Long-horizon coding that finishes the job

Anthropic reports a tester completing a 680,000-line code migration in under a day, and an early tester auditing and fixing a 200,000-line codebase in under three hours where Opus 5 took over 20 hours and used 2.5x as many tokens. On Terminal-Bench 4.0, Opus 5.5 scores 66.4% against 52.3% for Opus 5. GitHub reports it solved more terminal tasks than Opus 5 in less than half the steps in VS Code.

Business outcome: large-scale refactors and migrations that previously required weeks of engineering time become single-day, single-agent tasks.

40% lower running cost with fewer tokens per task

Input and output tokens are $4 and $20 per million, 20% below Opus 5, and cache reads drop to $0.20 per million — 60% less than Opus 5. Anthropic states the model also uses fewer tokens per task, netting out to a 40% cost drop on typical workloads. On FrontierCode, Opus 5.5 at default effort beats GPT-6 Astra at roughly 20% of the cost per task.

Business outcome: agentic workloads become economically viable at production scale rather than confined to pilot budgets.

Research and analysis you can verify

On GDPval-AA v2.1, which tests real-world work across 44 occupations, Opus 5.5 scores 1846 Elo against 1735 for Fable 5.1 and 1708 for Opus 5. In an internal test, 16 of 18 Opus 5.5 reports cleared Anthropic's quality bar on a research task where any invented figure or quote would have failed, while neither Fable 5.1 nor Opus 5 cleared it in any attempt.

Business outcome: research outputs carry a lower verification burden, reducing the review overhead that erodes the value of AI-assisted analysis.

Strongest alignment scores Anthropic has recorded

Opus 5.5 achieves the best scores of any model to date on Anthropic's automated behavioural audit, which tests across thousands of simulated scenarios. It is less likely than recent models to take hard-to-reverse actions or act outside given boundaries, and it matches or beats Opus 5 on prompt injection in every setting tested — coding, tool use, computer use, and web browsing. On a benchmark run by AI security firm Gray Swan, it ties Fable 5.1 for the lowest prompt injection success rate of any model tested.

Business outcome: autonomous agents can be granted longer unsupervised run windows with a documented, auditable safety rationale.

Output that reads like a colleague wrote it

Anthropic made major changes to how the model writes, addressing the most common feedback about Opus 5. It puts the most important information up front, is less likely to use jargon, and follows supplied writing rules. One early tester's summary was simply that it 'writes the way I do.' Anthropic frames clearer communication as a safety benefit as well as a practical one, because work becomes easier to check.

Business outcome: less time spent reformatting model output for stakeholders and faster review cycles on long working sessions.

Guardrails built for autonomous operation

Opus 5.5 ships with a classifier that screens every action before it runs, an open-source sandbox that security teams can audit, and code review that catches vulnerabilities before merge. Vetted organisations can apply to the Life Sciences Verification Program for biology research, and Anthropic is expanding access to its Cyber Verification Program for verified cybersecurity practitioners.

Business outcome: security and compliance teams get an auditable control surface rather than having to build their own around a black box.

• Frontier capability at mid-tier costOpus 5.5 leads Anthropic's published benchmark table on agentic coding, computer use, and knowledge work while costing 40% less to run than Opus 5.
• Token efficiency compounds at scaleFewer tokens per task plus a 60% cache read reduction means the savings multiply across high-volume agent workloads rather than applying once.
• Best-in-class alignment resultsThe highest scores Anthropic has recorded on its automated behavioural audit, with prompt injection resistance matching or beating Opus 5 in every tested setting.
• Documented, vetted access pathsLife Sciences and Cyber Verification Programs give regulated teams a legitimate route to sensitive-domain use rather than a blanket restriction.

• Safeguards can reroute your taskAnthropic states that when safeguards intervened in its own evaluations, cybersecurity tasks were completed by Claude Opus 4.8 and biology and frontier LLM development tasks by Claude Opus 5 — likely reducing measured performance on those benchmarks.
• Benchmark margins are narrowingAnthropic itself cautions that at these capability levels, benchmark margins have become a less reliable guide to real-world differences, and that the gap between Opus 5.5 and Fable 5.1 is narrower than the scores suggest.
• Independent reviewers place a rival ahead on some measuresThe independent AI Explained channel found GPT-6 Astra ahead on Terminal-Bench-Science 0.1 by around 6% and on Humanity's Last Exam Diamond by around 5%, so Opus 5.5 is not the outright leader on every axis.
• Professional RealityAnthropic's own system card states Opus 5.5 remains well below the level needed to substitute for its research scientists and engineers — on its internal Cobbench, Opus 5.5 succeeds around 56% of the time where Anthropic says at least 85% would be required for full substitution.

Codebase-wide migration and audit programmes

Engineering teams facing multi-repository refactors get a model that stays on task for extended runs. Clio reported handing Opus 5.5 a large engineering task across six repositories and letting it run unattended for over 18 hours, hitting milestones faster than Opus 5 with minimal reworking. Compare this against the broader field of AI coding tools before standardising.

Quantitative research and financial modelling

Walleye Capital reported that Opus 5.5 largely solved its evaluation suite on the lowest effort setting, and at higher settings detected that the minute indexing in the firm's own instructions was off by one — an error no model had previously caught. For firms where one wrong assumption undermines a result, that self-checking behaviour is the differentiator.

Production agent deployment with governance

Teams shipping autonomous agents into enterprise systems get a pre-execution action classifier, an auditable open-source sandbox, and stronger prompt injection defences. Optiver reported Opus 5.5 matched Opus 5's quality on agentic coding tasks in about half the turns, time, and output tokens, cutting that workload's cost by 40 to 50%.

Legal, consulting, and professional analysis

Deloitte Consulting LLP reported that even at its lowest effort setting, Opus 5.5 caught 72% of known bugs in code reviews versus 56% for Opus 5 at high effort, with fewer false alarms. LexisNexis Legal & Professional noted consistent identification of highly relevant citations and strength with statutes in initial evaluations.

Confirm your access route — Opus 5.5 is available through the Claude API, Claude Code, and the Claude Platform, and Anthropic has raised five-hour limits on Pro, Max, Team, and seat-based Enterprise plans.

Baseline one real workload before switching — run a representative long-horizon task on your current model and record tokens consumed, wall-clock time, and cost per completed task so the 40% saving claim can be verified against your own numbers.

Set effort level deliberately — Anthropic's benchmarks show Opus 5.5 at default (medium) effort beating Opus 5 at max effort on several coding benchmarks for a fraction of the cost, so start at medium and only raise it where quality demands.

Check whether your domain triggers safeguards — if your work touches cybersecurity, biology, or frontier LLM development, apply to the relevant verification programme before assuming Opus 5.5 will handle those tasks end to end.