Claude Opus 5.5: Fable 5.1-Level Scores at $4 and $20, With Benchmarks, Effort Costs, and Migration Guide
Claude Opus 5.5, released September 22, 2026, beats Claude Fable 5.1 on every benchmark Anthropic published and costs $4 and $20 per million tokens, 60% less than Fable 5.1. It scores 66.4% on Terminal-Bench 4.0 against 57.9% for GPT-6 Astra. Thinking is always on, the default effort is medium, and four API changes can break Opus 5 code.
Anthropic released Claude Opus 5.5 on September 22, 2026. It is the first model in the Claude 5.5 family, and on Anthropic's own benchmark table it beats Claude Fable 5.1 on every row while costing 60% less per token. It is available now as claude-opus-5-5 on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and in Claude Code (Anthropic).
The verdict: Opus 5.5 is the model to default to for agentic coding and knowledge work, and you rarely need to turn its effort up. At medium effort, the default, it matches or beats GPT-6 Astra at max effort on FrontierCode and GDPval-AA for roughly a fifth of the cost per task. Fable 5.1 no longer has a performance case for everyday coding at 2.5 times the price. Before you switch an API integration, read the breaking changes below: four of them return errors on code that works on Opus 5.
Key Takeaways
- Price: $4 per million input tokens and $20 per million output, 20% below Opus 5 and 60% below Fable 5.1. Cache reads fall 60% to $0.20 per million.
- Performance: Opus 5.5 scores 66.4% on Terminal-Bench 4.0, against 55.8% for Fable 5.1 and 57.9% for GPT-6 Astra. It leads 7 of the 9 benchmarks Anthropic published.
- Efficiency: Anthropic says it costs 40% less per task than Opus 5 on typical workloads and generates output more than 30% faster.
- Effort: Thinking is always on and the default effort is medium, down from high on Opus 5. Medium is already Anthropic's best-value setting on its coding charts.
- Breaking changes: Disabled thinking, forced tool use, and the older
computer_20251124tool all return 400 errors. Text between tool calls now arrives in thinking blocks. - Claude Code: Opus 5.5 is the default model on paid plans and the API, and needs version 2.1.280. Five-hour limits rise 20%, and subscribers get a banked rate-limit reset.
- Safeguards: Most cybersecurity work reroutes to Opus 4.8, and a biology classifier now runs as well. Handle
stop_reason: "refusal"in your code.
Benchmarks: Opus 5.5 vs Fable 5.1, Opus 5, and GPT-6 Astra
Opus 5.5 leads seven of the nine benchmarks on Anthropic's launch table. GPT-6 Astra leads the other two. Opus 5.5 was run at max effort unless noted, with production safeguards on. When the safeguards stepped in, the task was completed by an older model, which Anthropic says likely pulls Opus 5.5's scores down (Anthropic).
| Benchmark | Opus 5.5 | Fable 5.1 | Opus 5 | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|---|---|---|
| Terminal-Bench 4.0 (agentic coding) | 66.4% | 55.8% | 52.3% | 57.9% | 37.3% |
| FrontierCode v1.1 Main (agentic coding) | 54.4% | 50.3% | 48.0% | 53.3% | 47.5% |
| CursorBench 4.0 (agentic coding) | 57.8% | 51.8% | 46.6% | not reported | 41.7% |
| GDPval-AA v2.1 (knowledge work, Elo) | 1846 | 1735 | 1708 | 1542 | 1588 |
| AutomationBench (business workflows) | 40.0% | 31.4% | 26.9% | 41.4% | 28.8% |
| Humanity's Last Exam, with tools | 67.7% | 65.6% | 63.6% | 57.2% | not reported |
| Terminal-Bench-Science 0.1 | 58.7% | 52.6% | 29.0% | 64.6% | 22.4% |
| OSWorld 2.0, partial (computer use) | 81.8% | 80.7% | 74.0% | not reported | not reported |
| Chartography, with tools | 89.0% | 88.4% | 83.4% | not reported | not reported |
Anthropic reports Terminal-Bench 4.0 at xhigh effort for Opus 5.5 and at high effort for Astra, each model's best score. The AutomationBench runs, done by Zapier, counted every safeguard intervention as a failure. Terminal-Bench-Science has a standard error of 3.5 to 5 points per model, so Astra's lead there is less certain than it looks.
