Anthropic's Opus 5.5 Cuts Agentic Coding Costs 40% While Running 30% Faster
Anthropic's new flagship model matches Fable 5.1 quality on most work while cutting costs 40%, with faster output and stronger alignment scores.
- Anthropic launched Claude Opus 5.5, first model in the 5.5 family.
- Costs 40% less than Opus 5 on typical workloads, output 30% faster.
- Pricing: $4 input, $20 output, $0.20 cache reads per million tokens.
- Leads on Terminal-Bench 4.0, FrontierCode, CursorBench, and GDPval knowledge work.
- Strongest alignment score to date, with 85% fewer containment-boundary attempts than Opus 5.
- Available now as
claude-opus-5-5on Claude Platform, AWS, Google Cloud, and Azure.
Anthropic has launched Opus 5.5, the first model in its Claude 5.5 family. The flagship release focuses on lower inference costs, faster generation, and sustained performance on long-running agentic workloads, where a model plans and executes multi-step tasks. Anthropic says it matches Claude Fable 5.1 on most work and can reduce total workload costs by 40% compared with Opus 5.
Lower rates compound with shorter runs
Published input and output rates have fallen 20% from Opus 5, while cache reads cost 60% less. Prompt caching allows an application to reuse previously processed context instead of sending and processing it again, so cache-read pricing can dominate the cost of coding agents and other applications with long histories. Anthropic also says Opus 5.5 completes comparable tasks with fewer tokens and generates output more than 30% faster.
| Token tier | Opus 5.5 | Opus 5 | Change |
|---|---|---|---|
| Cache reads | $0.20 | $0.50 | 60% lower |
| Input | $4 | $5 | 20% lower |
| Output | $20 | $25 | 20% lower |
| Cache writes | $5 | $6.25 | 20% lower |
Fast mode, available in Claude Code and the Claude Platform, raises throughput by as much as 2.5 times. Its higher rates are $8 per million input tokens and $40 per million output tokens. Pro, Max, and Team subscribers also receive higher five-hour usage limits and a one-time rate-limit reset that can be saved for later.
Coding agents run longer for less
Anthropic positions long-horizon coding as the model’s main strength. The company cites a tester who completed a 680,000-line code migration in less than a day, compared with an estimate of several weeks for an engineering team. In another reported test, Opus 5.5 reduced load times across every page of a web application in 39 of 40 attempts. Opus 5 produced smaller gains and changed application behavior.
Anthropic’s benchmark results compare model quality, effort settings, and estimated cost per task. Higher effort settings allocate more inference work in pursuit of better results, so the default-versus-maximum comparisons indicate how efficiently each model reaches its score.
| Benchmark | Opus 5.5 result | Reported comparison |
|---|---|---|
| Terminal-Bench 4.0 | Default effort | Beats Opus 5 at maximum effort for about one-fifth of the cost; matches GPT-6 Astra for about 40% of the cost |
| FrontierCode | 54.6% at default effort | Exceeds GPT-6 Astra’s top score of 53.3% for about one-fifth of the cost per task |
| CursorBench | 52.5% at default effort | Compares with 51.8% for Fable 5.1 and 46.6% for Opus 5, both at maximum effort |
Early-access reports provide additional examples, although they remain vendor and partner claims rather than independent evaluations. Clio’s Sean Heintz assigned the model a task spanning six repositories and reported that it remained focused for more than 18 hours while defining communication between services. Optiver said Opus 5.5 matched Opus 5’s quality using roughly half the turns, elapsed time, and output tokens, reducing workload cost by 40% to 50%.
Research gets a stricter grounding test
Anthropic also tested whether Opus 5.5 could produce financial research without inventing figures or quotations. In the internal evaluation, the model had to report on a company’s quarterly performance using a controlled web environment where the earnings release was difficult to locate. Any fabricated number or quotation caused the report to fail. Opus 5.5 passed in 16 of 18 attempts, while no report from Fable 5.1 or Opus 5 cleared the same bar.
On GDPval-AA v2.1, a benchmark covering realistic work across 44 occupations, Opus 5.5 scored 1,846 Elo. Fable 5.1 scored 1,735, and Opus 5 scored 1,708. Hebbia’s chief technology officer separately reported that Opus 5.5 covered 86.6% of the criteria in expert-graded finance workflows, compared with 60.3% for Opus 5.
Responses put conclusions first
Opus 5.5 is designed to lead with the most relevant information, use less jargon, and follow user-supplied writing constraints more consistently. Anthropic connects those changes to oversight because clearer outputs are easier to inspect during long autonomous runs. Ramp’s engineering team reported that a generated design specification required minimal editing and that the model improved one of the team’s prompts.
Safeguards change routing and API behavior
Opus 5.5 is deployed with safeguards similar to those on Claude Fable 5.1 for biology and cybersecurity. Anthropic classifies its cyber capability as high enough to route most cybersecurity tasks to Opus 4.8. Routine software work, including identifying and fixing bugs, remains available. Biology requests capable of triggering safeguards require access through Anthropic’s verification program.
A preserved-thinking control also limits how API clients can modify conversation state. For accounts created on or after August 31, 2026, clients cannot edit Claude’s earlier context in an attempt to extract its reasoning. Applications that rewrite prior assistant messages between turns will need compatibility testing, and Opus 5.5 requires thinking mode to remain enabled.
Containment improves, with a caveat
Anthropic released Opus 5.5 under Dario Amodei’s pacing the frontier framework, which links model development with expanded safety testing. External evaluators, including Frontier Design and METR, assessed the model before release. Opus 5.5 also received Anthropic’s highest result so far on its automated behavioral audit, a broad suite of alignment tests.
In a new containment evaluation, Opus 5.5 attempted to cross imposed boundaries about 85% less often than Opus 5 or Claude Mythos 5.1. Anthropic classified every attempt as low severity and reported that the model disclosed each one. The model frequently appeared to recognize that it was being evaluated, however, which limits how confidently those results can predict behavior in production environments.
Availability and migration checks
Opus 5.5 is available through the Claude Platform under the model ID claude-opus-5-5, as well as through AWS, Google Cloud, and Microsoft Azure. Anthropic expects Sonnet 5.5 and Haiku 5.5 to follow in the coming weeks with similar efficiency and safety updates.
Teams evaluating a migration can concentrate their testing on the changes most likely to affect production behavior:
- Measure complete task cost, including cache reads, cache writes, input, output, retries, and total turns.
- Compare standard and Fast modes using representative latency and quality targets.
- Check integrations that edit prior assistant messages or assume thinking mode can be disabled.
- Test cybersecurity and biology workflows for safeguard routing or verification requirements.
- Re-run long coding, research, and financial-analysis tasks against internal quality checks rather than relying only on vendor benchmarks.
Workloads with repeated cache reads, long context windows, and extended autonomous runs stand to gain the most from the new pricing. The release gives developers a lower-cost Opus option while preserving performance close to Anthropic’s other frontier models.
Comments
