. Input and output are 20% off, and cache reads are 60% off. The announcement itself points out that cache reads make up most of the cost of agentic and coding work.
• Against Fable 5.1, the gap is small on cache reads. Input and output cost 40% of Fable 5.1's price, but cache reads are $0.20 versus $0.25, only a 20% difference.
• Against Sonnet 5, only cache reads cost the same. Input and output are twice as expensive, but cache reads are $0.20 for both, so the price gap narrows as the share of cached tokens grows.

One more point: the per-token rate and the bill are two different things. Anthropic's announcement says that at default settings it costs 40% less than Opus 5 on typical work, because the per-token price is lower and it also uses fewer tokens per task. That's Anthropic's own measurement. Meanwhile, the official documentation also says that at the same effort, Opus 5.5 tends to think more per turn than Opus 5, especially at xhigh and max. Thinking tokens are billed as output tokens even when they aren't displayed. Workloads you ran cheaply on Opus 5 with thinking disabled can't disable it on Opus 5.5, so output tokens may go up. After migrating, the only way to know is to measure usage and compare again on your own workload. Pricing across all Claude models is covered in our Opus, Sonnet and Haiku pricing comparison.

4. Benchmarks: comparing only within Anthropic's table

The Opus 5 announcement showed its comparison table as an image, but this announcement uses a text table on the page. It has five columns: Opus 5.5, Fable 5.1, Opus 5, GPT-6 Astra and GPT-5.6 Sol. First, the table's conditions:

• Unless noted otherwise, Opus 5.5 was measured at maximum effort (max). Terminal-Bench 4.0 is the one exception: Opus 5.5 at xhigh and GPT-6 Astra at high (OpenAI's reported figure), each model's best score
• GPT-6 Astra and GPT-5.6 Sol on Terminal-Bench 4.0, and GPT-6 Astra on Terminal-Bench-Science 0.1, are figures reported by OpenAI. For the GPT figures in the other rows, there's no note on who ran them
• AutomationBench figures were run and reported by Zapier (Opus 5.5 from Zapier's early-access evaluation; Opus 5 and the two GPT models from Zapier's public leaderboard). This is the only row the table explicitly says was measured by a third party
• Opus 5.5 was measured with the same safeguards enabled as in production. On problems where a safeguard intervened, Opus 4.8 solved them instead for cybersecurity, and Opus 5 for biology and frontier LLM development. Anthropic notes that this likely lowers Opus 5.5's scores

Source: comparison table in Anthropic's "Introducing Claude Opus 5.5" (September 22, 2026; checked September 24). Green means Opus 5.5 is the top score in that row; red means the same table has a higher score than Opus 5.5 in that row. "—" means the table has no value.

Opus 5.5 has the top score in 7 of the 9 rows. The widest margins are on Terminal-Bench 4.0 (10.6 points ahead of Fable 5.1) and GDPval-AA v2.1 (111 points ahead of Fable 5.1). Compared with the previous Opus 5, Terminal-Bench-Science 0.1 went from 29.0% to 58.7%, roughly double.

Scores by effort level, from the announcement's chart descriptions

The table values are at max (xhigh for Terminal-Bench 4.0), but the announcement also includes charts plotting score against cost per task at each effort level, and their descriptions give figures at the default medium setting.

Source: chart descriptions in Anthropic's "Introducing Claude Opus 5.5" (all measurements and comparisons are Anthropic's own)

FrontierCode is the one worth pausing on. Within the same announcement, medium (54.6%) slightly beats max (54.4%). The pattern seen since Opus 5, that raising effort doesn't always raise scores, continues with Opus 5.5 (section 6). For GPT-6 Astra's own announcement and figures, see our GPT-6 Astra release breakdown.

5. Where it doesn't win, and how to read the table

Release coverage tends to list only the rows a model wins, so here's the other side. Everything below comes from Anthropic's own announcement and documentation.

6. What the new medium default effort means

This is the change most likely to be overlooked. A request that doesn't specify effort runs at high on Opus 5 and at medium on Opus 5.5. With the exact same code, the model thinks one level less deeply.

So which level should you use? Here is the official documentation's guidance, summarized for Opus 5.5.

Sources: Anthropic, "Effort", the migration guide, Claude Code, "Model configuration"

The official advice is consistent: don't carry over settings from the previous model; re-tune effort on your own eval set. The announcement also quotes early-tester companies, such as Deloitte's report that even at the lowest effort it found more known bugs than Opus 5 at high, but these are results each company measured on its own work and Anthropic chose to publish; there's no guarantee your workload will behave the same. Check on your own work whether medium holds up on quality, and what improves when you raise it to high. On Opus 5.5, you can also change effort per message mid-conversation without breaking the prompt cache (beta).

