Overview
Claude Opus 5.5 is the first model in Anthropic’s Claude 5.5 family, released on September 22. Anthropic says it matches Claude Fable 5.1 on most work while cutting the running cost: $4 per million input tokens and $20 per million output, against $5 and $25 for Opus 5. Cache reads fall from $0.50 to $0.20, a 60% drop, and cache reads dominate the bill in agentic and coding workloads. Combined with fewer tokens per task, Anthropic puts the total operating cost of a typical workload about 40% below Opus 5, with output more than 30% faster. It is also the first model shipped under Anthropic’s pacing the frontier policy, tested before release by outside groups including Frontier Design and METR.
Key Features
- Reworked cost structure: Token prices down 20%, cache reads down 60%, and fewer tokens used per task, which together land around 40% below Opus 5 on total cost.
- Long-horizon coding: Anthropic pitches it for codebase-wide migrations and audits. One early tester audited and fixed a 200,000-line codebase in under three hours; Opus 5 took over 20 hours and 2.5x the tokens.
- Fast mode: Claude Code and the Claude Platform offer a fast mode up to 2.5x speed at $8 input and $40 output per million tokens.
- Higher subscription limits: Five-hour usage limits rise 20% on Pro, Max, Team and seat-based Enterprise plans, plus a limit reset subscribers can save for later.
- Writing changes: Anthropic says the model communicates more naturally and puts the most important information earlier, a response to feedback from financial services, legal and software teams.
Use Cases
- Codebase migrations and audits, including cross-file and cross-language rewrites such as Anthropic’s internal C to Rust port of HAProxy
- Long-horizon coding agents that run multi-step work in the terminal and editor, where cost falls along with token usage
- Knowledge work: 1,846 Elo on GDPval-AA v2.1, covering documents, spreadsheets and analysis
- Computer use: 81.8% on OSWorld 2.0 under the partial-completion scoring
Pros
- Token prices and cache prices come down together, which matters most for long-context and multi-turn agent workflows
- 66.4% on Terminal-Bench 4.0, 54.4% on FrontierCode v1.1 main set and 57.8% on CursorBench 4.0 in Anthropic’s benchmarks
- At default effort it clears GPT-6 Astra’s top FrontierCode score at roughly a fifth of the cost per task
- At default effort it beats GPT-5.6 Sol’s best CursorBench score by 11 points at about a third of the cost
- Sonnet 5.5 and Haiku 5.5 follow in the coming weeks, filling in the rest of the price and capability ladder
Pricing
Billed per token: $4 per million input, $20 per million output, $0.20 cache reads and $5 cache writes. Fast mode runs $8 input and $40 output. Anthropic also raised five-hour usage limits on Pro, Max, Team and some Enterprise plans and gives subscribers one bankable limit reset. Weights are not released, so the model is only available through the Claude Platform, AWS, Google Cloud and Microsoft Azure, with zero data retention available. Check the official page for current prices and quotas.
Summary
Claude Opus 5.5 opens Anthropic’s 5.5 family with a focus on running cost rather than raw scores: prices down 20%, cache reads down 60% and fewer tokens per task, which Anthropic frames as 40% lower cost than Opus 5. It still sits in the top tier on both first-party and third-party benchmarks, with 66.4% on Terminal-Bench 4.0, 54.4% on the FrontierCode main set, 57.8% on CursorBench 4.0 and a top score of 58 on the Artificial Analysis Intelligence Index. For teams running agents and coding pipelines every day, the point of this release shows up on the invoice more than on the leaderboard.
Version History
- Claude Opus 5.5 fast mode exits research preview (2026-10-15): Removes the waitlist and ends the research preview; no access request or beta header needed, just set speed: fast on the API request. Uses the same weights and behavior as standard Opus 5.5 with up to 2.5x output-token speed, priced at $8 per million input and $40 per million output tokens with separate capacity limits.
- September recap: Chat and Cowork become one Claude (2026-10-08): Claude published its September recap: Chat and Cowork merged into a single Claude, with workflows that run in the cloud and continue after the laptop lid closes. Claude Docs lets a team and Claude edit the same document together. Within the 5.5 series, Opus 5.5 handles heavy tasks while Sonnet 5.5 is tuned for quick edits.
- Anthropic releases Claude Haiku 5.5 (the fastest, cheapest, and strongest small model, aimed at high-throughput, low-cost scenarios), and expands C (2026-10-07): Anthropic releases Claude Haiku 5.5 (the fastest, cheapest, and strongest small model, aimed at high-throughput, low-cost scenarios), and expands the Cyber Verification Program; existing evaluations only cover Opus 5.5 topping the Arena, with no mention of Haiku 5.5 or the project update.
- Tops Arena Text Arena at 1,509 points (2026-09-26): Arena published its Text Arena results, with Claude Opus 5.5 (High) taking first place at 1,509 points, 18 above Opus 5 (High), which now sits at No. 11. Opus 4.6 (High) held second place four points back, and Anthropic occupies the top six. At a blended input and output price of $16 per million tokens, Opus 5.5 (High) sits on the Text Arena Pareto frontier.
- Tops Code Arena: WebDev and the Artificial Analysis Coding Agent Index (2026-09-24): Arena published its Code Arena: WebDev standings, where Claude Opus 5.5 (Max) took first place at 1,818 points, 26 ahead of second-placed GPT-6 Astra (Max) and 126 above Opus 5 (Max) at 1,692. Artificial Analysis measured it at 66 on the Coding Agent Index under Claude Code max effort, six points above Opus 5, with 63.1% on Terminal-Bench 4.0, 68.4% on DeepSWE v1.1 and 66.4% on SWE-Atlas-QnA. Cost per task rose to $13.04.
- Anthropic releases Claude Opus 5.5, the first Claude 5.5 model (2026-09-22): Anthropic released Claude Opus 5.5, which it says matches Claude Fable 5.1 on most work. It costs $4 per million input tokens and $20 per million output, down 20% from Opus 5. Cache reads drop from $0.50 to $0.20, and with fewer tokens per task the total operating cost of a typical workload falls about 40% while output generation runs more than 30% faster. Benchmark scores include 66.4% on Terminal-Bench 4.0, 54.4% on the FrontierCode v1.1 main set and 57.8% on CursorBench 4.0, with a top score of 58 on the Artificial Analysis Intelligence Index. Claude Code v2.1.280 made it the default Opus model the same day, Anthropic raised five-hour usage limits on Pro, Max, Team and some Enterprise plans, and Sonnet 5.5 and Haiku 5.5 are slated for the coming weeks.