Anthropic released Claude Opus 5.5 on September 22, 2026, making it the first model in the new Claude 5.5 family. The launch is unusually important because Anthropic is not just claiming a capability gain: it also cut the base API price to $4 per million input tokens and $20 per million output tokens, reduced cache-read pricing to $0.20 per million tokens, and says the model generates output more than 30% faster than Claude Opus 5.
Anthropic says Opus 5.5 performs at roughly the level of Claude Fable 5.1 on most work while costing about 40% less than Opus 5 on typical workloads. That 40% figure is not simply the 20% list-price cut; Anthropic says Opus 5.5 also uses fewer tokens per task in its tests. AI-XBlog has not independently benchmarked Opus 5.5, so vendor benchmark and efficiency claims below are labeled as Anthropic-reported unless another source is named.
For developers, the production Claude Platform model ID is claude-opus-5-5. Anthropic says the model is available on all platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure.
For Claude subscription users, the launch also changes practical access: Anthropic says it is increasing the five-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans. Subscription users also receive a rate-limit reset that can be saved and used when they choose. Anthropic has not published one universal numeric multiplier for those plan limits, so this page does not infer a new message or token quota.
Claude Opus 5.5 at a glance
| Item | Claude Opus 5.5 |
|---|---|
| Release date | September 22, 2026 |
| Claude Platform model ID | claude-opus-5-5 |
| Base input price | $4 / 1M tokens |
| Base output price | $20 / 1M tokens |
| 5-minute cache write | $5 / 1M tokens |
| Cache read | $0.20 / 1M tokens |
| Fast mode | $8 input / $40 output per 1M tokens, up to 2.5× speed |
| Output speed | Anthropic says more than 30% faster than Opus 5 |
| Typical-workload cost claim | Anthropic says about 40% lower than Opus 5 |
| Availability | Anthropic platforms plus AWS, Google Cloud, and Microsoft Azure |
| Subscription usage | Anthropic says five-hour limits are increasing on Pro, Max, Team, and seat-based Enterprise; subscription users also get a saveable rate-limit reset |
| Safety routing | Some cyber requests route to Opus 4.8; biology and frontier-LLM requests can route to Opus 5 |
What changed from Claude Opus 5?
The commercial story is stronger than a normal point release. Opus 5.5 lowers both the list price and the amount of compute Anthropic says it needs to complete difficult work.
| Area | Claude Opus 5 | Claude Opus 5.5 | Why it matters |
|---|---|---|---|
| Input price | $5 / 1M | $4 / 1M | 20% lower list price |
| Output price | $25 / 1M | $20 / 1M | 20% lower list price |
| Cache read | $0.50 / 1M | $0.20 / 1M | 60% lower, especially relevant to coding and agent loops |
| 5-minute cache write | $6.25 / 1M | $5 / 1M | Lower prompt-caching setup cost |
| Output speed | Previous baseline | More than 30% faster, per Anthropic | Potentially lower wall-clock latency for long jobs |
| Safeguards | Earlier Opus safety stack | Fable-class safeguards for cyber, biology, and distillation | More requests may transparently route to another Claude model |
For teams already using Opus 5 through the API or Claude Code, the biggest practical reason to evaluate 5.5 is not the version number. It is the combination of lower token rates, much cheaper cache reads, faster output, and Anthropic’s claim that the model needs fewer tokens to finish complex tasks.
Claude Opus 5.5 API pricing
Anthropic’s launch pricing is $4 per million input tokens and $20 per million output tokens. Cache reads cost $0.20 per million tokens, down from $0.50 on Opus 5; 5-minute cache writes cost $5 per million tokens, 1-hour cache writes cost $8 per million tokens, and Batch API input/output token pricing is 50% off the standard rates.
Refusal and fallback billing matters for real API cost. Opus 5.5 can return a successful HTTP 200 response with stop_reason: "refusal". According to Anthropic’s refusals and fallback documentation, pre-output refusals in the bio, frontier_llm, and reasoning_extraction categories are billed at normal model rates; other pre-output refusal categories are not billed, while mid-stream refusals bill the input plus any output already streamed. A fallback request can add a second model charge when the triggering refusal is billable, although Anthropic’s fallback credit offsets the extra prompt-cache setup cost. For production budgeting, treat this as part of Opus 5.5’s execution economics rather than assuming every blocked request costs $0.
Fast mode is also available in Claude Code and the Claude Platform. Anthropic lists it at $8 per million input tokens and $40 per million output tokens, with speeds up to 2.5× the standard mode.
