What it is
Claude Opus 5.5 is Anthropic's main high-end model, released on September 22, 2026. Anthropic's docs say to start with it for most workloads. It reads text and images, writes text, and has a 1M-token context window with up to 128K output tokens.
It always thinks before answering ("adaptive thinking"), and you steer how much with an effort setting that defaults to medium. The API price is $4 per million input tokens and $20 per million output tokens, down from $5/$25 on Opus 5. A faster mode costs $8/$40.
It is available through the Claude API, the Claude apps, Amazon Bedrock, Google Cloud and Microsoft Foundry.
Our review
Opus 5.5 is the model to beat for long agent jobs, and for once the independent numbers agree with the vendor's. Artificial Analysis puts it first on its Intelligence Index at 58 on max effort, ahead of Sonnet 5.5 at 56 and both GPT-6 Astra and Claude Fable 5.1 at 53. Anthropic's own table has it at 66.4% on Terminal-Bench 4.0 against 57.9% for Astra, though Astra still wins AutomationBench (41.4% to 40.0%) and Terminal-Bench-Science (64.6% to 58.7%).
The price cut is real, but the larger saving is meant to come from tokens. List rates fell from $5/$25 to $4/$20 and cache reads from $0.50 to $0.20, and Anthropic claims typical workloads cost 40% less than on Opus 5 because the model writes less. That is self-reported. Artificial Analysis offers a check: an average task on its index cost $1.34 at medium effort and $5.98 at max, so effort is the dial that sets the bill. GPT-6.1 Sol scored 48 at medium for $0.21 a task, which is the comparison a cost-conscious buyer has to make.
Moving from Opus 5 is not a drop-in change. The docs list four breaking changes: thinking cannot be turned off, forced tool use returns an error, thinking blocks are tied to the model and conversation that produced them, and the older computer_20251124 tool is rejected on the Claude API. A fifth change fails nothing but alters behaviour. Text between tool calls now arrives inside thinking blocks that are empty by default, so an app that streams progress to users goes quiet until it changes the display setting.
Security teams should read the safeguards before buying. Anthropic says most cybersecurity tasks are re-routed to Opus 4.8 unless the account is in its Cyber Verification Program, which Opus 5.5 was due to join "in the coming weeks" at launch. Anthropic also concedes that safeguard fallbacks probably pulled its own scores down, and Zapier's AutomationBench counted those interventions as failures. For a model sold on long autonomous runs, a hand-off to an older model mid-task is the failure mode to measure on your own traffic.
Where it earns its place
A platform team running an in-house coding agent over a 200,000-line codebase
Run overnight audits and multi-file refactors without someone steering each step
Run at medium effort and raise it only for tasks that fail. Anthropic says a 200,000-line audit took under 3 hours against over 20 on Opus 5 (vendor figure). With 90% of a 200K-token context cached, a turn that writes 20K tokens costs about $0.52.
A deal team at a bank or law firm
Analyse a full data room or a stack of long contracts in one pass
The 1M-token window, about 555,000 words, is billed at the standard per-token rate, so a data room fits without chunking. Anthropic cites a merger analysis that finished in 63 minutes rather than 93 at half the cost (vendor figure). Zero data retention is available by agreement for regulated material.
A head of engineering paying for Claude Fable 5.1
Cut the model bill without losing agent quality
Fable 5.1 costs $10/$50 and scored 53 on Artificial Analysis's index against Opus 5.5's 58, and Anthropic itself recommends starting with Opus 5.5. A $20,000-a-month Fable workload moved over at equal token counts becomes $8,000, before any token savings.
Before you commit
- Setup
- Model ID claude-opus-5-5 on the Claude API (anthropic.claude-opus-5-5 on Bedrock). Opus 5 code needs the migration guide: drop forced tool_choice, stop disabling thinking, and set a display value if you stream text between tool calls.
- Integrations
- Messages API with official SDKs, Batch API and prompt caching (512-token minimum). Fast mode gives up to 2.5x output speed on the Claude API, Google Cloud (Anthropic managed) and Azure-hosted Foundry, but not with Batch or Priority Tier.
- Data and privacy
- Anthropic says retained API data is never used for training without express permission; standard commercial retention deletes inputs and outputs within 30 days, and zero data retention is available by agreement. On Bedrock and Google-managed Vertex the cloud provider is the data processor.
- Cost at scale
- A coding-agent turn with 200K tokens of context, 90% cached, plus 20K output costs about $0.52, so 1,000 turns a day is about $15,500 a month; Batch halves anything that can wait. Artificial Analysis measured $1.34 a task at medium effort and $5.98 at max. Fast mode doubles the rate to $8/$40.
- Limits
- Start-tier organisations get 2M input and 400K output tokens a minute and a $500 monthly spend cap; Scale gets 10M/2M and $200,000. Cache reads do not count toward input rate limits. Retirement not before 22 September 2027.
What's good
- First on Artificial Analysis's Intelligence Index (58 at max effort)
- $4/$20 with $0.20 cache reads, down from $5/$25 and $0.50 on Opus 5
- Full 1M context at the standard rate; 128K output, 300K on Batch (beta)
- Same model ID on the Claude API, Bedrock, Google Cloud and Microsoft Foundry
What's not
- Thinking cannot be disabled, so short replies still pay for reasoning
- Four breaking API changes from Opus 5, plus a quiet change to streamed text
- Most cybersecurity work re-routes to Opus 4.8 without verification
- Several times GPT-6.1 Sol's cost per task on Artificial Analysis's index
Best for
Teams running long coding agents, code audits or document-heavy research where finishing the task matters more than the cost of each call.
Alternatives
Claude Opus 5.5: common questions
- Is Claude Opus 5.5 worth it?
- Anthropic's new default earns the title: first on Artificial Analysis's independent index, and a fifth cheaper per token than the Opus it replaces. We rate it 4.7 out of 5.
- How much does Claude Opus 5.5 cost?
- API: $4 per 1M input tokens, $20 per 1M output; cache reads $0.20, 5-minute cache writes $5; Batch API 50% off; fast mode $8/$40; US-only inference 1.1x.
- What are the downsides of Claude Opus 5.5?
- Thinking cannot be disabled, so short replies still pay for reasoning; four breaking API changes from Opus 5, plus a quiet change to streamed text; most cybersecurity work re-routes to Opus 4.8 without verification; several times GPT-6.1 Sol's cost per task on Artificial Analysis's index.
- Who is Claude Opus 5.5 best for?
- Teams running long coding agents, code audits or document-heavy research where finishing the task matters more than the cost of each call.
- How much does Claude Opus 5.5 cost for a team or in production?
- A coding-agent turn with 200K tokens of context, 90% cached, plus 20K output costs about $0.52, so 1,000 turns a day is about $15,500 a month; Batch halves anything that can wait. Artificial Analysis measured $1.34 a task at medium effort and $5.98 at max. Fast mode doubles the rate to $8/$40.
- Is Claude Opus 5.5 safe for company data?
- Anthropic says retained API data is never used for training without express permission; standard commercial retention deletes inputs and outputs within 30 days, and zero data retention is available by agreement. On Bedrock and Google-managed Vertex the cloud provider is the data processor.
- What are the best Claude Opus 5.5 alternatives?
- The closest alternatives we have reviewed are Claude Fable 5.1, Claude Sonnet 5.5, GPT-6 Astra, GPT-6.1 Sol and Gemini 4 Argon.
Answers checked Oct 22, 2026.
Make Claude Opus 5.5? Show this rating on your site
