Anthropic released Claude Opus 5.5 on 22 September 2026. OpenAI shipped GPT-6 Sol and Luna the same day. The two launches landed within hours of each other and both led with price, which tells you what this stage of the market is about.
Here is what Opus 5.5 costs, what it scores, and the four API changes that will break your code.
What It Costs
Standard API pricing is $4 per million input tokens and $20 per million output tokens. That is a 20% reduction from Opus 5, which ran at $5 in and $25 out.
- Cache reads: $0.20 per million - down 60% from Opus 5
- Cache writes (5 minute TTL): $5 per million
- Cache writes (1 hour TTL): $8 per million
- Batch API: $2 in / $10 out per million
There is also a faster serving mode at $8 in and $40 out for workloads that need lower latency.
The 40% Claim, Read Carefully
Anthropic says Opus 5.5 costs 40% less to run than Opus 5 on typical workloads. The list price only fell 20%. The other half comes from three places: the 60% cheaper cache reads, output generation that is roughly 30% faster, and fewer tokens spent reaching the same answer.
That means the 40% figure is a claim about your workload, not a number on a price sheet. If your pipeline is cache-heavy you will likely beat 40%. If every request is a cold single-shot call with no reuse, expect closer to the 20% list cut and nothing more.
Measure it before you rewrite any budget. Take a week of production traffic, replay it, and compare the invoice.
Context, Output and Cutoff
- Context window: up to 1M tokens
- Max output: 128K standard, 300K on the Batch API with a beta header
- Knowledge cutoff: June 2026
Benchmarks
These are the published numbers, with the strongest competing score alongside each one.
- Terminal-Bench 4.0: 66.4% - GPT-6 Astra 57.9%, Opus 5 52.3%
- FrontierCode v1.1: 54.4% - GPT-6 Astra 53.3%, Fable 5.1 50.3%
- GDPval-AA v2.1: 1846 Elo - Fable 5.1 1735, Opus 5 1708
- AutomationBench: 40.0% - GPT-6 Astra 41.4% (Astra leads), Opus 5 26.9%
- Terminal-Bench-Science 0.1: 58.7% - GPT-6 Astra 64.6% (Astra leads), Opus 5 29%
- Humanity's Last Exam, with tools: 67.7%
Where It Loses
Two of the six go to GPT-6 Astra: AutomationBench by 1.4 points and Terminal-Bench-Science by 5.9 points. Anthropic published both rather than omitting them, which is worth noting given how often launch pages do the opposite.
Anthropic also said directly that benchmark margins have become a less reliable guide to real-world differences at this capability level, and that the practical gap between Opus 5.5 and Fable 5.1 is narrower than the scores suggest. Take that at face value. A three-point lead on a coding benchmark is not a reason to migrate a working pipeline.
The AutomationBench jump is the number that does mean something: 26.9% to 40.0% over one release is a real change in how far an agent gets before a human has to step in.
Four Breaking Changes
These will return errors or behave differently against code written for Opus 5. Read them before you swap the model string.
- Thinking cannot be disabled. Omit the parameter or use adaptive mode. Code that explicitly turns thinking off needs changing.
- Forced tool use returns a 400. Setting
tool_choice: "any" is rejected. If you relied on forcing a tool call every turn, that pattern is gone.
- Thinking blocks are bound to the model and conversation. Replaying stored thinking blocks from another model or another conversation fails. Any system that caches and re-sends reasoning traces needs auditing.
- The old computer-use tool is rejected.
computer_20251124 no longer validates. Move to computer_toolset_20260801.
The third one is the quiet killer. Conversation-replay architectures that store and resubmit assistant turns are common, and this failure will not show up until a specific code path runs in production.
Where You Can Use It
Live from launch on the Claude API and Claude Platform, Amazon Bedrock, Google Cloud, and Microsoft Foundry. Anthropic says Claude Sonnet 5.5 and Haiku 5.5 follow in the coming weeks, which is the pair most teams will actually deploy at volume.
Should You Upgrade
- Already on Opus 5 with a cache-heavy pipeline - yes, the economics alone justify it. Budget an afternoon for the four breaking changes.
- Running long agentic sessions - yes. AutomationBench at 40.0% against 26.9% is the largest practical improvement in this release.
- On Fable 5.1 and happy - no rush. Anthropic said the gap is narrower than the benchmarks show.
- Cost is the binding constraint - look at GPT-6 Luna first. At $0.10 in and $0.50 out it is a different price category entirely, and it scores 66.6% on DeepSWE v1.1.
- Need computer use or long-horizon science work - benchmark GPT-6 Astra alongside it. Astra leads on both.
Sources