The phrase “40% cheaper” needs context. The token list price is 20% lower than Opus 5. Anthropic’s larger 40% typical-workload estimate includes the company’s observation that Opus 5.5 uses fewer tokens on many long, agentic tasks. That means real savings depend on workload shape. A short chat request may mainly reflect the 20% rate cut; a long coding agent that repeatedly reuses cached context can benefit more from the cache-read cut and lower token use.
Why the cache-read price matters for coding agents
The most consequential line item may be the cache read rate rather than the headline input price. Coding agents repeatedly send large repositories, system instructions, tool definitions, project memory, and previous context. Prompt caching lets those repeated tokens cost less after the initial write.
Dropping cache reads from $0.50 to $0.20 per million tokens is a 60% reduction. For large Claude Code sessions or long-running agents, that can have a larger effect on effective cost than a simple 20% cut to uncached input. Our Claude Code pricing guide explains how subscription usage, credits, API billing, and caching interact.
Model ID and availability
Anthropic’s launch page gives the direct Claude Platform model ID as claude-opus-5-5. The company says Opus 5.5 is available on all platforms and specifically names Amazon Web Services, Google Cloud, and Microsoft Azure.
Anthropic’s current platform pricing documentation says Claude 4.6 and later models use the company’s full 1M-token long-context policy at standard pricing. Anthropic’s dedicated Opus 5.5 model page now confirms a 1M-token context window, a 128K-token maximum output, and a June 2026 knowledge cutoff. The Batch API can support up to 300K output tokens in beta with Anthropic’s documented beta header.
Developer breaking changes from Opus 5
Opus 5.5 is not a drop-in model-ID swap for every API integration. Anthropic documents four breaking changes for code moving from Opus 5: adaptive thinking is always on and cannot be disabled; forced tool choice using any or a named tool now returns a 400 error; thinking blocks are bound to the model and conversation; and the older computer_20251124 computer-use tool is not accepted on the Claude API or Google Cloud. Anthropic also notes a response-shape change: text between tool calls is returned inside thinking blocks, so applications that stream that text as progress updates may need to set an appropriate thinking.display value.
For migrations, Anthropic recommends changing the model ID to claude-opus-5-5, removing disabled/manual thinking settings in favor of adaptive thinking plus the effort parameter, replacing forced tool choice with auto plus strict tool use where schema compliance matters, and updating older computer-use integrations. These changes are operationally important enough to test before switching a production agent.
What Anthropic says about performance
Anthropic reports strong gains in agentic coding, computer use, and professional knowledge work. The company also explicitly warns that small benchmark margins are becoming less reliable indicators of real-world differences at this capability level. That is the right way to read these launch numbers: as evidence of where Anthropic optimized the model, not as proof that Opus 5.5 is universally best.
| Benchmark | Opus 5.5 | Opus 5 | Source status |
|---|---|---|---|
| Terminal-Bench 4.0 | 66.4% | 52.3% | Anthropic-reported |
| FrontierCode v1.1 | 54.4% | 48.0% | Anthropic-reported |
| CursorBench 4.0 | 57.8% | 46.6% | Anthropic-reported |
| GDPval-AA v2.1 | 1846 Elo | 1708 Elo | Anthropic-reported |
| AutomationBench | 40.0% | 26.9% | Run and reported by Zapier in Anthropic’s launch materials |
Those scores are useful for deciding what to test first. They are not a substitute for measuring your own task completion rate, latency, token use, tool errors, and review burden.
Stronger safeguards change how some requests are handled
Opus 5.5 is the first Opus model to launch with safeguards similar to Anthropic’s Fable 5.1 system for cybersecurity, biology, and model-development risks. Anthropic says most cybersecurity tasks that trigger its classifiers are transparently routed to Claude Opus 4.8. Requests flagged for biology or frontier-LLM-development risk can be routed to Claude Opus 5.
This matters for evaluation. A team testing Opus 5.5 on security tasks may not always be measuring Opus 5.5 end to end if safety routing intervenes. Anthropic says it plans to expand its Cyber Verification Program so vetted security practitioners can use Opus 5.5 more directly, while its Life Sciences Verification Program is already open to vetted organizations.
For production agents, model safeguards should complement—not replace—system controls such as least privilege, approval gates, credential isolation, sandboxing, and audit logs. See our AI agent security guide for a broader control framework.
Preserved thinking can affect API integrations
Anthropic says Opus 5.5 launches with preserved thinking, the anti-distillation control introduced with Fable 5.1. For API accounts created on or after August 31, 2026, this prevents clients from editing Claude’s prior reasoning context in ways that could be used to extract internal reasoning.
That is an integration detail worth testing before a production migration. Teams that manipulate prior assistant reasoning or rely on unusual conversation-replay behavior should validate their request flow against Anthropic’s preserved-thinking documentation rather than assuming an Opus 5-to-5.5 model-name swap is the only change.
