This article has been updated since the audio was generated. The narration may differ from the current text.
AI SummaryGenerated from this article
Anthropic released Claude Opus 5.5, claiming 40% lower typical running costs than Opus 5 through 20% cheaper token prices, 60% cheaper cache reads, and reduced tokens per task. The model performs at roughly Claude Fable 5.1 levels on most work while showing gains on agentic coding benchmarks, scoring 66.4% on Terminal-Bench 4.0 and 57.8% on CursorBench 4.0. Opus 5.5 generates output over 30% faster and used 2.5 times fewer tokens on a HAProxy translation test. External evaluators from METR and Frontier Design tested the model before release, though actual costs and performance depend on individual workloads.
Anthropic has released Claude Opus 5.5, the first model in its new Claude 5.5 family. The company says it performs at roughly the level of Claude Fable 5.1 on most work while costing 40% less to run than Opus 5 on typical workloads at default settings.
The pricing behind that estimate is more complicated. In its Opus 5.5 announcement, Anthropic lists input and output token prices that are 20% lower than Opus 5’s. The larger estimated saving comes from cheaper cache reads and using fewer tokens to finish a task. An application’s actual bill will depend on its workload.
Anthropic also reports gains in agentic coding, computer use, knowledge work, output speed and model behavior. Its published results make a case for testing Opus 5.5, though they do not establish that it will be cheaper or better for every job.
Opus 5.5 Starts a New Claude Family
Opus 5.5 is the first Claude 5.5 release. Anthropic says Sonnet 5.5 and Haiku 5.5 will follow in the coming weeks; neither is available as part of this announcement.
Two comparisons frame the launch. Anthropic describes Opus 5.5 as performing at the level of Claude Fable 5.1 on most work. Its 40% lower typical running cost is measured against Opus 5. The cost claim does not compare Opus 5.5 with Fable 5.1, and “most work” does not mean the models behave alike on every task.
This is also Anthropic’s first model release since the company called for “pacing the frontier.” Anthropic says external evaluators, including METR and Frontier Design, tested Opus 5.5 before release. The announcement’s summary does not show that those organizations endorsed every safety conclusion Anthropic makes.
Coding Scores Improve, but Test Settings Matter
Anthropic reports its clearest performance gains on agentic coding, where a model uses tools and works through software changes over multiple steps. Its published benchmark table puts Opus 5.5 ahead of both Opus 5 and Fable 5.1 on Terminal-Bench 4.0 and CursorBench 4.0.
Agentic coding test
Opus 5.5
Fable 5.1
Opus 5
Terminal-Bench 4.0
66.4%
55.8%
52.3%
CursorBench 4.0
57.8%
51.8%
46.6%
These are Anthropic’s reported results, not measurements made for this article. Unless otherwise noted, the Opus 5.5 figures in its benchmark table use adaptive thinking at maximum effort; its Terminal-Bench 4.0 result uses xhigh effort. Anthropic’s separate cost-effectiveness claim concerns the default effort setting. The maximum-effort scores and default-effort cost estimate do not describe one test run.
At medium effort, Anthropic says Opus 5.5 scores 54.6% on FrontierCode v1.1, compared with a 53.3% top score it lists for GPT-6 Astra, at roughly one-fifth the cost per task. That company-reported cross-model comparison reflects the test setup and token use as well as posted token prices. Developers will need to validate the cost against their own jobs.
The broader results are mixed. Anthropic’s table lists Opus 5.5 at 40.0% on AutomationBench, below GPT-6 Astra’s listed 41.4%. It also lists GPT-6 Astra ahead on Terminal-Bench-Science 0.1. Anthropic attributes some comparison figures to other organizations and notes that production safeguards could intervene during parts of its evaluations. The coding gains stand, but no single leaderboard position tells the whole story.
The 40% Saving Is Not a Flat Price Cut
Anthropic’s posted token prices show where Opus 5.5 becomes cheaper than Opus 5:
Price per million tokens
Opus 5.5
Opus 5
Input
$4
$5
Output
$20
$25
Cache reads
$0.20
$0.50
Cache writes
$5
$6.25
Input, output and cache-write prices fall by 20%. Cache reads fall by 60%. Caching lets an application reuse previously processed context, which can save considerably more in long coding or agent sessions that repeatedly refer to the same material. Anthropic says cache reads account for much of the cost of agentic and coding work.
Its 40% typical-workload estimate combines those price changes with a further claim: Opus 5.5 uses fewer tokens per task than Opus 5. Every request, API bill and coding project will have its own cost profile. A short request with little cached context will differ from an agent repeatedly working through a repository.
Anthropic says Opus 5.5 generates output more than 30% faster than Opus 5, though that does not mean every tool-using task will finish 30% sooner. A separate fast mode is available in Claude Code and the Claude Platform, with Anthropic advertising up to 2.5 times the speed. Fast mode costs $8 per million input tokens and $40 per million output tokens, twice the standard Opus 5.5 input and output rates.
