Skip to content
NewsModel release

Anthropic Releases Claude Opus 5.5 with Fable-Level Performance at 40 Percent Lower Cost

· by Pondero Newsdesk

The short version

Anthropic launched Claude Opus 5.5 on September 22, 2026, cutting cache-read costs 60 percent and input tokens 20 percent below Opus 5, while recording the company's highest automated behavioral audit scores ever.

Anthropic Releases Claude Opus 5.5 with Fable-Level Performance at 40 Percent Lower Cost

Anthropic launched Claude Opus 5.5 on September 22, 2026, its first model release since Dario Amodei published a public call for "pacing the frontier" earlier this year. The new model matches Claude Fable 5.1 on most benchmarks, costs 40 percent less on typical workloads than its predecessor, and records the highest automated behavioral audit scores Anthropic has measured to date.

What changed

Opus 5.5 replaces Opus 5 at the top of Anthropic's API lineup. Input tokens are priced at $4 per million, down from $5 for Opus 5. Output tokens drop to $20 per million from $25. Cache reads fall to $0.20 per million from $0.50, a 60 percent reduction. Cache writes drop to $5 per million from $6.25, per Anthropic's product page. The model also generates output more than 30 percent faster than Opus 5 at standard settings.

A fast mode, available in Claude Code and the API, runs at $8 per million input tokens and $40 per million output tokens and delivers up to 2.5x the generation speed of standard mode.

Benchmark scores put Opus 5.5 ahead of Fable 5.1 across most agentic tasks, per the same page. Terminal-Bench 4.0 reaches 66.4 percent (Fable 5.1: 55.8 percent, GPT-6 Astra: 57.9 percent). FrontierCode v1.1 scores 54.4 percent versus Fable 5.1's 50.3 percent. On OSWorld 2.0 computer-use, Opus 5.5 scores 81.8 percent partial completion against Fable 5.1's 80.7 percent. Humanity's Last Exam, a multidisciplinary reasoning test run with tools, reaches 67.7 percent for Opus 5.5 against 57.2 percent for GPT-6 Astra.

Anthropic notes that benchmark margins have become a less reliable guide to real-world differences at these capability levels. The practical gap between Opus 5.5 and Fable 5.1 is narrower than the raw scores suggest.

Where it performs in practice

External testers ran real workloads on the model before launch. One completed a 680,000-line code migration in under a day, per Anthropic, work the company says would have taken an engineering team weeks. A separate tester asked Opus 5.5 to cut load times across every page of a web app. It succeeded on 39 of 40 pages. Opus 5 made smaller improvements on the same task and changed the app's behavior in the process.

Anthropic ran its own internal comparisons on two tasks. In a codebase audit, Opus 5.5 fixed a 200,000-line codebase in under three hours. Opus 5 took more than 20 hours on the same task and used 2.5 times as many tokens. In a second test, both Opus 5.5 and Fable 5.1 translated HAProxy, a widely used C-based load balancer, into Rust. Both passed nearly all of HAProxy's own regression tests. Opus 5.5 finished in 9.5 hours against 12 for Fable 5.1 and cost 51 percent less.

On cost-adjusted comparisons, Opus 5.5 beats GPT-6 Astra on FrontierCode at roughly 20 percent of Astra's per-task cost, and matches Astra on Terminal-Bench at roughly 40 percent of the cost, per Anthropic's release notes.

Safety and alignment

On Anthropic's automated behavioral audit, which runs Claude across thousands of simulated scenarios, Opus 5.5 achieved the highest scores of any model the company has evaluated, per the product page. The model is less likely than recent models to take hard-to-reverse actions or operate outside defined boundaries, and is more resistant than Opus 5 to prompt injection attacks. Anthropic extended its alignment testing this cycle to cover longer tasks, impossible tasks, and scenarios modeled on real incidents.

Pre-release evaluations were conducted by Frontier Design and METR. The full system card is published at anthropic.com.

Because Opus 5.5's capabilities in biology and cybersecurity are comparable to Claude Mythos 5.1, Anthropic deployed it with safeguards matching those on Fable 5.1. Organizations conducting biology research can apply to the Life Sciences Verification Program. Cybersecurity practitioners will gain verified access through an expanded Cyber Verification Program in the coming weeks.

Cache reads cost 60 percent less than on Opus 5

Cache reads are the dominant cost driver in agentic and coding workflows. At $0.20 per million tokens, the cache-read rate is 60 percent below Opus 5's $0.50. A team running 10 billion cache-read tokens per month would move from $5,000 to $2,000 in that line item alone.

The release also landed alongside a new GPT-6 model from OpenAI on the same day. CNBC described both as "the first models since [both companies] called for AI development slowdowns," per CNBC. Frontier pricing continues to compress, even as both vendors publicly endorsed slower development earlier in 2026. For teams evaluating which frontier model to anchor agentic coding pipelines to, the combination of the 60 percent cache-read cut and the 30 percent throughput gain makes a model migration straightforward to cost-model.

Fast mode's $8/$40 pricing is a different calculation. The 2.5x speed boost matters primarily in real-time pipelines where latency is the binding constraint, not cost-per-task efficiency. Teams without a latency problem are better served by standard mode.

Pro, Max, Team, and seat-based Enterprise plan subscribers receive an increase in five-hour usage limits with this release. Subscription users can now save and redeploy their rate limit reset at a time of their choosing, rather than having it expire automatically.

What to watch next

Sonnet 5.5 and Haiku 5.5 are expected within weeks, per Anthropic, carrying the same efficiency and safety improvements. The second variable is competitive: Opus 5.5's Terminal-Bench scores put it ahead of GPT-6 Astra on agentic coding tasks at roughly 40 percent of Astra's cost. Whether enterprise teams treat that differential as sufficient to standardize on Claude Code over competing platforms is the clearest near-term test of how the new pricing lands in practice.

Sources