Claude Opus 5.5 is Anthropic’s newest flagship model and the first release in its Claude 5.5 family, launched on 22 September 2026. It costs $4 per million input tokens and $20 per million output tokens, 20% less than Opus 5, and it scores 58 on the independent Artificial Analysis Intelligence Index at max effort.
It is available on Claude Pro, Max, Team and Enterprise plans, but not on the Free plan. Anthropic shipped it about an hour before OpenAI launched GPT-6 Sol and GPT-6 Luna. Prices, limits and plan details below are correct as of 24 September 2026.
What is Claude Opus 5.5?
Claude Opus 5.5 is Anthropic‘s top general-access model, released on 22 September 2026 as the first member of the Claude 5.5 family. It takes text and images as input, returns text, has a June 2026 knowledge cutoff and a 1M-token context window. Anthropic positions it as its everyday model for agentic coding and knowledge work.
In the Claude Opus 5.5 launch announcement, Anthropic says it “costs 40% less to run than Opus 5 on typical workloads” and performs at the level of Fable 5.1 on most work (Anthropic-reported). Fable 5.1 costs 2.5 times as much per token.
Opus 5 arrived on 24 July 2026. The API model ID is claude-opus-5-5 (anthropic.claude-opus-5-5 on Bedrock). Among other generative AI tools, treat it as Anthropic’s default pick.
What changed in Opus 5.5 vs Opus 5?
Opus 5.5 is cheaper per token, faster and scores higher than Opus 5, but it behaves differently by default. Token prices fell 20%, cache reads fell 60%, and output is over 30% faster (Anthropic-reported). The default effort level dropped from high to medium, and thinking can no longer be switched off.
| Item | Claude Opus 5 | Claude Opus 5.5 |
|---|---|---|
| Input price (per 1M tokens) | $5 | $4 |
| Output price (per 1M tokens) | $25 | $20 |
| Cache read (per 1M tokens) | $0.50 | $0.20 |
| Cache write, 5-minute | $6.25 | $5 |
| Default effort | High | Medium |
| Thinking switched off | Allowed | Not allowed (400 error) |
| Terminal-Bench 4.0 | 52.3% (Anthropic-reported) | 66.4% (Anthropic-reported) |
| GDPval-AA v2.1 (Elo) | 1708 (Anthropic-reported) | 1846 (Anthropic-reported) |
| Artificial Analysis Index (max) | 51 | 58 |
| Output tokens per AA task (max) | About 73K | About 119K |
This is the first Opus price cut, yet at max effort the model writes far more tokens per task, so Artificial Analysis found the cost per task roughly level with Opus 5.
How much does Claude Opus 5.5 cost?
Opus 5.5 pricing on the Claude API is $4 per million input tokens and $20 per million output tokens. Cache reads cost $0.20, 5-minute cache writes $5 and 1-hour cache writes $8. Batch processing halves the base rates, and fast mode costs $8 input and $40 output. There is no long-context surcharge.
| Price item (per 1M tokens) | Opus 5.5 | What to know |
|---|---|---|
| Input | $4 | Same rate across 1M window |
| Output | $20 | Includes thinking tokens |
| Cache read | $0.20 | 0.05x base input |
| Cache write, 5-minute | $5 | 1.25x base input |
| Cache write, 1-hour | $8 | 2x base input |
| Batch API | $2 input / $10 output | 50% off |
| Fast mode | $8 input / $40 output | Research preview, Claude API only |
| US-only inference | 1.1x every category | Data residency |
Figures come from Anthropic’s API pricing documentation. Fast mode promises up to 2.5 times the speed (Anthropic-reported), but is not offered on Bedrock, Google Cloud or Foundry.
The “40% cheaper” headline is per task, not per token: it assumes Opus 5.5 also uses fewer tokens. Artificial Analysis’s Opus 5.5 evaluation found the opposite at max effort, so your saving depends on the effort level you run.
Pick The Right Cache Length
If your agent pauses for more than five minutes between calls, one 1-hour cache write at $8 beats paying $5 twice to rewrite an expired 5-minute cache. For fast loops, stay on the 5-minute write. Prompts under 512 tokens cannot be cached at all.
What is the Opus 5.5 context window and max output?
Claude Opus 5.5 has a 1M-token context window by default and a maximum output of 128K tokens per response. On the Batch API, output can rise to 300K tokens with a beta header. The whole 1M window is billed at standard rates, so long documents and large codebases cost no premium.
In Claude Code, the 1M window applies on every plan, including Pro, and auto-compaction starts at about 967K tokens.
Can you turn off thinking in Claude Opus 5.5?
No. Adaptive thinking is always on in Opus 5.5 and cannot be disabled. Sending a request with thinking set to disabled, or with a fixed thinking budget, returns a 400 error. Instead, you control how hard the model thinks with five effort levels: low, medium, high, xhigh and max.
