Claude Opus 5.5 is a compelling upgrade candidate for long coding and knowledge-work tasks. Released on September 22, 2026, it lowers Anthropic's standard API price to $4 per million input tokens and $20 per million output tokens, while improving the published results over Opus 5. The catch is practical: some requests that worked on Opus 5 now fail, and an unchanged agent interface can lose its progress updates.

Start your evaluation at medium, the new default. Compare completed tasks, billable tokens and correction time before replacing a working integration. Routine extraction or short edits may still be cheaper on Sonnet 5.

Checked September 23, 2026. This review analyzes Anthropic's release, model documentation and migration guide. We did not run an independent benchmark or a paid latency test. Published scores below are attributed; cost examples are calculations at direct API list prices.

What is new in Claude Opus 5.5?

The headline is better economics for work that takes many turns: codebase migrations, difficult debugging, document analysis and agents that use tools. Anthropic reports clearer writing, stronger visual understanding and fewer steps on complex work. Its early-access examples are useful leads for evaluation, rather than promises about your repository.

Specification	Claude Opus 5.5
Release date	September 22, 2026
Direct API model ID	`claude-opus-5-5`
Context / standard maximum output	1M / 128K tokens
Input → output	Text and images → text
Reliable knowledge cutoff	June 2026
Thinking	Adaptive, always on
Default effort	`medium`; Opus 5 defaulted to `high`
Cache minimum	512 tokens

The 1M context and 128K standard output limits are unchanged from Opus 5. The separate 300K output option documented for the Batch API is a beta capability, not the synchronous limit. This is also not a native image or video generator: visual understanding and generating code that renders graphics are different capabilities.

Opus 5.5 vs Opus 5: the differences that matter

The price cut is real. Input and output rates fall 20%; cache reads fall 60%. That is attractive for an agent that repeatedly reads a large, stable repository context.

Default behavior changes. Leaving effort unspecified now selects medium. At the same named effort level, Opus 5.5 can think more per turn than Opus 5, especially at xhigh and max. Carrying over your old setting does not fix the compute budget.

An API migration needs more than a model-name swap. Thinking cannot be disabled, forced tool selection is rejected, and preserved thinking imposes rules on conversation edits. Those changes matter more to an existing app than another launch-day ranking.

For the previous generation's background, see our Claude Opus 5 review. For mixed workloads, the Claude Code model-selection guide explains when to route routine work to a cheaper model.

Benchmarks: a strong launch with settings to read

Anthropic's launch table reports the following. These are the results published in its announcement, not an OmniaKey rerun or a fresh independent leaderboard snapshot.

Evaluation	Opus 5.5	Opus 5	Fable 5.1
Terminal-Bench 4.0	66.4%	52.3%	55.8%
FrontierCode v1.1 Main	54.4%	48.0%	50.3%
CursorBench 4.0	57.8%	46.6%	51.8%
GDPval-AA v2.1	1846	1708	1735

Opus 5.5 uses adaptive thinking at max unless the announcement says otherwise; its Terminal-Bench result uses xhigh. Anthropic reports a ±2.6-point standard error for that score. Production safeguards were enabled, with fallback models handling some cybersecurity, biology and frontier-model-development tasks. The table therefore does not describe identical, unrestricted conditions for every model.

The default-effort results are especially relevant to a buyer: Anthropic reports 54.6% on FrontierCode and 52.5% on CursorBench at medium. The FrontierCode medium result is slightly above the headline max result, a useful reminder that more effort does not guarantee a higher score on every evaluation.

The sensible conclusion is that Opus 5.5 deserves a trial on difficult agent work. These scores do not establish your latency, failure rate or cost per accepted change. GDPval-AA v2.1 scores also should not be compared directly with older v2 numbers in historical articles.

Claude Opus 5.5 API pricing

All prices below are USD per million tokens, for Anthropic's standard global direct API. They exclude tools, taxes, data-residency modifiers and third-party gateway charges.

Token category	Opus 5.5	Opus 5	Rate reduction
Uncached input	$4	$5	20%
Output, including billed thinking	$20	$25	20%
Cache read	$0.20	$0.50	60%
5-minute cache write	$5	$6.25	20%
1-hour cache write	$8	$10	20%

Two examples show why “20% cheaper” and “40% cheaper” answer different questions. The output amounts below include all billed output, including thinking.