Anthropic adds its own caveat: "at these levels of capability we've found that benchmark margins have become a less reliable guide to real-world differences." It says the gap between Opus 5.5 and Fable 5.1 is narrower in its own use than these scores suggest.
Cost Per Task at Every Effort Level
Medium effort gets you most of Opus 5.5's quality, and on some benchmarks all of it. Anthropic published score-versus-cost charts for five benchmarks, one point per effort level. We pulled every point from the launch page's interactive charts (Anthropic):
| Benchmark | Low | Medium (default) | High | Xhigh | Max |
|---|---|---|---|---|---|
| Terminal-Bench 4.0 | 38.5% / $1.29 | 57.6% / $2.94 | 64.2% / $3.88 | 66.4% / $7.35 | 64.8% / $11.24 |
| FrontierCode v1.1 | 47.3% / $0.40 | 54.6% / $0.80 | 54.0% / $1.09 | 51.4% / $2.25 | 54.4% / $6.19 |
| CursorBench 4.0 | 43.7% / $1.18 | 52.5% / $2.90 | 56.0% / $3.97 | 56.0% / $6.99 | 57.8% / $13.43 |
| GDPval-AA v2.1 (Elo) | 1224 / $0.21 | 1576 / $0.86 | 1692 / $1.54 | 1820 / $4.21 | 1846 / $8.92 |
| AutomationBench | 23.3% / $0.50 | 28.6% / $0.64 | 32.0% / $0.70 | 34.4% / $0.86 | 40.0% / $1.37 |
Here is how we read that table:
- Code changes and reviews: use medium. FrontierCode, which grades whether a change is ready to merge, peaks at medium. Max costs nearly eight times as much for a slightly lower score.
- Terminal-based coding agents: use high. High reaches 64.2% on Terminal-Bench 4.0 for $3.88. Xhigh adds 2.2 points but almost doubles the cost.
- Knowledge-work deliverables and business workflows: use xhigh or max. GDPval-AA and AutomationBench keep improving to the top of the range, and AutomationBench stays cheap even at max ($1.37 a task).
- Low effort is for simple, high-volume work only. It falls 19 points behind medium on Terminal-Bench 4.0.
For how to set effort per task in your own prompts and Claude Code sessions, see our Opus 5.5 prompting guide.
The comparison with other models is more striking. Medium Opus 5.5 scores 57.6% on Terminal-Bench 4.0 for $2.94. That beats Fable 5.1 at max effort (55.8% for $19.50) and roughly matches GPT-6 Astra at high effort (57.9% for $7.21). On GDPval-AA, medium Opus 5.5 (1576 at $0.86) beats Astra at max (1542 at $4.53). All of these costs are Anthropic's estimates. Partner results point the same way: Factory calls Opus 5.5 "the first model we'd default to at medium effort," and says it matched Opus 5 on high effort with 20 to 25% fewer output tokens.
How Opus 5.5 Compares
Against Fable 5.1: Opus 5.5 wins every benchmark on Anthropic's table, at $4 and $20 per million tokens against $10 and $50. In one internal test, Anthropic had both models translate HAProxy from C to Rust. Both rewrites passed nearly all of HAProxy's regression tests, but Opus 5.5 finished in 9.5 hours against 12 and cost 51% less. Choose Fable 5.1 only where your own evals show it ahead.
Against Opus 5: Opus 5.5 improves every score and costs less per task. An early tester audited and fixed a 200,000-line codebase with it in under three hours; Opus 5 took more than 20 hours and 2.5 times as many tokens. There is no reason to stay on Opus 5 other than the breaking changes below.
Against GPT-6 Astra: Opus 5.5 wins four of the six benchmarks both models are scored on, at 40% of Astra's per-token price ($10 and $50, per our Astra launch coverage). The biggest leads are GDPval-AA (1846 vs 1542), Humanity's Last Exam (67.7% vs 57.2%), and Terminal-Bench 4.0 (66.4% vs 57.9%). Astra leads AutomationBench (41.4% vs 40.0%) and Terminal-Bench-Science (64.6% vs 58.7%). OpenAI still calls Astra "the world's best model for computer use." Nobody has published Opus 5.5 on OpenAI's computer-use benchmark, OSWorld offline.