The default is now medium, where Opus 5 defaulted to high, so requests that omit effort run one level lower than before.
Forced tool use also returns a 400 error now. In Claude Code, Opus 5.5 needs v2.1.280 or later, a common reason it is missing from the model picker. For alternatives, see our guide to AI tools for developers.
How does Claude Opus 5.5 perform on benchmarks?
On Anthropic’s own table, Opus 5.5 leads on seven of nine benchmarks, including Terminal-Bench 4.0 at 66.4% and GDPval-AA v2.1 at 1846 Elo (Anthropic-reported). OpenAI’s GPT-6 Astra beats it on AutomationBench and Terminal-Bench-Science. All of Anthropic’s Opus 5.5 results use adaptive thinking at max effort.
Astra’s leads are narrow on AutomationBench (41.4% vs 40.0%) and wider on Terminal-Bench-Science (64.6% vs 58.7%). Anthropic’s table compares against GPT-5.6 Sol, not GPT-6 Sol, which launched later that day.
Do independent tests back up Anthropic’s numbers?
Mostly, though independent scores run lower. Artificial Analysis scores Opus 5.5 at 58 on its Intelligence Index v4.3.2 at max effort, ahead of GPT-6 Astra and Fable 5.1 on 53. At the default medium effort it scores about 51, and at low effort 42.
Read more: GPT-6 Sol vs Luna vs Astra
Cost per Index task runs from $0.55 at low effort to $1.34 at medium and $5.98 at max. Independent re-runs of Terminal-Bench 4.0 came in below Anthropic’s 66.4%: Artificial Analysis measured 59.6% and Vals AI 61.6%.
Vals AI ranks Opus 5.5 first overall but lower on legal research, tax and medical coding. Anthropic’s system card also admits a regression: the model is “more likely than previous models to follow malicious instructions in text that a user pastes into their own prompt.”
Is Claude Opus 5.5 free, and which plans get it?
No, Claude Opus 5.5 is not free. It is included on Pro, Max, Team and seat-based Enterprise plans, and the Free plan only gets Sonnet and Haiku models. Pro costs $20 a month, or $17 a month billed annually. Unlike Fable 5.1 on Pro, Opus 5.5 needs no extra usage credits.
Read more: Claude Pricing 2026
Outside claude.ai, it runs on the Claude API, Amazon Bedrock, Google Cloud’s Vertex AI and Microsoft Foundry. GitHub Copilot is rolling it out on Pro+, Max, Business and Enterprise plans at provider list pricing. It is also the default model in Claude Code.
What changed with Claude Opus 5.5 usage limits?
Anthropic raised five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans, but gave no percentage for claude.ai chat. Anthropic’s Claude Code developer account said five-hour session limits in Claude Code rose 20%, and that the lower price lets Opus 5.5 go 25% further within those limits.
Subscribers also got a one-off rate-limit reset. According to MacRumors’ Opus 5.5 launch report, it can be used until 22 October. Anthropic’s help centre says it sits under Settings > Usage on web and desktop, not mobile or Claude Code, and is lost if you downgrade first.
Why did Claude switch models in my conversation?
Anthropic rates Opus 5.5 as comparable to Claude Mythos 5.1 in cybersecurity and biology, so it applies Fable 5.1-style safeguards. Flagged cybersecurity requests fall back to Opus 4.8, while flagged biology and frontier AI development requests fall back to Opus 5. On claude.ai you see a notice and a label showing which model answered.
The wording varies. The announcement says “most cybersecurity tasks will be re-routed to Opus 4.8”, while Anthropic’s help article on model switching describes “a narrow set of higher-risk requests”, such as exploit generation and penetration testing.
On the API, fallback is an opt-in beta; without it, a flagged request returns a refusal. Anthropic says its Cyber Verification Program will cover Opus 5.5 “soon”, but it does not yet. Life sciences organisations can apply to the new Life Sciences Verification Program.
How does Opus 5.5 compare with Fable 5.1 and GPT-6 Sol?
Opus 5.5 beats Fable 5.1 on every row of Anthropic’s own table and on Artificial Analysis (58 vs 53), at 40% of Fable’s per-token price. GPT-6 Sol is half Opus 5.5’s per-token price, at $2 input and $10 output, but scores lower on Artificial Analysis at 48 (max).
Fable 5.1 still uses fewer tokens per task, and some reviewers prefer it for the hardest coding problems. Our Claude Opus 5.5 vs Fable 5.1 comparison covers when Fable is worth paying for, and our Claude Fable explainer covers its plan rules.
Sol’s edge shrinks on cache-heavy work, since both charge $0.20 per million cached tokens. Our GPT-6 Sol vs Claude Opus 5.5 head-to-head has the full breakdown.
When is Sonnet 5.5 coming out?
Anthropic has not given a Sonnet 5.5 release date. It says Claude Sonnet 5.5 and Haiku 5.5 will arrive “in the coming weeks” with many of the same improvements to performance, efficiency and safety as Opus 5.5. No pricing has been announced for either, as of 24 September 2026.
