Claude Opus 5.5: What's New, Pricing and Who It's For
Claude Opus 5.5 launched on 22 September 2026 at $4/$20 per million tokens, 20% below Opus 5, with cache reads down 60%. Here's what changed, how it compares to Fable 5.1 and GPT-6 Astra, and which teams should switch.
On this page
Claude Opus 5.5 is Anthropic's new flagship model, released on 22 September 2026. It costs $4 per million input tokens and $20 per million output tokens, 20% less than Opus 5. Cache reads fall 60% to $0.20. On Anthropic's own benchmarks it matches or beats Claude Fable 5.1 on agentic coding and knowledge work, at 40% of Fable's price.
That makes this an unusual launch. Anthropic's recent releases have held the price and added capability: Opus 5 kept Opus 4.8's $5/$25 rate card in July, and Fable 5.1 kept Fable 5's $10/$50 in September. Opus 5.5 is the first Opus price cut since Opus 4.5 took the line from $15/$75 to $5/$25 in November 2025, and it adds capability at the same time. On Anthropic's own numbers, it even beats the model above it in the line-up.
This guide covers what shipped and what it costs per task, in dollars and pounds. It also covers the breaking API changes to check first, and which teams should switch now.
Claude Opus 5.5 At A Glance
Opus 5.5 is the first model in Anthropic's 5.5 family. Anthropic said Sonnet 5.5 and Haiku 5.5 were due "in the coming weeks", and both have since launched: Sonnet 5.5 on 28 September at $2/$10 and Haiku 5.5 on 7 October at $0.10/$0.50. Our roundup of every Claude release in October 2026 covers both. Anthropic's model guide now tells developers to start with Opus 5.5 for most workloads. It suggests Fable 5.1 only when Opus 5.5 at higher effort still falls short.
- Released: 22 September 2026, API ID claude-opus-5-5.
- Price: $4 input and $20 output per million tokens, down from $5/$25 on Opus 5.
- Cache reads: $0.20 per million tokens, 5% of the input rate instead of the usual 10%.
- Batch API: $2/$10, half the standard rate.
- Fast mode: $8/$40, down from $10/$50 on Opus 5.
- Context and output: 1M token context window and 128K max output tokens.
- Thinking: Always on. Effort is the only control, and the default is now medium.
- Availability: Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud Vertex AI and Microsoft Foundry.
- Retirement: Not sooner than 22 September 2027 on Anthropic-operated platforms.
What's New In Opus 5.5
Benchmark Gains Over Opus 5 And Fable 5.1
Anthropic published scores against its own models and OpenAI's. These are Anthropic-reported figures from the Opus 5.5 announcement, not independent tests, so treat them as a guide rather than a verdict. Anthropic ran Opus 5.5 at max effort (xhigh on Terminal-Bench 4.0), and it cautions that in its own use the gap to Fable 5.1 is narrower than these scores suggest.
| Benchmark | Opus 5.5 | Fable 5.1 | Opus 5 | GPT-6 Astra |
|---|---|---|---|---|
| Terminal-Bench 4.0 (agentic coding) | 66.4% | 55.8% | 52.3% | 57.9% |
| FrontierCode v1.1 | 54.4% | 50.3% | 48.0% | 53.3% |
| CursorBench 4.0 | 57.8% | 51.8% | 46.6% | Not reported |
| GDPval-AA v2.1 (knowledge work, Elo) | 1846 | 1735 | 1708 | 1542 |
| OSWorld 2.0 (computer use) | 81.8% | 80.7% | 74.0% | Not reported |
| Humanity's Last Exam | 67.7% | 65.6% | 63.6% | 57.2% |
| AutomationBench | 40.0% | 31.4% | 26.9% | 41.4% |
| Terminal-Bench-Science 0.1 | 58.7% | 52.6% | 29.0% | 64.6% |
The pattern is clear. Opus 5.5 leads on coding, knowledge work and computer use. GPT-6 Astra keeps a narrow lead on AutomationBench and a wider one on scientific terminal tasks. The biggest single jump is Terminal-Bench 4.0, where Opus 5.5 scores 14 points higher than Opus 5.
