Claude Opus 5.5 pricing and benchmarks, read as an agent harness decision

Opus 5.5 is the first model in the Claude 5.5 family; Anthropic says Claude Sonnet 5.5 and Claude Haiku 5.5 follow in the coming weeks with many of the same improvements to performance, efficiency and safety. Artificial Analysis lists a 1M token context window, text and image input and text output; neither source states a maximum output token limit. The two levers that set an agent harness bill, price per token and tokens per task, both move down in this release.

The per-million prices Anthropic published are below. The cache read line is the one that matters for agent loops, since Anthropic says cache reads make up the majority of agentic and coding work costs.

There is also a Fast mode in Claude Code and the Claude Platform at up to 2.5x speed, priced at $8 per million input and $40 per million output, exactly double the standard rate. Anthropic says Opus 5.5 generates output more than 30% faster than Opus 5. Artificial Analysis reports no output tokens per second figure for the model, so there is no independent latency measurement in these sources to check that against.

Anthropic's table puts Opus 5.5 ahead on seven of its nine lines: Terminal-Bench 4.0 at 66.4% against GPT-6 Astra's 57.9% and Opus 5's 52.3%; FrontierCode v1.1 main set at 54.4% against 53.3% and 48.0%; CursorBench 4.0 at 57.8% against Claude Fable 5.1's 51.8% and GPT-5.6 Sol's 41.7%; GDPval-AA v2.1 at 1846 against 1708; Humanity's Last Exam at 67.7% with tools against Astra's 57.2%; OSWorld 2.0 at 81.8% partial against Opus 5's 74.0%; and Chartography at 89.0% with tools. It trails on two: AutomationBench, run and reported by Zapier, at 40.0% against Astra's 41.4%, and Terminal-Bench-Science 0.1 at 58.7% against Astra's 64.6%.

The footnotes qualify the headline row. Terminal-Bench 4.0 carries ±2.6 points of standard error for Opus 5.5 and ±1.6 to 2 points for the other Claude models; Terminal-Bench-Science 0.1 runs ±3.5 to 5 points per model, against Astra's 5.9-point lead there. The Zapier AutomationBench runs were performed without fallback models, so safeguard interventions counted as failures, which Anthropic says produced a lower score than Opus 5.5 would achieve in practice. On the main table, when safeguards intervened, cybersecurity tasks were completed by Claude Opus 4.8 and biology and frontier LLM development tasks by Claude Opus 5, which Anthropic says likely reduced the reported scores. Anthropic states that benchmark margins have become a less reliable guide to real-world differences, and that in its own use the gap to Fable 5.1 is narrower than these scores suggest.

Anthropic reports Opus 5.5 at default effort beating Opus 5 at max effort on Terminal-Bench 4.0 for about a fifth of the cost per task, and matching GPT-6 Astra for about 40% of the cost. On FrontierCode at default (medium) effort it reports 54.6%, above Astra's top score of 53.3%, for roughly a fifth of the cost per task. On CursorBench at default effort it reports 52.5% against Fable 5.1's 51.8% at max effort, and beats GPT-5.6 Sol's top score of 41.7% by 11 points for about a third of the cost. The chart captions and the summary table do not agree - the table lists 54.4% on FrontierCode and 57.8% on CursorBench against the captions' 54.6% and 52.5% at default effort - and the release does not reconcile them.

Long-running jobs show the same pattern in wall-clock terms. An early tester audited and fixed a 200,000-line codebase in under three hours where Opus 5 took over 20 hours and used 2.5x as many tokens. In an internal test translating HAProxy from C into Rust, both Opus 5.5 and Fable 5.1 passed nearly all of HAProxy's own regression tests, but Opus 5.5 finished in 9.5 hours against 12 and cost 51% less. One tester completed a 680,000-line code migration in less than a day.