Same token workload	Opus 5.5 calculation	Opus 5.5	Opus 5
100K uncached input + 10K output	0.1 × $4 + 0.01 × $20	$0.60	$0.75
100K uncached + 900K cache-hit input + 10K output	0.1 × $4 + 0.9 × $0.20 + 0.01 × $20	$0.78	$1.20

The second example saves 35%, but excludes the earlier cache write. A cache hit must actually be reported by the provider; resending the same text does not itself prove one occurred.

Anthropic's 40% lower typical task cost combines token rates and token consumption at default settings. It is a vendor measurement, not a universal discount. Its claim of more than 30% faster output is also not a guarantee about end-to-end completion time.

Fast mode is a research preview with $8/$40 input/output rates and an advertised speed of up to 2.5×; the API option is first-party only. Batch processing has $2/$10 input/output rates for asynchronous work. Neither should be confused with ordinary interactive pricing. A Claude subscription and metered API billing are separate purchases; see our Claude Code pricing guide.

API migration: six checks before switching

Remove disabled or manually budgeted thinking. Both thinking.type: "disabled" and thinking.type: "enabled" with budget_tokens return HTTP 400. Omit thinking or use adaptive; control effort with output_config.effort.
Replace forced tool selection. tool_choice types any and tool are unsupported. Use auto or none. strict: true validates a selected tool's arguments; it does not force a call. Use structured outputs when the requirement is schema-valid JSON.
Preserve the conversation. On accounts created from August 31, 2026 at 00:00 UTC, replaying thinking blocks after changing their earlier system prompt, tools or messages can return 400. Keep history append-only and follow the documented preserved-thinking controls when edits are necessary.
Read blocks by type. Progress text between tools now arrives in thinking blocks. The default display: "omitted" leaves that text empty. The migration guide documents display: "summarized" and beta "updates" for interfaces that show progress; pass thinking blocks back unmodified in tool loops.
Check computer-use tools by platform. Claude API and Google Cloud require computer_toolset_20260801 instead of computer_20251124. The earlier tool still works on Amazon Bedrock.
Handle refusals and model switches. HTTP 200 with stop_reason: "refusal" is not a completed answer. A switch back to Opus 5 also does not preserve 5.5's thinking; Fable 5.1 and Mythos 5.1 on the Claude API are the documented exceptions that can read it.

Here is a minimal direct Anthropic API request. It requires an Anthropic key, not an OmniaKey key. It follows the documented contract; we did not execute a paid request.

bash

curl https://api.anthropic.com/v1/messages \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-opus-5-5",
    "max_tokens": 4096,
    "output_config": {"effort": "medium"},
    "messages": [{"role": "user", "content": "Review this migration plan and list its three main risks."}]
  }'

Leave enough max_tokens for both thinking and visible output. A client should collect text blocks rather than assuming content[0] contains the answer.

Who should upgrade, and where can you use it?

Opus 5 users: prioritize testing long refactors, debugging and expensive failures. Run the same tasks at medium and high, keeping tools, permissions and acceptance checks fixed. Record success, retries, tokens, elapsed time and human corrections. Use fresh conversations for clean model comparisons.

Sonnet users: keep routine requests on the cheaper route unless Opus measurably reduces retries or review time. Fable users: test whether Opus 5.5 delivers the required quality at its lower price; the launch table alone is not enough to retire a model that works for your task.

Anthropic lists availability in Claude and Claude Code, the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry. Exact entitlements and regional terms belong to each platform. This review does not establish a live Opus 5.5 route on OmniaKey: check the current model catalog for the exact model ID and price before configuring a gateway.

Frequently asked questions

When was Claude Opus 5.5 released?

September 22, 2026. It is an official release, and its direct API ID is claude-opus-5-5.

How much does Opus 5.5 cost?

Standard direct API rates are $4 input and $20 output per million tokens. Cache reads cost $0.20; 5-minute and 1-hour writes cost $5 and $8. Subscription allowances are separate.

Is Opus 5.5 better than Opus 5 for coding?

Anthropic's published coding evaluations favor 5.5, and its token rates are lower. It is a strong upgrade candidate; test your own tool loops and acceptance criteria before moving production traffic.

Can I turn thinking off?

No. Adaptive thinking is always on. Use a lower effort level when latency or cost is the concern, and test whether quality still meets your requirements.

Does the 1M context window mean unlimited usage?

No. It is a capacity limit, not free tokens. Input, output, caching, tools and account rate limits still apply.

Sources

Explore OmniaKey See available models

Claude Opus 5.5 review

OPUS 5.5