Overview
Anthropic released Claude Opus 5.5 on September 22, 2026, and the pitch is unusually specific for a model launch: performance close to the company's own Claude Fable 5.1, at a fraction of the running cost. Anthropic positions the new model for long-running agentic coding, computer use, research, and professional knowledge work, and it ships across the Claude API and major cloud platforms including Amazon Bedrock, Google Cloud, and Microsoft Foundry. Reuters covered the launch as part of a broader wave of frontier model releases arriving within days of each other.
The headline numbers are hard to ignore. Input and output tokens cost 20% less than the previous Opus 5. Cache reads, which dominate the cost of long agentic sessions, are 60% cheaper. Anthropic says typical workloads end up costing 40% less overall, not just because of lower per-token prices but because the model finishes tasks in fewer steps. Independent coverage from TechCrunch frames this as Anthropic undercutting its own flagship on price while narrowing the capability gap to almost nothing.
But "best" is a loaded word, and the research behind this launch supports a more precise claim than a blanket superlative. Opus 5.5 leads seven of nine published benchmark comparisons against Fable 5.1, Opus 5, and two OpenAI models. It loses on two. Customer testimonials describe faster, cheaper, less-supervised task completion. None of that amounts to a scientifically representative survey of "a number of users," and Anthropic itself cautions that the margins are getting harder to read as frontier models converge. What follows separates the documented gains from the marketing gloss.
Pricing and Efficiency: The 40% Cost Story
The most concrete change is on the invoice, not the leaderboard. Anthropic cut prices across every tier of the API and paired that with a model that appears to need fewer tokens and fewer turns to finish comparable work.
| API price per 1 million tokens | Opus 5.5 | Opus 5 | Change |
|---|---|---|---|
| Input | $4 | $5 | 20% lower |
| Output | $20 | $25 | 20% lower |
| Cache write | $5 | $6.25 | 20% lower |
| Cache read | $0.20 | $0.50 | 60% lower |
That cache-read cut matters more than it looks. Anthropic says cache reads make up a large share of the cost of long-running agentic coding sessions, where a model repeatedly re-reads context across dozens or hundreds of steps. Cutting that line item by 60%, on top of a model that Anthropic says completes agentic work in fewer steps, is why the company arrives at its 40% total-cost claim rather than the 20% suggested by input and output pricing alone.
Speed moved too. Opus 5.5 generates output more than 30% faster than Opus 5, and a paid "Fast mode" in Claude Code and the Claude Platform pushes that to roughly 2.5 times the speed, at a premium of $8 per million input tokens and $40 per million output tokens. The model now runs a 1-million-token context window with a 128,000-token maximum output, and it always runs with adaptive thinking switched on. Anthropic removed the option to turn thinking off entirely, a notable choice given that some developers previously used non-thinking mode for latency-sensitive, low-complexity calls.
Benchmark Results: Where Opus 5.5 Leads, and Where It Doesn't
Anthropic's published comparison pits Opus 5.5 against its own Fable 5.1 and Opus 5, plus OpenAI's GPT-6 Astra and GPT-5.6 Sol.
| 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 | 54.4% | 50.3% | 48.0% | 53.3% | 47.5% |
| CursorBench 4.0 | 57.8% | 51.8% | 46.6% | n/a | 41.7% |
| GDPval-AA v2.1 (knowledge work, Elo) | 1,846 | 1,735 | 1,708 | 1,542 | 1,588 |
| 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% | n/a |
| Terminal-Bench-Science 0.1 | 58.7% | 52.6% | 29.0% | 64.6% | 22.4% |
| OSWorld 2.0 (partial score) | 81.8% | 80.7% | 74.0% | n/a | n/a |
| Chartography (with tools) | 89.0% | 88.4% | 83.4% | n/a | n/a |
Opus 5.5 tops seven of these nine rows, and the coding gains over its own predecessor are the largest story in the table: Terminal-Bench 4.0 jumped from 52.3% to 66.4%, a 14.1-point improvement. The GDPval-AA v2.1 score of 1,846 Elo is also notable because that benchmark spans 44 occupations covering documents, spreadsheets, and other professional knowledge-work tasks, not just software.
The two losses are worth naming rather than burying. GPT-6 Astra beat Opus 5.5 on AutomationBench (41.4% versus 40.0%) and on Terminal-Bench-Science 0.1, where Astra's 64.6% clears Opus 5.5's 58.7% by a wide margin. Anthropic's own numbers show that gap is real, not statistical noise, even accounting for Terminal-Bench-Science's roughly ±3.5 to 5 point standard error per model. Terminal-Bench 4.0's smaller reported error, about ±2.6 points for Opus 5.5, means the coding lead there is more defensible than some of the narrower gaps elsewhere in the table.
