# Anthropic's Claude Opus 5.5 Matches Its Larger Sibling While Costing 40 Percent Less to Run

Priced at $4 per million input tokens and $20 per million output tokens, the model scores 89.9 percent on SWE-bench Pro and 66.4 percent on Terminal-Bench 4.0, according to Anthropic's own figures.

- Published: 2026-09-27T06:30:19.445Z
- Canonical: https://polylog.news/ai/2026-09-27/anthropic-s-claude-opus-5-5-matches-its-larger-sibling-while
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Anthropic News](https://www.anthropic.com/claude-opus-5-5), [Anthropic News](https://www.anthropic.com/claude-fable-and-mythos-5-1)

Anthropic released [Claude Opus 5.5](https://www.anthropic.com/claude-opus-5-5) on September 22, positioning it as matching [Claude Fable 5.1](https://www.anthropic.com/claude-fable-and-mythos-5-1)'s performance on most tasks while costing 40 percent less to run than Opus 5 and generating output more than 30 percent faster. Pricing is [$4 per million input tokens and $20 per million output tokens](https://www.benzinga.com/markets/private-markets/26/09/61933271/anthropic-launches-claude-opus-5-5-cuts-costs-40-and-scraps-5-hour-usage-caps), with cache reads at $0.20 per million. Anthropic also removed the five-hour usage caps on its subscription tiers.

The published scores, compiled by [third-party trackers](https://www.vellum.ai/blog/claude-opus-5-5-benchmarks-explained), include 89.9 percent on SWE-bench Pro, 66.4 percent on Terminal-Bench 4.0, and 1846 Elo on GDPval-AA v2.1. The computer-use figure deserves attention. On OSWorld 2.0, the model scores 81.8 percent under partial credit but only 48.7 percent under strict completion. That gap measures how far agentic desktop work still is from reliable, fully automated execution, and it comes from Anthropic's own testing. No independent lab has reproduced it.

The model is available through the Claude platform as well as Amazon Web Services, Google Cloud and Microsoft Azure. Anthropic says external evaluators tested the model before release and that it scored better than its other models on the company's automated behavioral audit.

## What this means

What matters here is capability per dollar, not capability alone. A cheaper model that holds the prior tier's quality resets the price of every agentic workload that calls a model thousands of times in a loop, which is where most token spending concentrates. Anthropic gains ground in coding and long-running agent tasks, where it already has strong developer adoption. The parties exposed are inference resellers and application companies whose margins were built on the old price of frontier-quality output, along with rival labs that now have to compete on price rather than on a new benchmark score. The precedent is the pattern of the past year, in which each lab's cheaper mid-cycle model was matched within weeks by a competitor's price cut, compressing profit margins across the serving layer. The likelier outcome in the coming weeks is a matching price move from OpenAI or Google on their business-tier models, provided they have the capacity to serve it. This reading would be wrong if Anthropic quietly limits how much Opus 5.5 throughput it makes available, which would mean the cost savings come from hardware allocation rather than genuine model efficiency.

## What to watch

- Whether independent evaluations reproduce the SWE-bench Pro and Terminal-Bench figures, since matching the larger model's performance is the basis for the entire price claim.
- Whether OpenAI or Google cuts prices on a comparable tier within weeks, which would confirm that price, not capability, is now the main point of competition.
- Whether the strict OSWorld completion rate improves in the next release, since that number determines whether computer-use agents can be trusted without a person reviewing each step.