For subscribers, Anthropic says it is increasing five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans. It is also offering a rate-limit reset that users can save and apply when they choose. Those access changes are separate from the per-token pricing comparison.
Long Coding Jobs Are the Strongest Practical Test
Anthropic’s examples suggest the efficiency gains may be most valuable when a model must sustain work across a large codebase. It says an early tester used Opus 5.5 to audit and fix a 200,000-line codebase in under three hours. In the same account, Opus 5 took more than 20 hours and used 2.5 times as many tokens.
In an internal test, Anthropic asked Opus 5.5 and Fable 5.1 to translate HAProxy from C to Rust. Both rewrites reportedly passed nearly all of HAProxy’s regression tests, while Opus 5.5 finished in 9.5 hours rather than 12 and cost 51% less. Passing nearly all regression tests is a useful result, though it does not demonstrate a production-ready replacement.
These examples involve extended, checkable work, but they remain selected tests reported by Anthropic. A team considering Opus 5.5 for migrations or audits should compare models on its own repositories, record total cost and completion time, and review the changes against its existing tests. A faster run has little value if it misses a behavioral regression.
Safety Claims Rest on More Than One Kind of Test
Anthropic says Opus 5.5 achieved its strongest result to date on the company’s automated behavioral audit, which tests models across thousands of simulated scenarios. It reports that the model is less likely than recent predecessors to take hard-to-reverse actions or exceed assigned boundaries, and that it resists prompt injection better than Opus 5.
The findings concern Anthropic’s evaluation suite. Simulated scenarios can test boundaries, but a high score cannot establish that an agent will respect every boundary in a live system. Anthropic says it has expanded testing to longer tasks, impossible tasks and scenarios modeled on real incidents, while acknowledging the limits of that testing.
METR and Frontier Design’s external evaluations add scrutiny. Their participation is distinct from Anthropic’s own reported audit score. Anthropic also says biological and cybersecurity capabilities warrant safeguards similar to those used for Fable 5.1. Vetted organizations can apply for its Life Sciences Verification Program, and Anthropic says it plans to expand access through its Cyber Verification Program in the coming weeks. Access controls are part of the release; they do not mean the underlying risks have disappeared.
Clearer Writing Could Make Long Sessions Easier to Check
Anthropic says Opus 5.5 responds more naturally to writing instructions and puts important information earlier in its answers. The evidence it describes comes from early testers and its own use, without a comparable public score for writing quality. It is a reported improvement, not a measured guarantee for every style or prompt.
Clear explanations are useful in the long-running tasks behind the coding claims. When an agent describes what it changed, what failed and what needs review, a person can inspect its work more easily. Tests and oversight are still necessary.
Final Thoughts
For developers, the useful test is whether Opus 5.5 finishes real tasks reliably enough to justify the total cost. Lower standard token prices, sharply cheaper cache reads and reported reductions in tokens used per task give them reason to make that comparison on their own work.
Safety calls for a separate judgment. External testing and a stronger internal audit are welcome for a model that can act over longer periods, but neither assures safe behavior in deployment. Its efficiency will be most useful where teams can check the work and keep safeguards and human review in place.
Frequently Asked Questions
5 questions
1
What Is Claude Opus 5.5?
Claude Opus 5.5 is the first model in Anthropic’s Claude 5.5 family. Anthropic says it performs at roughly the level of Claude Fable 5.1 on most work and improves on Opus 5 in areas including agentic coding. Sonnet 5.5 and Haiku 5.5 are planned for later release.
2
How Much Does Claude Opus 5.5 Cost?
Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens at standard rates. Cache reads cost $0.20 per million tokens, and cache writes cost $5. Anthropic estimates that typical workloads at default settings cost 40% less than with Opus 5, but individual bills will vary with caching and token use.
3
Is Claude Opus 5.5 Faster Than Opus 5?
Anthropic says Opus 5.5 generates output more than 30% faster than Opus 5. That figure describes output generation, not the completion time of every multi-step task. Anthropic also offers a separately priced fast mode in Claude Code and the Claude Platform, advertised at up to 2.5 times the speed.
4
Does Opus 5.5 Outperform Claude Fable 5.1?
Opus 5.5 leads Fable 5.1 on several coding benchmarks published by Anthropic, including Terminal-Bench 4.0 and CursorBench 4.0. Anthropic nevertheless describes the models as performing at a similar level on most work. Benchmark settings vary, and the company’s 40% typical cost-saving claim compares Opus 5.5 with Opus 5, not Fable 5.1.
5
Was Claude Opus 5.5 Externally Evaluated for Safety?
Yes. Anthropic says METR and Frontier Design were among the external evaluators that tested Opus 5.5 before release. Its claim that the model achieved its strongest alignment score to date refers to Anthropic’s own automated behavioral audit. External participation does not independently confirm every result or guarantee safe behavior in deployment.