Faster Output And Fewer Steps
Anthropic says Opus 5.5 generates output more than 30% faster than Opus 5. Early testers report efficiency gains too: GitHub says that in VS Code the model solved more terminal tasks than Opus 5 in less than half the steps.
That second claim matters more for your bill than the speed does. An agent that finishes in fewer turns re-reads its context fewer times and writes fewer tokens.
It also explains Anthropic's headline figure. Anthropic says Opus 5.5 costs 40% less than Opus 5 on typical workloads. Only 20% of that comes from the rate card. The rest depends on the model using fewer tokens, which you should measure on your own traffic before you count on it.
1M Context, 128K Output And Always-On Thinking
The core specs carry over from Opus 5. The 1M token context window bills at standard rates throughout, so a 900K token request costs the same per token as a 9K one. Max output is 128K tokens on the Messages API, or up to 300K on the Batch API with a beta header.
Thinking is the big change. On Opus 5 you could switch thinking off at high effort or below. On Opus 5.5 it is always on, and the effort setting is the only way to control how much the model thinks.
The default effort also drops from high to medium. Any request that does not set effort will now think less than it did on Opus 5.
Claude Opus 5.5 Pricing Explained
Input, Output, Cache And Batch Rates
All rates below are US dollars per million tokens, taken from Anthropic's pricing page on 23 September 2026.
| Rate | Opus 5.5 | Opus 5 | Change |
|---|---|---|---|
| Input | $4.00 | $5.00 | 20% lower |
| Output | $20.00 | $25.00 | 20% lower |
| 5-minute cache write | $5.00 | $6.25 | 20% lower |
| 1-hour cache write | $8.00 | $10.00 | 20% lower |
| Cache read | $0.20 | $0.50 | 60% lower |
| Batch input / output | $2.00 / $10.00 | $2.50 / $12.50 | 20% lower |
| Fast mode input / output | $8.00 / $40.00 | $10.00 / $50.00 | 20% lower |
The cache read rate deserves attention. Most Claude models charge 10% of the input price for a cache hit. Opus 5.5 charges 5%, which is why cache reads fall 60% while everything else falls 20%. If your prompts reuse a large system prompt, codebase or document set, your saving will be bigger than the headline cut.
A few smaller rules apply:
- Fast mode: Research preview, Claude API only, and not available with the Batch API.
- US-only inference: Setting inference_geo to US adds a 1.1x multiplier to every rate.
- Bedrock and Vertex AI: Regional endpoints carry a 10% premium over global endpoints, and those platforms bill you directly.
You can compare every Claude model on our Claude pricing page. The pricing history explorer shows how these rates have moved.
What Opus 5.5 Costs In Pounds
Anthropic bills in US dollars only, so UK teams pay whatever their bank or card charges to convert. At the 22 September 2026 rate of $1.3357 to the pound, Opus 5.5 works out as follows:
- Input: About £2.99 per million tokens.
- Output: About £14.97 per million tokens.
- Cache read: About £0.15 per million tokens.
- Batch: About £1.50 input and £7.49 output.
- Fast mode: About £5.99 input and £29.95 output.
For comparison, Opus 5 comes to roughly £3.74/£18.72. UK VAT may apply to your invoice, depending on how your account is set up. A weaker pound also raises your real cost, even when Anthropic's prices stay the same.
Real-World Cost Examples: A Coding Agent Vs A Support Chatbot
Per-token prices only tell you so much. Here are two worked examples at list price. In both, we charge fresh input at the base rate and leave out cache-write premiums for simplicity. Those premiums add a few percent in practice.
Coding agent task: 1.5M input tokens across the whole task, 90% of them cache reads, plus 100K output tokens. On Opus 5.5 that is 150K fresh input ($0.60), 1.35M cache reads ($0.27) and 100K output ($2.00).