Against GPT-6 Sol and Luna: OpenAI released GPT-6 Sol and Luna the same day at $2 and $10, and $0.10 and $0.50, per million tokens. Opus 5.5 scores higher on the two benchmarks both labs report: FrontierCode (54.4% vs Sol's 49.3%) and AutomationBench (40.0% vs 33.2%). Sol is cheaper per task on routine automation. Our September 22 newsletter lines up all three models at every effort level. A week later OpenAI shipped GPT-6.1 Sol, its upgrade to GPT-6 Sol, at the same $2 and $10 per million tokens with cached input cut to $0.10; that page compares it with Opus 5.5 on OpenAI's charts.
Pricing and Specs
| Item | Claude Opus 5.5 | Claude Opus 5 | Claude Fable 5.1 |
|---|---|---|---|
| Input (per million tokens) | $4 | $5 | $10 |
| Output (per million tokens) | $20 | $25 | $50 |
| Cache read | $0.20 | $0.50 | $0.25 |
| 5-minute cache write | $5 | $6.25 | not listed here |
| 1-hour cache write | $8 | not listed here | not listed here |
| Context window / max output | 1M / 128K | not listed here | 1M / 128K |
| Default effort | medium | high | high |
| Reliable knowledge cutoff | June 2026 | not listed here | June 2026 |
Sources: Anthropic's launch page and the Opus 5.5 model page. Batch requests cost half: $2 and $10. The Batch API also allows up to 300K output tokens with the output-300k-2026-03-24 beta header. Fast mode, a research preview, costs $8 and $40 for up to 2.5 times the speed. It runs on the Claude API and in Claude Code, but not on Bedrock, Claude Platform on AWS, Google Cloud, or Foundry (What's new in Opus 5.5). Anthropic commits to keeping the model available until at least September 22, 2027.
Opus 5.5 Is Also Available in Cursor
Cursor announced that Claude Opus 5.5 is now available in Cursor. Cursor's post says it is "the new top model on CursorBench at 57.8% (Max)" and "costs 40% less per task than Opus 5." The leaderboard's own per-task cost column does not show that saving at Max: Opus 5.5 Max is $13.43 per task against $11.95 for Opus 5 Max. The saving appears one effort level down, where Opus 5.5 Extra High costs $6.98 per task against $11.43 for Opus 5 Extra High (39% less) while scoring 56.0% against 46.1%, and at High, where Opus 5.5 costs $3.97 against $9.00. (Cursor on X; CursorBench)
Cursor's official model documentation says Opus 5.5 replaces Opus 5 at 20% lower input and output rates: $4 and $20 per million tokens for Opus 5.5 versus $5 and $25 for Opus 5. (Cursor model docs). Its benchmark page reports that Opus 5.5 scores 57.8% at Max, 56.0% at Extra High and High, 52.5% at Medium, and 43.7% at Low on CursorBench 4.0. The page warns that results are subject to variance and small score differences may not be statistically meaningful. (CursorBench)
The result is a new distribution path, not a new Anthropic model variant. Teams already using Cursor should re-run their own repository evals: CursorBench measures ambiguous, multi-file tasks from Cursor sessions, and its per-task cost depends on the model's effort level and token usage.
What Changes in Claude Code
Opus 5.5 is now the default model on Pro, Max, Team, Enterprise, and Anthropic API setups, and it requires Claude Code 2.1.280 or later. Run claude update if you do not see it (Claude Code model configuration).
- Limits: Five-hour session limits increase 20%. Because Opus 5.5 is cheaper, Anthropic says the same limit goes 25% further (ClaudeDevs). The launch page extends the higher limits to seat-based Enterprise plans.
- Banked reset: Pro, Max, and Team subscribers get a rate-limit reset they can use whenever they choose. You use it from Settings, Usage on the web or in the desktop app by clicking "Reset for free." It is not available in the Claude Code terminal, IDE extensions, or mobile, and any expiry date appears in Settings (Claude Help Center).
- Migration: Run
/claude-api migrateto move an API-backed project toclaude-opus-5-5(ClaudeDevs).