Should Opus 5 users migrate?
For many API users, Opus 5.5 is a strong candidate for immediate evaluation because the direction of the economics is favorable: base token rates are lower, cache reads are substantially cheaper, and Anthropic reports higher speed and lower task-level token use.
That still does not justify a blind production switch. A controlled migration should compare:
- Task completion: does the model finish the same job correctly with less intervention?
- Total tokens: measure input, output, thinking, and cache behavior rather than only list price.
- Latency: standard mode may already be faster; Fast mode trades more money for even lower latency.
- Tool behavior: verify tool selection, retries, and error recovery in your own agent loop.
- Safety routing: specialized cyber or biology workloads may be transparently handled by another Claude model.
- Regression risk: prompt style, output format, and long-running behavior can change across model generations.
Where Opus 5.5 fits in the Claude lineup
Anthropic’s positioning is increasingly about cost-performance tiers rather than a simple “bigger model is always better” ladder. Opus 5.5 is priced well below Fable 5.1 while Anthropic says it reaches Fable-level performance on much everyday high-end work. Sonnet remains the lower-cost general workhorse, while Fable and Mythos occupy the highest capability and controlled-access tiers.
If you are choosing a subscription rather than an API model, our Claude pricing guide covers Free, Pro, Max, Team, Enterprise, and premium-model usage rules. For a broader platform decision, see ChatGPT vs Claude. For the current OpenAI model-level alternative, see GPT-6 Sol and Luna. Developers choosing an environment rather than just a model can also compare Cursor vs Claude Code.
AI-XBlog assessment
The most important part of Claude Opus 5.5 is the cost structure. A 20% cut to base input/output pricing, a 60% cut to cache reads, and Anthropic’s claim of lower token use make this a more meaningful release for production economics than a benchmark-only upgrade would be.
The safety design is also operationally important. Opus 5.5 is more capable, but Anthropic is pairing that capability with stronger classifiers and transparent fallbacks for sensitive domains. Teams should understand that routing behavior before using benchmark numbers to estimate exactly which model will execute a particular workload.
AI-XBlog has not run independent Opus 5.5 benchmarks, so we would not call it the “best model” based on launch-day vendor scores. For existing Opus 5 users, however, the lower official rates alone create a clear reason to test 5.5 now.
FAQ
How much does Claude Opus 5.5 cost?
The standard Claude Platform price is $4 per million input tokens and $20 per million output tokens. Cache reads are $0.20 per million tokens and 5-minute cache writes are $5 per million.
What is the Claude Opus 5.5 API model ID?
Anthropic lists the Claude Platform model ID as claude-opus-5-5.
Is Claude Opus 5.5 cheaper than Opus 5?
Yes. The base token prices are 20% lower, and cache reads are 60% cheaper. Anthropic says typical workloads can cost about 40% less because the new model also uses fewer tokens on many tasks.
Does Claude Opus 5.5 have Fast mode?
Yes. Anthropic says Fast mode is available in Claude Code and the Claude Platform at $8 per million input tokens and $40 per million output tokens, with speeds up to 2.5× standard mode.
Is Claude Opus 5.5 available on AWS, Google Cloud, and Azure?
Yes. Anthropic says the model is available on all platforms and specifically lists Amazon Web Services, Google Cloud, and Microsoft Azure.
Is Opus 5.5 the same as Fable 5.1?
No. They are separate models. Anthropic says Opus 5.5 performs at the level of Fable 5.1 on most work, but Fable remains a different higher-tier model with its own pricing and safeguards.
Should Claude Opus 5 users switch immediately?
It is worth testing immediately because the official pricing is lower, but production migration should still include regression testing for task quality, latency, token use, tool behavior, safety routing, and integration compatibility.
Primary sources
- Anthropic — Claude Opus 5.5 launch announcement
- Anthropic Claude Platform Docs — pricing
- Reuters — Anthropic unveils Claude Opus 5.5
- The Verge — Claude Opus 5.5 safeguards
Source check: Opus 5.5 API pricing and refusal/fallback billing were rechecked September 27, 2026 (Taipei time); broader subscription limits and cloud-distribution details remain last fully reviewed September 24, 2026. Pricing, model availability, subscription usage limits, cloud distribution, safeguards, and documentation can change; this page is maintained as living content.
AI-XBlog Weekly Brief
Keep up with AI that actually works
Join the AI-XBlog Weekly Brief for major AI updates, practical workflows, useful tools, and editor’s picks. No daily noise.
Reader discussion
Join the discussion
Have you tried this tool or workflow? Share your experience, corrections, or questions. Useful reader feedback may help us improve this article.
All comments are reviewed before publication. Your email address will not be published. Promotional links and low-value spam are removed.