Support chatbot: 100,000 conversations a month. Each one has a 6K token cached system prompt and knowledge base, 2K tokens of fresh input and a 500 token reply.
| Model | Coding agent, per task | Support chatbot, per month |
|---|---|---|
| Claude Opus 5.5 | $2.87 (£2.15) | $1,920 (£1,437) |
| Claude Opus 5 | $3.93 (£2.94) | $2,550 (£1,909) |
| Claude Fable 5.1 | $6.84 (£5.12) | $4,650 (£3,481) |
| GPT-6 Astra | $7.85 (£5.88) | $5,100 (£3,818) |
| Claude Sonnet 5 | $1.57 (£1.18) | $1,020 (£764) |
| Claude Haiku 4.5 | $0.79 (£0.59) | $510 (£382) |
At identical token counts, Opus 5.5 is 27% cheaper than Opus 5 for the agent and 25% cheaper for the chatbot. The agent saves more because it leans harder on cache reads. If Opus 5.5 also completes the task in fewer steps, as Anthropic and early testers report, the gap widens towards the 40% it quotes.
The chatbot row carries a warning, though. Thinking cannot be switched off on Opus 5.5, so every reply adds thinking tokens on top of the visible 500. At low effort that overhead is small, but it is never zero.
For short, high-volume replies, Sonnet 5 or Haiku 4.5 will usually be the better buy. Try your own numbers in the AI pricing calculator.
Opus 5.5 Vs Opus 5, Fable 5.1 And GPT-6 Astra
Against Opus 5, the choice is simple. Opus 5.5 is cheaper on every rate, faster, and ahead on every benchmark Anthropic published. The only reasons to wait are the breaking changes covered below. Opus 5 remains available as a legacy model while you migrate.
Against Fable 5.1, Opus 5.5 costs 60% less per token and leads on Anthropic's coding, knowledge work and computer use figures.
Fable 5.1 still has two real advantages. Its cache reads are cheaper still at $0.25 per million, just 2.5% of its input rate. Anthropic also positions it as the model for the hardest long-horizon reasoning. If your evals show Opus 5.5 falling short at high effort, Fable 5.1 is the step up.
Against GPT-6 Astra, Opus 5.5 is 60% cheaper per token and leads on most of the shared benchmarks. Astra launched on 3 September at $10/$50.
Astra's cached input costs $1.00, five times Opus 5.5's rate, so heavily cached workloads favour Claude even more. Astra's leads on AutomationBench and scientific tasks are worth testing if that is your domain. Our GPT vs Claude comparison and AI API comparison cover the wider field.
Breaking Changes To Check Before You Migrate
Moving from Opus 5 is more than a model name swap. Anthropic's Opus 5.5 migration guide lists four breaking changes:
- Thinking can't be disabled: Both thinking: {"type": "disabled"} and manual thinking budgets return a 400 error. Remove the field and set an effort level instead.
- Forced tool use is not supported: tool_choice types any and tool return a 400 error. Use auto with strict tool use or structured outputs, and say in the prompt when the tool should be used.
- Thinking blocks are tied to the model: Only Fable 5.1 and Mythos 5.1 can read Opus 5.5's thinking blocks. A router that hands a conversation to another model loses them, and editing earlier turns mid-conversation can return a 400 error on newer accounts.
- Old computer use tool rejected: On the Claude API and Google Cloud, computer_20251124 no longer works. Use the computer_toolset_20260801 toolset. Bedrock is unaffected.
Three quieter changes will not break requests but can surprise you:
- Default effort is now medium: Set effort explicitly and re-run your quality and cost tests.
- Narration moved into thinking blocks: Text the model writes between tool calls now arrives as thinking blocks, empty by default. Apps that show progress updates to users will go quiet unless you change the display setting.
- Broader refusals: Safety classifiers now cover bio and reasoning extraction as well as cyber.
On that last point, a refusal is not a silent swap. You get a normal response with stop_reason: "refusal". Anthropic offers an opt-in server-side fallback, currently in beta, that retries the request on a recommended model. It says most cybersecurity tasks will go to Opus 4.8.
Anthropic's refusals and fallback documentation says each attempt that produces output is billed at the rates of the model that ran it. A refusal before any output is free. If you run security tooling, budget for Opus 4.8's $5/$25 on those turns.