Breaking Changes for API Users
Four changes return errors on code that runs today on Opus 5, and a fifth quietly changes the response shape. The first three also apply on Fable 5.1 (What's new in Opus 5.5).
| Change | What fails | What to do |
|---|---|---|
| Thinking cannot be disabled | thinking: {"type": "disabled"} or a manual budget_tokens returns a 400 | Omit thinking or send "adaptive", and control depth with effort |
| No forced tool use | tool_choice of any or a named tool returns a 400 | Use auto with strict tool use, or structured outputs, and say in the prompt when to call the tool |
| Thinking blocks are tied to the model and conversation | Replaying a block after editing the system prompt, tools, or earlier messages returns a 400 on accounts created on or after August 31, 2026 | Keep conversations append-only; use mid-conversation system messages instead of edits |
computer_20251124 rejected | Declaring the old computer-use tool returns a 400 on the Claude API and Google Cloud | Switch to computer_toolset_20260801; Bedrock still accepts the old tool |
| Text between tool calls moves | No error, but progress notes arrive as empty thinking blocks at the default display setting | Set thinking.display if your UI streams those notes |
Two behavior changes affect cost. A request without effort now runs at medium instead of high. And at the same effort level, Opus 5.5 tends to think more per turn than Opus 5, especially at xhigh and max. Anthropic's advice is to re-run your effort sweep rather than carry an Opus 5 setting over, and to leave room in max_tokens for thinking. Opus 5.5 also reads dense charts and screenshots much more precisely without tools, so older vision workarounds may no longer be needed.
Safeguards and Fallback
Anthropic says Opus 5.5 is comparable to Claude Mythos 5.1 in biology and cybersecurity, so it ships with Fable 5.1-class safeguards. You can find and fix bugs in your own code as part of normal development, but most other cybersecurity tasks are rerouted to Opus 4.8. A biology classifier runs alongside the cyber one. Requests that try to get the model to reproduce its internal reasoning can be declined under a reasoning_extraction category (Anthropic; What's new in Opus 5.5).
A declined request returns HTTP 200 with stop_reason: "refusal" and a stop_details object naming the policy area. Your code should handle it by setting fallbacks: "default" (beta), using the SDK middleware, or retrying yourself. Vetted organizations can apply to the Life Sciences Verification Program now. Anthropic says the Cyber Verification Program will expand to Opus 5.5 in the coming weeks. Opus 5.5 is available with zero data retention.
On alignment, Anthropic says Opus 5.5 scored better than any recent Claude model on nearly every misalignment measure in its roughly 2,000-scenario behavioral audit. It also tried to cross containment boundaries about 85% less often than Opus 5 or Mythos 5.1. Anthropic also notes that Opus 5.5 "often suspects it is being evaluated," which limits how far the audit can be trusted.
What Early Testers Report
The partner reports repeat one pattern: the same or better results in fewer turns and tokens.
- Optiver: matched Opus 5's quality "in about half the turns, time and output tokens," cutting that workload's cost by 40 to 50%.
- Deloitte: at its lowest effort, Opus 5.5 caught 72% of known bugs in code review, against 56% for Opus 5 at high effort.
- Box: used a third of Opus 5's tokens, with answers 40% less verbose.
- Quantium: a coding task that took 38 prompts over four days finished in 11 prompts over three hours.
- Stripe: one Opus 5.5 session directed a dozen others through a multi-day rebase of 40 stacked pull requests, and all 40 passed CI.
- Hebbia: covered 86.6% of its finance-workflow rubric, against 60.3% for Opus 5.
All of these are vendor-selected quotes from the launch page. Anthropic's own test backs up the accuracy point. When Anthropic's models wrote an earnings report from a copy of the web where the source was hard to find, 16 of 18 Opus 5.5 reports passed a grader that fails any invented figure or quote. Neither Fable 5.1 nor Opus 5 passed in any attempt.
When to Choose Opus 5.5
Choose Opus 5.5 at medium effort as your default for agentic coding, code review, and research. It is the strongest model on Anthropic's table and among the cheapest per task at the frontier.
Raise it to high for terminal-driven agents, and to xhigh or max for finished knowledge-work deliverables, where scores keep rising.
Choose GPT-6.1 Sol or GPT-6 Luna instead for high-volume, routine automation, where their lower per-token price wins. GPT-6.1 Sol, OpenAI's September 29 upgrade to GPT-6 Sol, keeps Sol's price with cheaper cached input.
Keep Astra only if your own computer-use evals favor it.
Keep Fable 5.1 only where your own evals show it ahead.
Before you migrate an API client, fix the four breaking changes, add refusal handling, and set effort explicitly. Then compare cost per completed task, not per token. The rest of the Claude 5.5 family has since shipped: Claude Sonnet 5.5 on September 28 and Claude Haiku 5.5 on October 7.
Corrections and Updates
- 2026-10-09: This page said Claude Sonnet 5.5 and Haiku 5.5 would follow in the coming weeks; Anthropic released them on September 28 and October 7. The model-choice advice now points to GPT-6.1 Sol, released September 29, rather than GPT-6 Sol.
