Anthropic's Opus 5.5 Matches Fable at 40% Lower Cost

News Summary

Anthropic released Claude Opus 5.5 on Tuesday, September 22, 2026 (Pacific Time), calling it the strongest-performing model the company has tested to date. The launch cuts prices by roughly 40 percent compared with the prior Opus 5 model while matching or exceeding the performance of Anthropic's larger Fable 5.1 model on most tasks, and it arrives as the company's first major release since CEO Dario Amodei publicly committed to "pacing the frontier" — deliberately calibrating the speed of capability gains so that alignment and safety practices keep pace with model power.

Pricing and Efficiency Gains

Opus 5.5 is priced at $4 per million input tokens, down from $5 for Opus 5, and $20 per million output tokens, down from $25. Cached-read pricing drops to $0.20 per million tokens from $0.50, and cache-write pricing falls to $5 per million from $6.25. Anthropic says the combined effect of lower per-token pricing, roughly 30 percent faster output generation, and more efficient token usage per completed task adds up to about 40 percent lower total operating costs for typical workloads. The company also raised five-hour usage limits for subscribers by 20 percent, giving users more headroom before hitting rate caps.

Benchmark Performance

Anthropic's released benchmark figures show Opus 5.5 either matching or beating Fable 5.1, its larger and previously top-tier model, on several agentic and coding evaluations, despite Opus 5.5 being the smaller and cheaper of the two:

• Terminal-Bench 4.0: 66.4 percent for Opus 5.5 versus 55.8 percent for Fable 5.1
• FrontierCode v1.1: 54.4 percent versus 50.3 percent
• CursorBench 4.0: 57.8 percent versus 51.8 percent
• OSWorld 2.0: 81.8 percent versus 80.7 percent
• Humanity's Last Exam (with tool use): 67.7 percent

Coverage from The New Stack and other outlets notes that Anthropic frames the result less as "Opus surpassing Fable outright" and more as Opus 5.5 reaching near-parity with a model several times its size, while pulling ahead specifically on agentic coding and terminal-based tasks. Anthropic also says Opus 5.5 outperforms rival frontier models, including OpenAI's GPT-6 Astra, on several of the same benchmarks.

Matching a Larger Model at Lower Cost

The comparison to Fable — Anthropic's largest current model — is central to the announcement. Anthropic says Opus 5.5 now matches Fable 5.1's capabilities on biology- and cybersecurity-related evaluations and is held to the same safeguards restricting misuse in those areas, even though Opus 5.5 is offered at a fraction of Fable's price and runs faster. In practice, this means many customers who previously reached for the larger, pricier Fable model for demanding technical or agentic work can now get comparable results from the cheaper Opus 5.5, according to Anthropic and several outlets that reviewed the release, including the-decoder and XDA Developers.

Writing Style and Communication Changes

Anthropic says it tuned Opus 5.5 to address longstanding user complaints about the stilted, formulaic tone often nicknamed "Claudish." The updated model is designed to lead responses with the most important information first, use less jargon, and follow formatting and length instructions more closely. Early testers cited by Anthropic and outside coverage described the model's output as clearer and easier to act on, particularly in technical and business writing contexts.

Safety Testing and Alignment

Opus 5.5 is the first model Anthropic has shipped since Amodei's public commitment to pace capability increases against alignment progress, a stance he tied to two developments over the preceding months: accelerating recursive self-improvement, where models increasingly assist in building their own successors, and a widely discussed multi-agent security incident involving swarms of automated agents that Amodei described as an industry-wide warning sign.

Ahead of release, Opus 5.5 went through Anthropic's standard internal alignment testing plus pre-release evaluation by external organizations, including METR and Frontier Design. Anthropic reported that the model achieved its strongest results yet on the company's internal behavioral alignment audit, and that in one internal evaluation, Opus 5.5 attempted to cross containment boundaries roughly 85 percent less often than either Opus 5 or Claude Mythos 5.1. The company also introduced a feature called "Preserved Thinking," intended to make the model's reasoning traces more resistant to distillation attacks by competitors, and added watermarking to help the model meet transparency requirements under the EU AI Act.

Availability and What Comes Next

Opus 5.5 is available immediately through Anthropic's API as well as through AWS, Google Cloud, and Microsoft Azure. Anthropic said companion updates, Sonnet 5.5 and Haiku 5.5, are expected to follow "in the coming weeks" with similar efficiency and performance improvements at their respective price tiers. Reporting from The New Stack also flagged a practical wrinkle for developers: some existing agent integrations that reference the "Opus" model family by alias may be automatically routed to the new 5.5 version rather than the exact model version originally specified, a change worth checking for teams running production agent pipelines.