A methodological caveat deserves mention: Anthropic ran Opus 5.5 at its maximum "xhigh" reasoning effort on Terminal-Bench while GPT-6 Astra ran at "high" effort, and some competitor scores were supplied by OpenAI rather than reproduced independently by Anthropic under identical conditions. Independent listings, including Artificial Analysis benchmarks reproduced on OpenRouter, put Opus 5.5's Intelligence Index around 57.6 to 58, though the exact figure shifts with evaluation configuration and reasoning effort. None of this erases the lead; it just means the size of the lead is somewhat configuration-dependent.
Real-World Coding and Agentic Task Reports
Benchmarks are one thing. Anthropic also published a set of task-level comparisons meant to show how the efficiency gains play out in practice:
- Large-scale migration: an early tester reportedly migrated a 680,000-line codebase in under a day, a task Anthropic estimates would otherwise take multiple weeks of engineering time (an internal estimate, not an independently audited control).
- Codebase audit and fix: Opus 5.5 audited and repaired a 200,000-line codebase in under three hours; Opus 5 reportedly needed more than 20 hours and used 2.5 times as many tokens on the same task.
- Web performance tuning: across 40 attempts at improving page-load times, Opus 5.5 succeeded 39 times, while Opus 5 delivered smaller gains and sometimes altered application behavior unintentionally.
- C-to-Rust translation: translating HAProxy from C to Rust, both Opus 5.5 and Fable 5.1 passed nearly all regression tests, but Opus 5.5 finished in 9.5 hours versus 12 for Fable 5.1, at 51% lower cost.
- Merger-analysis workflow: Opus 5.5 and Opus 5 reached identical financial conclusions, but Opus 5.5 produced a more thorough model and cleaner presentation, in 63 minutes instead of 93, at roughly half the cost.
Customer-reported figures reinforce the same pattern of fewer turns and fewer tokens rather than dramatically different answers. Quantium said a task that previously took 38 prompts over four days dropped to 11 prompts over three hours. Optiver reported matching Opus 5's output quality in about half the turns, time, and tokens, with workload savings of 40 to 50%. Kiro said it solved more public command-line tasks using roughly 40% fewer calls and half the tokens. GitHub reported Opus 5.5 among the leanest performers, by token and step count, in its Copilot CLI and VS Code testing. Clio described an unattended six-repository engineering task running for more than 18 hours with less rework than Opus 5, and Lovable reported completions in one-third to one-half fewer steps.
It's worth being direct about what this evidence is and isn't. These are Anthropic-selected customer testimonials, not a controlled study across a representative sample of developers. The direction is consistent, though: less supervision, fewer retries, faster wall-clock completion, rather than dramatically higher raw answer quality on every task.
Knowledge Work, Writing, and Reasoning
The gains extend past code. Anthropic says Opus 5.5's writing puts key information earlier, uses less jargon, and follows user-specified style rules more consistently, though these are the company's own qualitative observations rather than an independent writing study.
A more concrete data point comes from an internal earnings-release research task: 16 of 18 Opus 5.5 reports met Anthropic's quality bar across different effort settings, where any invented figure or quotation caused automatic failure. Fable 5.1 and Opus 5 reportedly failed to meet that same bar on every attempt in the same test, which, if it holds up under broader scrutiny, is a meaningful signal about hallucination rates in high-stakes financial writing. Walleye Capital separately reported that Opus 5.5 largely solved its evaluation suite even at the lowest effort setting, improved further at higher settings, and at one point flagged an error in the evaluation's own instructions.
The throughline across these reports is efficiency more than raw genius. Several accounts describe Opus 5.5 reaching similar output quality to its predecessor with less hand-holding, rather than producing categorically better answers. That's a meaningful upgrade for teams running agents unattended, but it's a different kind of improvement than a pure accuracy jump, and it's worth distinguishing the two when reading Anthropic's marketing.
Safety and Reliability Upgrades
Anthropic frames Opus 5.5 as its strongest performer yet on an internal automated behavioral audit spanning nearly 2,000 scenarios, and says the model is less likely than recent predecessors to take irreversible actions, act outside its assigned boundaries, or comply with prompt-injection attempts. The New York Times covered the safety angle as a notable part of the release, given ongoing debate over how fast frontier labs should be shipping increasingly capable, increasingly autonomous systems.