- 2026-09-24: The Cursor section said Opus 5.5 "costs 40% less per task than Opus 5" next to its 57.8% Max-mode score, which implied a Max-mode cost comparison. CursorBench's own cost column shows Opus 5.5 Max at $13.43 per task against $11.95 for Opus 5 Max; the 39% to 56% saving is at Extra High, High, and Medium effort. The paragraph now shows the rows.
Sources
- Anthropic, "Introducing Claude Opus 5.5" (benchmarks, effort charts, pricing, tester quotes, safeguards; September 22, 2026): https://www.anthropic.com/claude-opus-5-5
- Anthropic, "Introducing Claude Sonnet 5.5" (September 28, 2026): https://www.anthropic.com/claude-sonnet-5-5
- Anthropic, "Introducing Claude Haiku 5.5" (October 7, 2026): https://www.anthropic.com/claude-haiku-5-5
- OpenAI, GPT-6.1 Sol model page (pricing; read October 9, 2026): https://developers.openai.com/api/docs/models/gpt-6.1-sol
- Anthropic, "Claude Opus 5.5" model page (context window, output limits, cutoffs, cache pricing, availability; checked September 22, 2026): https://platform.claude.com/docs/en/models/opus-5-5/overview
- Anthropic, "What's new in Claude Opus 5.5" (breaking changes, behavior differences, fast mode, refusals; checked September 22, 2026): https://platform.claude.com/docs/en/models/opus-5-5/whats-new-opus-5-5
- Claude Code docs, "Model configuration" (default model, version requirement; checked September 22, 2026): https://code.claude.com/docs/en/model-config
- Claude Help Center, "What is a limit reset?" (how and where to use a reset; checked September 22, 2026): https://support.claude.com/en/articles/17007452-what-is-a-limit-reset
- ClaudeDevs on X (Claude Code limits and reset; September 22, 2026): https://x.com/ClaudeDevs/status/2102438800836489554
- ClaudeDevs on X (platform pricing and migrate command; September 22, 2026): https://x.com/ClaudeDevs/status/2102438805253091551
- ClaudeDevs on X (effort default, compaction, fast mode; September 22, 2026): https://x.com/ClaudeDevs/status/2102438806465245246
- Cursor on X (Opus 5.5 availability, CursorBench score, and per-task cost; September 22, 2026): https://x.com/cursor_ai/status/2102448392773435706
- Cursor model docs, "Claude Opus 5" (Opus 5.5 recommendation and pricing; checked September 23, 2026): https://cursor.com/docs/models/claude-opus-5
- Cursor, "CursorBench" (Opus 5.5 scores, per-task costs, and variance note; checked September 23, 2026): https://cursor.com/cursorbench
- OpenAI, "Introducing GPT-6 Sol and Luna" (Sol and Luna pricing and benchmarks; September 22, 2026): https://openai.com/index/introducing-gpt-6-sol-and-luna/
Read next
Related coverage- OpenAI
GPT-6 Sol and Luna: Half the Price of GPT-5.6, With Benchmarks and Cost at Every Effort Level
OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, 2026, at $2 and $10 and $0.10 and $0.50 per million tokens, half the price of GPT-5.6. Luna scores 66.6% on DeepSWE for about $0.22 a task. Since October 7 the pair has been rolling out as ChatGPT's Chat models, and GPT-6.1 Sol has succeeded Sol at the same price.
- Claude
Claude Fable 5.1 and Mythos 5.1: Same Price, Cheaper Cache Reads, Fewer Safeguard Interventions
Anthropic released Claude Fable 5.1 on September 1, 2026 as claude-fable-5-1 at $10 per million input and $50 per million output tokens, with cache reads cut 75% to $0.25 per million. Anthropic estimates typical workloads cost about 25% less than Fable 5. Claude Mythos 5.1 is the same model with fewer cyber and life-science safeguards, limited to vetted US organizations.
- OpenAI
OpenAI Launches GPT-6 Astra for Computer Use and Complex Work
OpenAI launched GPT-6 Astra on September 3, 2026, as a model for computer use, browsing, software engineering, cybersecurity, science, and complex professional workflows. Access started with a limited set of organizations; Astra is now available in ChatGPT Work and Codex, the OpenAI API, Microsoft Foundry, and Amazon Bedrock.