What are reviewers saying about Claude Opus 5.5?
Most reviewers praise its writing and long coding sessions, and several have made it their default model. Some call it a smaller Fable. The main complaints are heavy token use at max effort, mandatory thinking, cyber and biology routing, and weaknesses on deadline-bound tasks and brand-guideline adherence.
In Every’s Opus 5.5 vibe check, Katie Parrott called it “a pleasure to write with… takes feedback without a fight.” Dan Shipper described it as a “smaller Fable… faster and cheaper and therefore more usable for day-to-day work.” The review also found it buries key points, manages deadlines poorly and follows brand guidelines inconsistently. Writers should test it against other AI tools for content creation.
Zvi Mowshowitz’s system card analysis concludes: “You want your new default model to be Opus 5.5.” He also disputes Anthropic’s cyber classification, arguing “This is a Tier 2 cyber model.”
Simon Willison’s launch-day write-up shows the max-effort downside: his test hit the 128K limit “while it was still reasoning”. He still made it one of his defaults.
Should you switch to Claude Opus 5.5?
Switch if you use Opus 5 or pay for Fable 5.1 on general work: Opus 5.5 is cheaper per token and scores higher on both vendor and independent tests. Hold off if your work centres on offensive security, which is routed to Opus 4.8, or if you need the Free plan.
If your pipelines paste untrusted text into prompts, add your own checks, given the system card’s prompt-injection regression.
Pin Effort Before You Migrate
Code that called Opus 5 without an effort setting now runs at medium. Set effort explicitly, sweep low, medium and high on your own evals, and raise max_tokens, because that limit covers thinking as well as the answer. Keep max for tasks that already failed at xhigh.
Conclusion
Claude Opus 5.5 is the model most Claude users should now default to: cheaper than Opus 5 and ahead of Fable 5.1 on published and independent tests. The trade-offs are mandatory thinking, heavy token use at max effort, and cyber and biology requests that fall back to older models.
Your next step: on a paid plan, check your reset in Settings > Usage before the reported 22 October expiry, then run your three most common tasks at medium and high effort and keep the cheaper setting that passes. For a wider stack, compare options in our AI tools directory.
FAQ
Q1. Is Claude Opus 5.5 better than Claude Fable 5.1?
On most published measures, yes. Opus 5.5 beats Fable 5.1 on every row of Anthropic’s own benchmark table and scores 58 against 53 on the Artificial Analysis Intelligence Index. It also costs $4 input and $20 output per million tokens, against $10 and $50 for Fable 5.1. Fable still uses fewer tokens per task, and some reviewers prefer it for the hardest coding problems.
Q2. Is Opus 5.5 available in GitHub Copilot?
Yes. GitHub is rolling Claude Opus 5.5 out gradually to Copilot Pro+, Max, Business and Enterprise plans, billed at provider list pricing. It is also available on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud Vertex AI and Microsoft Foundry. Fast mode, however, is only offered on the Claude API, as of 24 September 2026.
Q3. Why isn't Opus 5.5 showing up in Claude Code?
The most common reason is an outdated install. Claude Opus 5.5 needs Claude Code version 2.1.280 or later, so update first. Once updated, it becomes the default model for Pro, Max, Team, Enterprise and Anthropic API users. Note that it defaults to medium effort in Claude Code, and the legacy top-level effortLevel setting does not apply to it.
Q4. How do I use the Claude rate limit reset, and when does it expire?
On claude.ai web or desktop, go to Settings, then Usage, and select the reset button. It sets either your five-hour or weekly limit back to full, and it cannot be undone. It is not available on mobile or in Claude Code. MacRumors reports the reset can be used until 22 October 2026, and you lose it if you downgrade or cancel first.
Q5. What is the Life Sciences Verification Program?
It is Anthropic’s vetting programme for organisations such as academic labs, startups and pharmaceutical companies whose biology work is blocked by Opus 5.5’s safeguards. Announced on 17 September 2026, it has a Standard Use tier and a project-specific High-risk Use add-on that removes the biology safeguards. Without it, flagged biology requests fall back to Claude Opus 5.
Q6. Is Opus 5.5 max effort worth it?
For most work, no. Artificial Analysis measured about $5.98 per Intelligence Index task at max effort against $1.82 at high and $1.34 at the default medium, while the score rises only from 51 to 58. Simon Willison saw max-effort runs hit the 128K output limit mid-reasoning. Start at medium or high and move up only for tasks that fail.
Q7. Does Claude Opus 5.5 cost extra for long prompts?
No. Anthropic bills the whole 1M-token context window at standard rates, so there is no long-context surcharge: $4 per million input tokens and $20 per million output tokens throughout. The only multipliers are optional ones, such as US-only inference at 1.1x and fast mode at $8 input and $40 output. Batch processing halves the standard rates.
