Anthropic's Claude Opus 5.5 Bets on Efficiency Over Benchmark Bragging Rights

Anthropic has released Claude Opus 5.5, the first model in what it calls the Claude 5.5 family, and the pitch is unusual for a frontier launch: the company spends more of its announcement on cost per task than on leaderboard position. Anthropic says the model costs about 40% less to run than Opus 5 on typical workloads, generates output more than 30% faster, and posted the best score of any model it has tested on its internal alignment audit. It also concedes, in its own words, that benchmark margins at this capability level have become a less reliable guide to real-world differences.

The pricing story is the headline

Anthropic published the full price list. Input tokens drop to $4 per million from $5, output to $20 from $25 — a 20% cut on both. The bigger move is cache reads, which fall to $0.20 per million from $0.50, a 60% reduction. The company notes that cache reads make up the majority of costs in agentic and coding work, which is where that discount lands hardest. Cache writes go to $5 per million from $6.25.

The claimed 40% total savings comes from two compounding effects: lower per-token pricing and, Anthropic says, fewer tokens consumed per task. That second factor is the one worth scrutinizing, because it is harder to verify and highly workload-dependent. A team whose prompts are long and whose outputs are short will see a different curve than one running multi-hour agent loops.

There is also a Fast mode for Opus 5.5 in Claude Code and the Claude Platform, offering up to 2.5x speed at $8 per million input tokens and $40 per million output tokens — double the standard rate. Anthropic is separately raising five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans, and giving subscription users a rate limit reset they can bank and spend when they choose. Claude Sonnet 5.5 and Haiku 5.5 are promised in the coming weeks with similar improvements; no date was given.

Benchmarks, with an asterisk from the vendor itself

Anthropic reports Opus 5.5 leading on agentic coding, computer use and knowledge work. Selected figures: 66.4% on Terminal-Bench 4.0, 54.4% on FrontierCode v1.1, 57.8% on CursorBench 4.0, 1846 Elo on Artificial Analysis's GDPval-AA v2.1, 67.7% with tools on Humanity's Last Exam, and 81.8% partial on OSWorld 2.0. On Zapier's AutomationBench it scores 40.0%, behind the 41.4% Anthropic attributes to GPT-6 Astra. On Terminal-Bench-Science 0.1 it scores 58.7% against 64.6% for Astra.

Several caveats are buried in the footnotes and deserve promotion. Comparison figures for GPT-6 Astra and GPT-5.6 Sol are as reported by OpenAI, not independently rerun. Terminal-Bench 4.0 carries a standard error of roughly ±2.6 points for Opus 5.5, and Terminal-Bench-Science ±3.5 to ±5 points per model — wide enough that some gaps are closer to noise than the table implies. The WANDR results from Perplexity's benchmark were run under a modified setup with offline search tools and a 980k-token budget, which Anthropic says makes them non-comparable to Perplexity's published numbers.

Most interesting is the safeguard footnote. Opus 5.5 was evaluated with production safeguards enabled, and when those safeguards intervened, cybersecurity tasks were handed to Claude Opus 4.8 and biology and frontier LLM development tasks to Opus 5. Anthropic says this likely depressed scores. Zapier's AutomationBench runs had no fallback, so interventions counted as outright failures. That is a candid disclosure, but it also means the published numbers measure a deployed system with guardrails, not a bare model — and buyers cannot easily separate the two.

Anthropic frames Opus 5.5 as performing at the level of Claude Fable 5.1 on most work, and says it is comparable to Claude Mythos 5.1 in biology and cybersecurity. The announcement does not explain what those models are or where they sit in the lineup, which makes the comparison hard to interpret from outside.

Safety positioning and the pacing question

This is Anthropic's first release since it called for pacing the frontier, and the safety framing is deliberate. External evaluators including Frontier Design and METR tested the model before release. Anthropic says Opus 5.5 is less likely than recent models to take hard-to-reverse actions or step outside given boundaries, and more resistant to prompt injection than Opus 5. On a benchmark run by security firm Gray Swan, the company says Opus 5.5 ties Fable 5.1 for the lowest prompt injection success rate tested.

Because of its biology and cybersecurity capability, the model ships with safeguards similar to those on Fable 5.1. Vetted organizations can apply to a Life Sciences Verification Program now; a Cyber Verification Program is expanding in the coming weeks. Anthropic also describes a classifier that screens every action before execution, an auditable open-source sandbox and automated code review.

The company says its broadened alignment testing now covers longer tasks, impossible tasks and scenarios modeled on real incidents, while acknowledging the suite still has limits. Details are in the system card, which was not summarized further in the announcement.

What customers are reporting

Anthropic published testimonials from a long list of early testers. GitHub's Mario Rodriguez says the model used among the fewest tokens and steps the company measured, solving more terminal tasks in VS Code than Opus 5 in under half the steps. Clio's Sean Heintz describes an unattended 18-hour run across six repositories. Optiver reports matching Opus 5's quality in roughly half the turns, time and output tokens, cutting that workload's cost 40% to 50%. Deloitte Consulting's Carl Bennett says the lowest effort setting caught 72% of known bugs in code reviews versus 56% for Opus 5 at high effort. Hebbia's Aabhas Sharma cites 86.6% rubric coverage on finance workflows against 60.3% for Opus 5.

These are vendor-selected quotes from early-access partners, measured on private evaluations with undisclosed methodology. The consistent theme across them — fewer steps, fewer tokens, less rework — is more credible than any single number, but none of it is reproducible by a third party.

Anthropic's own demonstrations follow the same pattern: a 680,000-line migration completed in under a day, a 200,000-line audit in under three hours versus 20-plus for Opus 5, an HAProxy C-to-Rust translation finished in 9.5 hours against 12 for Fable 5.1 at 51% lower cost. None of these are independently verified, and the comparison systems are Anthropic's own.

What to watch

• Whether the 40% cost reduction holds on your workload mix, or only on cache-heavy agentic patterns where the 60% cache-read discount dominates.
• How often production safeguards intervene in biology and cybersecurity-adjacent work, and what the fallback behavior costs in quality — Anthropic's own benchmark footnotes suggest the effect is measurable.
• Independent reruns of Terminal-Bench, FrontierCode and CursorBench against OpenAI's models, since the cross-vendor figures here are self-reported and several gaps sit near the stated error bars.
• Ship dates and pricing for Sonnet 5.5 and Haiku 5.5, which will determine whether the efficiency gains reach cost-sensitive production tiers.
• Whether the Cyber Verification Program's expansion meaningfully widens access, and what verification actually requires — neither the criteria nor the timeline were disclosed.