7. Four things that break when you move from Opus 5, and the fixes

The official What's new in Claude Opus 5.5 lists four breaking changes that affect code running on Opus 5. All of them return a 400 invalid_request_error.

① Thinking can't be disabled

Opus 5 accepted thinking: {"type": "disabled"} at effort high or below. On Opus 5.5, thinking is always on, and both disabled and a manual budget, {"type": "enabled", "budget_tokens": N}, return an error at every effort level. The message you get back is:

"thinking.type.disabled" is not supported for this model. Use "thinking.type.adaptive" and "output_config.effort" to control thinking behavior.

Fix: remove the thinking parameter (it's then treated the same as {"type": "adaptive"}). Where you disabled thinking to save money, lower the effort instead.

# Before: works on Opus 5, returns 400 on Opus 5.5
client.messages.create(
    model="claude-opus-5",
    max_tokens=16000,
    thinking={"type": "disabled"},
    messages=[...],
)

# After: thinking is always on; effort sets how much
client.messages.create(
    model="claude-opus-5-5",
    max_tokens=16000,
    output_config={"effort": "low"},
    messages=[...],
)

Because thinking always runs, a response can start with a thinking block. Code that reads by position, such as content[0].text, will break, so select the blocks whose type is "text". In tool-use loops, send thinking blocks back untouched, even when they're empty (editing, reordering or dropping some of them returns 400). max_tokens caps thinking and the response combined, so workloads that had thinking disabled need their limit revisited.

② Forced tool use returns an error

Setting tool_choice to {"type": "any"} or {"type": "tool", "name": "..."} returns 400. The token counting API applies the same check. Only auto (the default) and none are allowed.

tool_choice: type "tool" and "any" are not supported for this model.

Fix: keep tool_choice at auto and add strict: true to the tool definitions (strict tool use), or move the schema to structured outputs. Then state in the prompt when the tool should be used. If you were using forced tool use just to guarantee JSON output, moving to structured outputs is the cleaner path. One caveat: strict tool use only accepts a subset of JSON Schema, and every object in the schema needs additionalProperties: false.

# After: auto + strict; say in words when to use the tool
client.messages.create(
    model="claude-opus-5-5",
    max_tokens=1024,
    tools=[{**tool, "strict": True} for tool in tools],
    tool_choice={"type": "auto"},
    messages=[{"role": "user",
               "content": "What's the weather in Paris? Use the get_weather tool."}],
)

③ Thinking blocks are bound to the model and the conversation

Each thinking block records which model produced it. Opus 5.5 can read thinking blocks from Opus, Sonnet and Haiku models up to Opus 5, but not from Fable or Mythos. In the other direction, only Fable 5.1 and Mythos 5.1 on the Claude API can read Opus 5.5's thinking blocks.

• Switching Opus 5 → Opus 5.5, or Opus 5.5 → Fable 5.1 (Claude API) carries the earlier reasoning over
• Any other switch (for example, Opus 5.5 → Sonnet 5) means the turns after the switch run without the previous model's reasoning. The request itself succeeds, and the dropped blocks aren't billed

In addition, the API checks that nothing before an Opus 5.5 thinking block (system, tools, earlier messages) has changed since the block was created. For accounts created on or after August 31, 2026, 00:00 (UTC), this is enforced by default, and sending blocks on top of a history you rewrote mid-conversation returns 400.

Fix: advance the conversation by appending only. When you want to change instructions or tools, use a mid-conversation system message instead of rewriting the history. According to the migration guide, Claude Code, claude.ai, Claude Managed Agents and the Claude Agent SDK already work append-only, so if you're building on top of them, nothing needs to change. If rewriting can't be avoided, add the beta header thinking-binding-controls-2026-08-01 and set thinking.block_binding.prefix_mismatch_behavior to "drop_block" to have the affected blocks dropped instead of getting an error. The mechanism is the same as in Fable 5.1, which we covered in detail in our article on Fable 5.1's breaking changes.

④ The old computer use tool isn't available on the Claude API and Google Cloud

Opus 5 accepted computer use through both the new toolset computer_toolset_20260801 and the older tool computer_20251124 with a beta header. Opus 5.5 on the Claude API and Google Cloud accepts only the toolset, and declaring the old tool returns 400. The message starts with:

'claude-opus-5-5' does not support tool types: computer_20251124.