For step-by-step help with model upgrades, see our migration guides.
What Changes For Claude Pro And Max Subscribers
If you use Claude through a subscription rather than the API, per-token prices do not affect you directly. Two things do.
- Higher five-hour limits: Anthropic has raised five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans. It has not published a percentage.
- A saveable reset: Subscribers get a one-off rate limit reset to save and use whenever they choose. It is handy on a heavy deadline day.
Because Opus 5.5 is cheaper to run and needs fewer steps, your allowance should also stretch further on long coding or research sessions. If you are weighing a plan against API usage, our cost per seat tool compares the two.
Who Claude Opus 5.5 Is For (And Who Should Stick With Sonnet)
Switch now if you run:
- Coding agents: The Terminal-Bench and CursorBench gains are the largest in the release, and cheaper cache reads suit long agent sessions.
- Knowledge work and research: Opus 5.5 leads on GDPval-AA and Humanity's Last Exam at well under Fable's price.
- Computer use and browser agents: It edges out Fable 5.1 on OSWorld 2.0, once you have moved to the new toolset.
- Anything currently on Fable 5.1 for cost reasons: Test Opus 5.5 first. For many workloads it will do the same job for 40% of the price.
Stick with Sonnet 5 or Haiku 4.5 if you run:
- High-volume, short-reply chat: Sonnet 5 at $2/$10 costs about half as much and responds faster.
- Classification, extraction and routing: Haiku 4.5 at $1/$5 is a quarter of Opus 5.5's price and fast enough for real-time use.
- Latency-critical features: Anthropic rates Opus 5.5's latency as moderate, against fast for Sonnet 5 and fastest for Haiku 4.5.
It is also worth waiting a few weeks if Sonnet is your main model. Anthropic has said Sonnet 5.5 and Haiku 5.5 are coming soon, and we will cover them on our Anthropic provider page when they land. Follow our price changes page to see the moment they do.
Frequently Asked Questions
Is Claude Opus 5.5 Cheaper Than Opus 5?
Yes. Opus 5.5 costs $4/$20 per million input and output tokens against $5/$25 for Opus 5, a 20% cut. Cache reads fall 60% to $0.20. Anthropic says typical workloads cost 40% less overall because the model also uses fewer tokens per task, though that part depends on your workload.
Can You Turn Off Thinking In Opus 5.5?
No. Thinking is always on, and requests that try to disable it return a 400 error. Use the effort setting to control how much the model thinks. Low effort is the closest equivalent to thinking disabled on earlier models.
Is Opus 5.5 Available On Amazon Bedrock And Google Vertex AI?
Yes. Opus 5.5 launched on the Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud Vertex AI and Microsoft Foundry on the same day. Fast mode is only available on the Claude API, and Bedrock and Google Cloud set their own pricing and retirement dates.
When Was Claude Opus 5.5 Released?
Anthropic released Claude Opus 5.5 on 22 September 2026, two months after Opus 5 launched on 24 July. You can read what Opus 5 introduced in our Claude Opus 5 guide. Our September 2026 pricing roundup covers every other price change this month.
Fun Fact: The Cheaper Model Beat The Flagship
Opus 5.5 costs 40% of Fable 5.1's per-token price, yet on Anthropic's own Terminal-Bench 4.0 results it scores 66.4% to Fable 5.1's 55.8%. It also beats GPT-6 Astra's 57.9%, a model that costs 2.5 times as much. Anthropic's own model guide now tells developers to start with Opus, not its premium Fable tier.
The Bottom Line
Claude Opus 5.5 is the easiest upgrade decision Anthropic has offered this year. It is cheaper than Opus 5 on every rate and faster. It also beats Fable 5.1 on most of the work people buy Opus for. The catch is migration work: thinking you cannot turn off, no forced tool use, and a new computer use toolset.
Plan for a short test cycle. Set effort explicitly, re-run your evals, and measure cost per task rather than cost per token. Then use the Compare AI pricing calculator to see what Opus 5.5 would save on your actual traffic.
