Claude Opus 5.5: Fable-level performance at 40% lower cost

Anthropic launched Claude Opus 5.5 on 22 September 2026, just two months after Opus 5 arrived. The headline claim is a meaningful one: on Anthropic’s own benchmarks, Opus 5.5 matches or exceeds Fable 5.1 across agentic coding, computer use, knowledge work, visual chart recognition, and multidisciplinary reasoning. Fable is Anthropic’s highest-capability tier, sitting above Opus in price and power. Getting Fable-level results from an Opus-priced model matters.

What it costs

The pricing shift is the most immediately practical part of this announcement.

	Input	Cache reads	Cache writes (5 min)	Output
Claude Opus 5	$5 / MTok	$0.50 / MTok	$6.25 / MTok	$25 / MTok
Claude Opus 5.5	$4 / MTok	$0.20 / MTok	$5 / MTok	$20 / MTok

Input and output tokens are 20% cheaper. Cache reads drop by 60%. For applications that rely heavily on prompt caching, as many production agentic systems do, the effective saving on a typical workload can reach 40%. Anthropic also says output generation is more than 30% faster, with a “Fast mode” reaching up to 2.5 times the speed of Opus 5 (priced at $8 input / $40 output per million tokens).

The model ID is claude-opus-5-5 and is available on the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Azure.

Always-on adaptive thinking

This is the behaviour change most likely to catch existing users off guard. On Opus 5, you could disable thinking entirely. On Opus 5.5, you cannot. Sending thinking: disabled returns a 400 invalid_request_error. Thinking is always on.

Control now comes through the effort parameter, with five levels: low, medium, high, xhigh, and max. The default is medium, and Anthropic says medium effort matches or beats Opus 5 at higher settings. The good news is that medium effort is also cheaper, so most teams will not need to change their effort level to stay within budget.

The practical implication: if your code pins effort: high on Opus 5 and you simply swap the model ID to claude-opus-5-5, the model will think more per turn than it did before at that setting. You will still benefit from the per-token price cut, but some of those savings may be absorbed by increased thinking token usage at high, xhigh, or max effort.

Four breaking changes to know about

Anthropic has published a migration guide, and there are four breaking changes relative to Opus 5:

1. Thinking cannot be disabled. As above, sending thinking: disabled is now an error.
2. Forced tool use returns an error. The tool_choice: any pattern that forces tool invocation is no longer supported.
3. Thinking blocks are tied to the model and conversation. You cannot mix thinking blocks from different models or insert them from an external source.
4. The computer_20251124 computer use tool is not accepted on the Claude API or Google Cloud. Use the newer computer_toolset_20260801 instead.

If you are running Opus 5 in production, those four points are worth checking before you update the model ID.

Refusals work differently too

Opus 5.5 is the first Opus model to carry a similar class of safeguards to Fable 5.1, covering cybersecurity, biology, and distillation. Rather than returning a hard error, a declined request returns HTTP 200 with stop_reason: "refusal" and a stop_details object naming the policy area. A server-side fallback (currently in beta) can automatically retry on a different model. For most cybersecurity tasks specifically, Anthropic routes these to Opus 4.8 transparently.

What changes for Pro, Max, and Team subscribers

Alongside the API launch, Anthropic announced increased five-hour usage limits for Pro, Max, Team, and seat-based Enterprise plans. The five-hour limit governs how much usage a subscriber can consume within a session window, separate from any weekly cap.

Every qualifying subscriber also receives a one-time rate-limit reset that can be activated at any point before 22 October 2026. If you have been hitting usage walls on demanding projects or extended conversations, this gives you additional headroom to use whenever it suits you.

The bigger picture

Opus 5.5 is the first model Anthropic has released since CEO Dario Amodei publicly called for a more measured approach to capability advancement. The release does not slow down, but it does place unusual emphasis on safety evaluation. Anthropic says Opus 5.5 underwent external review by Frontier Design and METR before launch, scored better than any recent Claude model on its automated behavioural audit across nearly 2,000 scenarios, and is meaningfully more resistant to prompt injection than Opus 5.

Sonnet 5.5 and Haiku 5.5 are expected in the coming weeks, with the same generation of performance and efficiency improvements.

What this means for you

If you are an API developer: the cost reduction is real and significant, particularly if your application leans on prompt caching. The breaking changes are narrow but require attention before migration. Check the What’s New guide, update your computer use tool reference, and test your effort level to confirm you land where you expect on the thinking-versus-cost trade-off.

If you are a Pro, Max, or Team subscriber: you have more room to run long sessions and demanding projects, and a bonus reset to use before 22 October 2026. There is no action required to receive the higher limits; they apply automatically to your plan.

If you are evaluating Claude against other providers: Anthropic is claiming Opus 5.5 outscores OpenAI’s GPT-5.6 Sol on SWE-bench Pro (89.9% versus a lower figure) at roughly one-third the cost. That is a benchmark, not a guarantee, but it is a useful reference point when comparing options for agentic coding workloads.

The context window remains 1,000,000 tokens with a 128,000-token synchronous output limit, or up to 300,000 tokens via the Batch API with a beta header. Everything else you relied on in Opus 5, including vision, PDF support, the Files API, and prompt caching, carries over.