Fix: remove the beta header and replace tools with [{"type": "computer_toolset_20260801"}] (no name, no screen size). The agent loop needs changes too, in three places: the action type is in the block's name, not in input.action; a single turn can contain multiple actions; and every result must be returned with toolset_name. On Amazon Bedrock, the old tool keeps working with Opus 5.5, so no change is needed there.

8. Changes that happen without an error

The four breaking changes return a 400, so you'll notice them. The tricky ones are the changes where only the behavior shifts, with no error.

Sources: What's new in Claude Opus 5.5, Migrating to Claude Opus 5.5

9. What changes in Claude Code and claude.ai

This is the part that matters if you don't call the API directly. Claude Code added Opus 5.5 support in v2.1.280 of the CHANGELOG.

Sources: Claude Code, "Model configuration" and the CHANGELOG (v2.1.280 and v2.1.281, checked September 24, 2026)

How Claude Code stores effort is explained in our guide to the effort setting, and how thinking itself works is covered in adaptive vs. extended thinking. According to Anthropic's announcement, fast mode (up to 2.5x faster, $8 input / $40 output) is also available in Claude Code.

claude.ai and the apps

Opus 5.5 has been selectable on claude.ai and the iOS and Android apps since launch day (Anthropic published an Opus 5.5 system prompt dated the same day). The announcement also says that, alongside the price cut, it will raise the five-hour usage limits for Pro, Max, Team and seat-based Enterprise, and give subscribers one usage-limit reset to use whenever they like. The announcement doesn't give a number for how much the limits go up. Per-plan pricing is covered in our Claude vs ChatGPT pricing comparison.

10. Choosing between Opus 5.5, Fable 5.1, Sonnet 5 and Opus 5

The guidance in the official models overview is short: if you're unsure, start with Opus 5.5; use Fable 5.1 for demanding reasoning and long-running agentic work, or when Opus 5.5 still falls short after you've evaluated it at high effort. Layer pricing and API constraints on top, and the choice breaks down like this.

If you're unsure, the order is simple: ① run Opus 5.5 at medium and high on your own work → ② if that's not enough, try xhigh → ③ if that's still not enough, move to Fable 5.1 → and if medium turns out to be plenty, check whether low or Sonnet 5 would also do. Raising effort and moving up a model are both ways of paying for quality, and which one is cheaper for the gain depends on the work.

11. Migration steps (for API users)

From the migration guide's checklist, here are the items that apply when moving from Opus 5, in the order you'd do them. Claude Code also has /claude-api migrate to help with this (a bundled skill mentioned in the migration guide; it confirms the scope with you before editing).

Source: items from the "Every starting model" section of Anthropic, "Migrating to Claude Opus 5.5", reordered into working order

If you're coming straight from Opus 4.8 or earlier, you also need to deal with thinking now running on requests that never specified it. It's worth checking the migration steps in our Opus 5 release breakdown as well.

Summary

Claude Opus 5.5 is a release where a price cut and a change in how you handle the model arrived together. Rates dropped 20% on input and output and 60% on cache reads, and in Anthropic's own comparison table it beats Fable 5.1 on many rows. On the other hand, thinking can no longer be turned off, and forced tool use, the old computer use tool and history rewriting now cause 400 errors.

And the drop in default effort to medium is easy to miss precisely because it doesn't produce an error. On the API, set effort explicitly; in Claude Code, check that your Opus 5 settings haven't silently stopped applying. Whether it actually got cheaper is something to measure with usage, not per-token rates.

Finally, benchmark figures are values measured in that table, under those conditions. On AutomationBench and Terminal-Bench-Science, GPT-6 Astra sits higher in the same table, and Anthropic itself says the gap with Fable 5.1 is smaller than the numbers suggest. Start by comparing medium and high on your own work.

FAQ

* Figures in this article are based on Anthropic's official announcement "Introducing Claude Opus 5.5" (benchmarks from the comparison table and chart descriptions on that page), the official documentation "Models overview," "Claude Opus 5.5," "What's new in Claude Opus 5.5," "Migrating to Claude Opus 5.5" and "Pricing," and Claude Code's "Model configuration" and CHANGELOG (all checked September 24, 2026). Specifications and pricing may change, so please confirm the final details in the official documentation.

Related articles: Claude Opus 5 release breakdown, Claude Fable 5.1 breaking changes and migration.