- Published on
Claude 5.5 Released, Performance Nearly Matches Fable, and It's Competing on Price Too
- Authors
- Name
- DeepHorus
Anthropic has officially launched Claude Opus 5.5, the first model in the Claude 5.5 family, with Sonnet 5.5 and Haiku 5.5 set to arrive in the coming weeks. According to the official announcement, it matches Claude Fable 5.1 performance on most tasks, cuts typical task costs by 40% compared to Opus 5, and boosts output speed by over 30%. The model is now available on Claude, Claude Code, and Claude Platform, as well as on AWS, Google Cloud, and Microsoft Azure. Its API model name is claude-opus-5-5, and it has become the default model for Claude paid plans.
This is no routine half-generation upgrade. Anthropic positions it as the workhorse model for long-running coding agents and knowledge work, with only about two months between Opus 5 and Opus 5.5. Official benchmarks show Opus 5.5 scoring 66.4% on Terminal-Bench 4.0, ahead of GPT-6 Astra's 57.9% and Opus 5's 52.3%; 54.4% on FrontierCode v1.1, slightly above Astra's 53.3%; and improving from 46.6% to 57.8% on CursorBench 4.0. In the knowledge work benchmark GDPval-AA v2.1, it scored 1846 Elo, surpassing Fable 5.1's 1735 and Opus 5's 1708. On OSWorld 2.0, it climbed from 74.0% to 81.8%.
The advantages do not extend across every category. On AutomationBench, Opus 5.5 scored 40.0%, slightly below GPT-6 Astra's 41.4%. On Terminal-Bench-Science 0.1, it scored 58.7%, still trailing Astra's 64.6%. Anthropic cautions that gaps on leaderboards among frontier models no longer fully capture the real-world experience, noting that the gap with Fable 5.1 in internal use is smaller than some evaluations suggest.
What truly showcases the upgrade are long tasks lasting hours. Early testers used Opus 5.5 to complete in a single day a migration involving 680,000 lines of code; an audit of a 200,000-line codebase took less than three hours, while Opus 5 took over 20 hours to complete a comparable task, consuming roughly 2.5 times as many tokens. When rewriting HAProxy from C to Rust, Opus 5.5 took 9.5 hours versus Fable 5.1's 12 hours, at 51% lower cost. In quarterly earnings report testing, 16 of the 18 reports submitted by Opus 5.5 passed, while Fable 5.1 and Opus 5 never met the same standard.
A more tangible change than any benchmark comes from pricing. Opus 5.5's API input and output prices are $4 and $20 per million tokens respectively, 20% lower than Opus 5; the cached read price drops from $0.50 to $0.20, a 60% reduction. The model also offers a Fast mode with speeds up to 2.5x, priced at $8 per million input tokens and $40 per million output tokens, supporting a 1 million-token context and a maximum output of 128,000 tokens. Anthropic raised the five-hour usage limits for Pro, Max, and Team plans by 20%. ClaudeDevs notes that the lower pricing will further extend usable time under the same quota by about 25%, and subscribers get an additional quota reset opportunity valid through October 22. However, some developers, citing tests from Artificial Analysis, point out that Opus 5.5's output token consumption in max mode ranks highest among tested models.
The community's earliest hands-on tests focused on creative coding and interactive simulations. After a week of continuous use, Wes Bos concluded that Opus 5.5 had reached Fable-level performance, and had the model render 500 pieces of fitness equipment to build a gym builder. BridgeMind used a single prompt to generate a playable Mario Kart game. Jake Eaton had it write approximately 7,500 lines of Python to draw various art styles pixel by pixel. Creative ability does not dominate across the board: Auggie's Bach Benchmark revealed issues with voice spacing and parallel octaves in its four-part chorales. Wharton School professor Ethan Mollick called it the first model that made him feel close to Fable-level without belonging to the Fable or Astra family, though the recently common problem of overly dense text output from Claude has not completely disappeared.
This is also the first model released after Anthropic proposed slowing down the pace of the frontier race. Before the launch, external organizations including Frontier Design and METR participated in evaluation, and the company also ran automated behavioral audits covering nearly 2,000 simulated scenarios to check for boundary-crossing, deception, and high-risk behavior. Official assessments show it attempts to push beyond safety boundaries about 85% less frequently than Opus 5 and Claude Mythos 5.1, though Anthropic acknowledges the model can sometimes detect that it is being tested. Since its biological and cybersecurity capabilities are already close to Mythos 5.1, the model is equipped with safety measures comparable to Fable 5.1: most cybersecurity tasks are diverted to Opus 4.8, and anti-distillation preserved thinking, a zero-data-retention option, and pre-action classifier checks are enabled.
From Opus 5 to Opus 5.5, the model designation only moved forward half a step, but the competitive logic of the model market has already changed. As the gaps between leading models on common tasks narrow, vendors can no longer sustain premium pricing on a handful of leaderboard results, and competition is shifting toward call frequency, sustained working duration, and enterprise budgets. Continued price reductions for flagship capabilities mean advanced intelligence is transitioning from a scarce service into a general-purpose productive resource available at scaleāand Personal Agents thereby gain the economic foundation to run around the clock.