Early access partners reported the saving as fewer calls rather than cheaper calls, which is the shape that matters if your harness pays a fixed overhead per turn. GitHub says that in VS Code the model solved more terminal tasks than Opus 5 in less than half the steps, and used among the fewest tokens and steps GitHub measured across Copilot CLI and VS Code. Lovable reports finishing in a third to half fewer steps. Optiver reports matching Opus 5's quality in about half the turns, time and output tokens, cutting the cost of that workload by 40 to 50%. Quantium describes a complex coding task that previously took 38 prompts over four days coming in at 11 prompts over three hours.

Anthropic's efficiency claim is that Opus 5.5 uses fewer tokens per task, and its per-task cost figures are measured at default effort. Artificial Analysis ran the adaptive-reasoning, max-effort, default-fallback configuration and reports 260M output tokens generated across its Intelligence Index against a median of 88M, which it labels very verbose, ranking the model 95th of 212 on verbosity. That evaluation cost $8,708.20 in total, or $5.98 per Intelligence Index task, with a 95% cache discount listed. Both numbers can hold at once, and the effort setting decides which of them a harness inherits.

Commenters on Hacker News are skeptical that the benchmark deltas mean much: one points out that Opus 5 was also described as Fable-level and says they no longer think benchmarks tell us very much. Another asked for token efficiency data before assessing the release, citing rumors that it generates more output tokens per task, which the Artificial Analysis verbosity figure answers only for the max-effort configuration. Several treated Opus 5's prose style as the real blocker, one saying most of their team went back to Opus 4.6, while one commenter who pasted a chunk of code into Opus 5.5 reported the output was not substantially different. Anthropic's release says Opus 5.5 puts the most important information up front and that early testers found its writing clearer, which is the only evidence offered on that point.

The same thread surfaced the day's other operational fact. Anthropic's status page logged elevated errors from 5:50pm PT / 00:50 UTC to 7:10pm PT / 02:10 UTC affecting Claude Mythos 5.1, Claude Fable 5.1 and Claude Opus 5 across claude.ai, the Claude API, Claude Code and Claude Cowork, now resolved. Commenters argued over whether that counted as an outage or a service degradation; the incident record shows Opus 5 as the last model to return to normal success rates. Opus 5.5 is not named in the incident.

OpenAI published a post titled GPT-6 Sol and Luna on the same day, 22 September 2026. Anthropic's comparison table names neither model: it lists GPT-6 Astra and GPT-5.6 Sol, with those figures as reported by OpenAI and Astra's Terminal-Bench 4.0 result run at high effort. These sources carry no benchmark result and no price for GPT-6 Sol or Luna, so the cost-per-task comparisons hold only against Astra and GPT-5.6 Sol until someone runs the new pair.

The case for moving a harness is strongest where cache reads dominate the bill and tasks run long enough that fewer turns compounds: the cache read line alone is 60% cheaper, and GitHub, Lovable, Optiver and Kiro all report roughly halving steps or tokens on real workloads. The case is weakest on AutomationBench-style SaaS workflows, where Astra scored 41.4% to Opus 5.5's 40.0%, and on Terminal-Bench-Science 0.1, where Astra leads by 5.9 points inside a ±3.5 to 5 point standard error. Security posture is a separate argument: Anthropic reports a classifier that screens every action before it runs, an open-source sandbox security teams can audit, code review that catches vulnerabilities before they merge, and a tie with Fable 5.1 for the lowest prompt injection success rate on a benchmark run by Gray Swan. Anthropic says the model was tested before release by external evaluators including Frontier Design and METR.

What these sources do not give you: a maximum output token limit, an independent tokens-per-second figure, a release date for Sonnet 5.5 or Haiku 5.5 beyond the coming weeks, or any numbers for GPT-6 Sol and Luna. Access for biology work runs through the Life Sciences Verification Program, with the Cyber Verification Program expanding in the coming weeks, so safeguard interventions and their fallbacks to Opus 4.8 and Opus 5 are a live variable in regulated domains. On the subscription side, Anthropic raised five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans and made the rate limit reset something a user can save and trigger later; one commenter on Hacker News reports a Reset for free button in the Claude app that expires 22 October.