Claude Opus 5.5 gives developers a stronger and potentially cheaper option for coding and knowledge work, but its best setting is not necessarily its most expensive one. Anthropic says the model completes comparable tasks at 40 percent lower total cost than Claude Opus 5. Benchmark results and early-use tests support a serious look at switching, while also showing why teams should measure completed work rather than rely on a headline score.
Where performance improved—and where it did not
Claude Opus 5.5 is built for agentic coding, in which a model uses tools to carry out multistep development tasks. It posts sizable gains over Claude Fable 5.1 on terminal-based coding and knowledge-work evaluations. An Elo score is a relative rating used to compare performance; a higher GDPVal-AA rating indicates a stronger result in that evaluation.
| Evaluation | Claude Opus 5.5 | Claude Fable 5.1 | GPT-6 Astra |
|---|---|---|---|
| Terminal-Bench 4.0 | 66.4% | 55.8% | 57.9% |
| FrontierCode v1.1 | 54.4% | 50.3% | 53.3% |
| GDPVal-AA v2.1 | 1846 Elo | 1735 Elo | 1542 Elo |
| AutomationBench | 40.0% | 28.9% | 41.4% |
| Terminal-Bench Science | 58.7% | 52.6% | 64.6% |
| Humanity’s Last Exam with tools | 67.7% | 65.6% | 57.2% |
The pattern matters more than any single rank. Claude Opus 5.5 leads these comparisons on general terminal workflows, FrontierCode, GDPVal-AA and Humanity’s Last Exam with tools, but GPT-6 Astra remains ahead on AutomationBench and Terminal-Bench Science. Elsewhere, Opus 5.5 scores 57.8 percent on CursorBench 4.0, against 51.8 percent for Claude Fable 5.1 and 46.6 percent for Claude Opus 5. Its smaller gains on computer use and chart recognition—81.8 percent versus 80.7 percent, and 89.0 percent versus 88.4 percent, respectively—suggest that improvement varies by task.
The cost question is bigger than token prices
An API, or application programming interface, lets software send requests to the model. API bills count tokens, the units of text the model reads and produces. Claude Opus 5.5 lowers each listed token price compared with Claude Opus 5:
| Price per million tokens | Claude Opus 5.5 | Claude Opus 5 |
|---|---|---|
| Input | $4.00 | $5.00 |
| Output | $20.00 | $25.00 |
| Prompt cache read | $0.20 | $0.50 |
| Prompt cache write | $5.00 | $6.25 |
A prompt cache stores reusable context so an application does not have to process it in full on every request. Cache reads are 60 percent cheaper, while standard input and output prices are 20 percent lower. Anthropic also says Opus 5.5 generates tokens more than 30 percent faster than Opus 5. Its stated 40 percent reduction in total cost for comparable tasks reflects not just list prices but fewer tokens used and faster execution. That figure describes the reported comparison, not a saving every deployment will achieve. Pro, Max and Team subscribers also receive five times higher usage limits.
▲ Reasoning effort and task cost
Reasoning effort—the amount of deliberation allocated to a response—changes the cost calculation again. The available settings are low, medium, high, extra-high and max. On AutomationBench, medium effort reaches approximately a 30 percent pass rate at well under $0.80 per task. GPT-6 Astra reaches 40 percent, but at nearly $2.00 per task. On Terminal-Bench 4.0, Opus 5.5 at high effort exceeds 66 percent accuracy at roughly $4.00 per attempt, while Astra at maximum effort reaches about 58 percent at more than $10.00.
More effort does not reliably mean a better answer. On FrontierCode v1.1, the Opus 5.5 medium setting, costing under $1.00, outperforms its max setting, costing over $5.00. The curve is uneven rather than steadily rising. For a team selecting a default, medium and high are more promising starting points than max; the right setting still depends on the task and the cost of an unsuccessful attempt.
What early work tests reveal
Benchmarks do not fully describe how a model behaves inside an application. In one early enterprise case, a tester migrated a 680,000-line code repository in under 24 hours. In full-stack web development testing, Claude Opus 5.5 completed 39 of 40 tasks on its first attempt. These are notable individual results, not forecasts for another organization’s repository or application.
Reviewability is another practical change. In a comparison of explanations for a billing-refactor fix, Opus 5.5 put the finding and the code changes up front, rather than leading with conversational text. That format may help people inspect changes when several agents are working at once. Improvements in visual design, front-end styling and 3D web work also appear promising, though those qualities are harder to capture in a single benchmark score.
The model’s surrounding setup matters too. A harness is the combination of system instructions, tool definitions and execution environment that lets a model act. Opus 5.5 shows improved tool calling, or choosing and using the right software tools. Anthropic has streamlined portions of its default system prompts and released a plugin evaluation suite to test whether custom tools and skills help or constrain the model. For developers, the practical lesson is to test existing instructions and tool descriptions, not assume that adding more will improve results.
Safety controls remain part of deployment
Anthropic reports its highest score to date on its comprehensive alignment evaluations, which assess how well a model follows intended behavior. Pre-release testing also involved external evaluators, automated behavioral audits and longer-running scenarios. Those checks included prompt injection, where instructions embedded in material a model processes try to redirect its behavior, and deliberately impossible tasks designed to reveal unreliable shortcut-seeking.
▲ Model safeguards and evaluations
Anthropic applies strict safeguards and access controls to advanced biology and cybersecurity capabilities. Access to certain capabilities is restricted to approved organizations through its Life Sciences Verification Program and Cyber Verification Program. For a business considering deployment, stronger coding performance does not remove the need to account for those controls or to evaluate how the model behaves within its own tools and permissions.
How to decide whether to switch
Claude Opus 5.5 makes the clearest case as a default candidate for routine coding and knowledge work: it combines stronger results on several major evaluations with lower API prices and faster token generation. It does not lead every test, and the early-use cases cannot establish how reliably it will handle a different workflow. Teams already using Claude Fable 5.1 for major planning, architecture or critical audits may want to compare those tasks separately rather than move them all at once.
A useful trial is straightforward: run representative tasks from the current workflow at medium and high effort, check the completed work and its human-review burden, and compare total cost per successful task with the existing model. If an application uses repeated context, include prompt caching in that calculation. If it relies on custom tools, use the plugin evaluation suite and review whether lengthy instructions still help. Those measurements offer a firmer switching decision than either a benchmark lead or a lower token price alone.