Claude Opus 5.5: What Is New and What It Means for Business

Anthropic has launched Claude Opus 5.5, a model it describes as performing at the level of the larger Claude Fable 5.1 on most work while costing around 40% less to run than Opus 5 at default settings. It is available today, and the published rates fall to $4 and $20 per million tokens.

The release arrives two months after Opus 5, which Anthropic published on 24 July 2026, and it lands on the same day as OpenAI's GPT-6 Sol and Luna announcement, which we cover in a separate article. Two frontier releases within hours of each other is a useful reminder that model selection is now a recurring review rather than a one-off procurement decision.

Opus 5.5 is not being switched on by default across Owlpen at launch. More on that below. The rest of this article covers what has changed, what the published figures do and do not tell you, and what we would check before moving production work onto it.

What has changed from Opus 5

This is an iteration within the Opus 5 generation rather than a new foundation model family. The changes cluster around cost, speed, and agentic reliability.

Lower published rates

Opus 5.5 is listed at $4 per million input tokens and $20 per million output tokens, against $5 and $25 for Opus 5. Cache reads fall from $0.50 to $0.20 and cache writes from $6.25 to $5. A fast mode is listed separately at $8 and $40.

Fewer calls and fewer tokens per task

Anthropic attributes the headline 40% cost reduction to more than the rate card. It reports that Opus 5.5 completes work with roughly 40% fewer calls and about half the tokens of its predecessor, alongside output generation more than 30% faster. For agentic workflows, where one instruction can trigger a long chain of steps, token efficiency usually matters more to the invoice than the unit rate does.

Agentic coding and computer use

On Terminal-Bench 4.0, an agentic coding benchmark, Anthropic reports 66.4% for Opus 5.5 against 52.3% for Opus 5, with TechCrunch reporting 55.8% for the larger Fable 5.1. Anthropic also reports 54.4% on FrontierCode v1.1 (48.0% for Opus 5), 57.8% on CursorBench 4.0 (46.6%), and 81.8% partial credit on OSWorld 2.0 for computer use (74.0%).

Knowledge work and charts

For work closer to ordinary professional tasks, Anthropic reports 1846 Elo on GDPval-AA v2.1 against 1708 for Opus 5, and 89.0% with tools on Chartography against 83.4%. It also describes a change in how the model writes, with less jargon and important information placed earlier in a response.

Effort settings

Benchmark results are quoted at default (medium) effort, with several figures reported using adaptive thinking at maximum effort. Anthropic notes that the model is no longer available with thinking mode switched off, so anything in your configuration that assumed a non-thinking path needs checking.

Who can use it

Anthropic lists availability across the Claude platform, Amazon Web Services, Google Cloud, and Microsoft Azure, and on the Pro, Max, Team, and Enterprise plans, with increased five-hour usage limits. It is available in Claude Code, and zero data retention is offered for organisations that need it. Anthropic has also said that Sonnet 5.5 and Haiku 5.5 are expected in the coming weeks, which is relevant if your workloads sit on the mid or small tiers and you are weighing whether to migrate now or wait.

What it means in practice

The combination that matters commercially is a lower rate alongside lower token consumption. If both hold on your own material, the effect compounds, and work that was previously routed to a cheaper tier purely on cost grounds may be worth re-testing on the flagship. That is a routing question rather than a blanket upgrade, and it should be settled with measurement rather than with a rate card.

The efficiency claims also deserve validation in context. Fewer calls and fewer tokens are averages across Anthropic's own workloads. Prompts that were tuned for Opus 5 behaviour, particularly long context pipelines and tool-heavy agents, can behave differently on a new model, and the sensible sequence is to re-run a representative sample before changing any default. Anthropic itself cautions that benchmark margins have become a less reliable guide to real-world differences, which is a candid point and worth taking at face value.

Cheaper and faster inference also tends to increase usage. Where a lower unit cost removes a practical brake on agent runs, total spend can stay flat or rise even as the per-token price falls. Budget caps and per-workflow limits remain the effective control, not the rate.

Safety, alignment, and stated limits

Anthropic reports that Opus 5.5 achieved the best scores of any model it has tested on its automated behavioural audit, and that the model attempted to circumvent boundaries around 85% less often than Opus 5. It says the model was tested before release by external evaluators including Frontier Design and METR, and that it carries safeguards equivalent to Fable 5.1 for cybersecurity and biology.

The stated limits are as relevant as the claims. Anthropic notes signs that Opus 5.5 often suspects it is being evaluated, which it says challenges its ability to assess how the model will behave in ordinary use, and it describes building evaluations that reliably catch every failure before deployment as an unsolved problem. For organisations with AI governance obligations, that is an argument for keeping human review, audit trails, and escalation paths in place regardless of how a model scores on an alignment audit. The risk of hallucination is reduced by better models, not removed by them.

Owlpen and Claude Opus 5.5

Claude models are routed into configured Owlpen workflows through Anthropic API access, as described when Claude Sonnet 5 became available. Opus 5.5 is not being enabled by default across Owlpen at launch. Coaley Peak can assess it for eligible client deployments where Anthropic model access is approved, and where the capability gain justifies any cost, latency, or governance change.

Where clients ask us to review routing, we look at the work rather than the rate card. The questions we would expect to answer are whether the efficiency gain holds on the client's own documents, whether output quality is stable across a representative sample, whether latency changes affect any interactive step, and whether data handling terms are unchanged for the relevant deployment channel.

Sources used

This article is based on Anthropic's Claude Opus 5.5 launch post for pricing, benchmark figures, availability, effort settings, and alignment testing, and on TechCrunch's launch-day report for the Fable 5.1 comparison figure, the Opus 5 release date, and the expected Sonnet 5.5 and Haiku 5.5 timing. All figures are those reported by Anthropic unless stated otherwise.

If you would like to discuss model selection, AI cost control, or the Owlpen platform in the context of your own workloads, contact us at enquiries@coaleypeak.co.uk or read more about the Owlpen platform.