- Claude Code
Claude Code Projects Turn One Conversation Into Parallel Cloud Sessions
Anthropic's redesigned Projects beta gives Claude Code one coordinating conversation that delegates work to parallel threads. Threads are usually cloud sessions, which Anthropic's docs now list as available, no longer in research preview, on Pro, Max, and Team and for eligible Enterprise seats. A project can also run a thread on your own machine through Remote Control.
- Claude
Prompting Claude Opus 5.5: Delete 'Think Carefully', Name the Finish Line, Start at Medium
Prompt Claude Opus 5.5 by handing over the whole task with a clear finish line, deleting “think carefully” lines, and starting at medium effort. Thinking is always on, so effort is your depth control. Add a CLAUDE.md rule that names when to stop, and set effort explicitly instead of carrying Opus 5 settings over.
- Claude
Claude Sonnet 5.5 Launches With Lower Task Costs and Cursor Support
Anthropic introduced Claude Sonnet 5.5 on September 28 with model ID `claude-sonnet-5-5`. It keeps Sonnet 5's $2 input and $10 output price per million tokens; Anthropic reports output generation more than 30% faster and up to 30% lower cost per task. Cursor lists the model in Settings > Models. Cache reads dropped to $0.10 per million tokens on October 7.
- Anthropic
Anthropic Releases Claude Haiku 5.5, Its Fastest Model and First Haiku With Effort Levels
Anthropic released Claude Haiku 5.5 on October 7, 2026, its fastest model to date and the first Haiku with effort levels. It scores 72.4% on OSWorld 2.1 against Haiku 4.5's 15.7%, and costs $0.10 per million input and $0.50 per million output tokens for prompts up to 100,000 tokens, 90% less than Haiku 4.5.
- OpenAI
GPT-6.1 Sol: Astra-Level Coding at a Fifth of the Price, With Cost at Every Effort Level
OpenAI released GPT-6.1 Sol on September 29, 2026, at $2 input and $10 output per million tokens, a fifth of GPT-6 Astra's price, with cached input at $0.10. On OpenAI's charts it edges past Astra's best DeepSWE score at high effort and trails Astra by 2.1 points on OSWorld offline. The API model ID is gpt-6.1-sol.
Frequently Asked Questions
What is Claude Opus 5.5?
Claude Opus 5.5 is Anthropic's newest Opus model and the first in the Claude 5.5 family, released September 22, 2026. Anthropic says it performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5.
How much does Claude Opus 5.5 cost?
Opus 5.5 costs $4 per million input tokens and $20 per million output tokens. Cache reads are $0.20, 5-minute cache writes $5, and 1-hour cache writes $8 per million. Batch requests are half price, and Fast mode costs $8 and $40.
Is Claude Opus 5.5 better than Claude Fable 5.1?
On Anthropic's published benchmarks, yes. Opus 5.5 scores higher on all nine, including 66.4% vs 55.8% on Terminal-Bench 4.0, at 40% of Fable 5.1's per-token price. Anthropic says the real-world gap is narrower than the scores suggest.
How does Claude Opus 5.5 compare with GPT-6 Astra?
Opus 5.5 leads GPT-6 Astra on four of the six benchmarks both are scored on, including Terminal-Bench 4.0 and GDPval-AA. Astra leads AutomationBench and Terminal-Bench-Science. Opus 5.5 costs $4 and $20 per million tokens against Astra's $10 and $50.
What effort level should I use with Opus 5.5?
Start at medium, the default. On Anthropic's charts, medium reaches 54.6% on FrontierCode for $0.80 a task, the same as max effort. Use high for terminal-based coding agents and xhigh or max for knowledge-work deliverables, where scores keep climbing.
What breaks when I move from Opus 5 to Opus 5.5?
Four things: thinking cannot be disabled, forced tool use returns a 400 error, thinking blocks are tied to the model and conversation, and the older computer_20251124 tool is rejected on the Claude API and Google Cloud. Text between tool calls also moves into thinking blocks.
What changed for Claude Code subscribers?
Opus 5.5 is the default model on Pro, Max, Team, Enterprise, and API setups, and needs Claude Code 2.1.280 or later. Five-hour limits rise 20%, and subscribers get a rate-limit reset they can use from Settings, Usage on the web or desktop app.
Is Claude Opus 5.5 available in Cursor?
Yes. Cursor says Opus 5.5 is available in Cursor, where it reports a 57.8% CursorBench 4.0 score at Max effort. Cursor's docs say it replaces Opus 5 at 20% lower input and output rates.