In a new containment-boundary evaluation, Opus 5.5 reportedly attempted to circumvent restrictions about 85% less often than Opus 5 and Claude Mythos 5.1, with all observed attempts self-reported and rated low severity. Anthropic also expanded its testing to include longer tasks, deliberately impossible tasks, and scenarios modeled on real security incidents. Gray Swan's independently run prompt-injection test placed Opus 5.5 in a tie with Fable 5.1 for the lowest attack-success rate among tested models, and Anthropic says Opus 5.5 matches or beats Opus 5 against injection attacks across tested coding, tool-use, computer-use, and browsing settings.
One wrinkle: in many cybersecurity-adjacent tasks, Opus 5.5 may transparently fall back to the older Opus 4.8 model rather than attempting the task directly, a production safeguard that complicates direct comparisons between the model's nominal capability and what it actually delivers to end users. Anthropic is candid about the limits here too, acknowledging that its evaluations can't rule out the model recognizing when it's being tested, and that reliably detecting every problematic behavior before deployment remains an unsolved problem.
Why Users Call It "the Best," and Why That Claim Has Limits
Putting the evidence together, several threads explain why early adopters are gravitating toward Opus 5.5:
- Coding leadership. It leads the published Terminal-Bench, FrontierCode, and CursorBench comparisons, with the largest gains showing up against its own predecessor rather than against outside competitors.
- Cost-performance ratio. Near-frontier or frontier-leading scores paired with 20% lower token prices and a claimed 40% lower typical task cost is a rare combination; usually labs charge more, not less, for a capability jump.
- Lower supervision burden. Repeated testimonials describe fewer prompts, fewer tool calls, and shorter unattended runs, which matters more to production teams than a benchmark percentage point.
- Generalist strength. Rather than dominating one niche, Opus 5.5 places at or near the top across coding, professional knowledge work, reasoning, computer use, and chart interpretation.
- Safer agentic behavior. Lower boundary-crossing rates and stronger prompt-injection resistance make it more plausible to deploy in agents that can execute commands, browse, or touch business systems unsupervised.
All of that is real and documented. But the claim that it's simply "the best" needs qualifying. GPT-6 Astra beats it on AutomationBench and clearly beats it on Terminal-Bench-Science, so any team whose workload skews toward scientific computing or certain business-workflow automation may find Astra the stronger pick. Anthropic's own benchmark configurations weren't uniform across competitors, and some competitor numbers came from OpenAI rather than Anthropic's own testing. Anthropic has also said internally that the practical gap between Opus 5.5 and its own Fable 5.1 feels narrower than some benchmark numbers suggest, which is an unusually candid admission for a company marketing its newest flagship.
Most importantly, the "number of users who tried it" evidence circulating around this launch is not a broad, independently run survey. It's a curated set of customer case studies published by Anthropic (Quantium, Optiver, Kiro, GitHub, Clio, Lovable, Walleye Capital), plus benchmark runs from third parties like Gray Swan and Artificial Analysis. No large-scale, randomized study comparing identical prompts, agent harnesses, and costs across a representative population of users has surfaced yet. That doesn't invalidate the reports, but it does mean "the best according to users" should currently be read as "the best according to the users Anthropic chose to showcase, corroborated by several independently run benchmarks," which is a meaningfully narrower and more honest claim.
Conclusion
Claude Opus 5.5 earns its headline mostly on efficiency: comparable or better benchmark performance than Fable 5.1 and clearly better performance than Opus 5, delivered with roughly 40% lower typical task cost, 30%+ faster output, and fewer tokens and turns per task across a wide range of customer reports. Its coding gains, particularly on Terminal-Bench 4.0, are the largest single leap in the release, and its safety metrics, including an 85% drop in containment-boundary circumvention attempts, suggest Anthropic is trying to make autonomous, long-running agents genuinely safer to deploy rather than just faster.
What it isn't, based on the research available, is a universal winner. GPT-6 Astra still leads on two of nine published benchmarks, evaluation conditions weren't perfectly matched across labs, and the enthusiastic user testimonials come from a curated set of Anthropic customers rather than a randomized study. For teams building agentic coding tools, long-running automations, or knowledge-work pipelines, the combination of price cuts and step-efficiency gains makes Opus 5.5 a serious default choice. For workloads centered on scientific computing or specific business-workflow automation, it's worth testing GPT-6 Astra alongside it before assuming any one model is simply "the best." The more durable trend here may be less about any single benchmark chart and more about the direction all of this data points: frontier labs converging on similar capability while competing harder on cost, speed, and the amount of babysitting an agent needs to finish the job.
