Anthropic Introduces Claude Opus 5.5

Anthropic has released Claude Opus 5.5, the first foundation model in its new Claude 5.5 family. The model delivers performance comparable to Claude Fable 5.1 on technical workloads while reducing operational costs by 40% relative to Opus 5. According to the official Anthropic announcement, external evaluation teams from METR and Frontier Design tested the system prior to public release. The release emphasizes practical utility in agentic software engineering, long-horizon data work, and multi-file code refactoring. Anthropic combined lower per-token API pricing with reduced token consumption per task, which lowers compute expenses for teams running autonomous developer workflows.

Operational costs and token pricing breakdown

Opus 5.5 requires less compute to serve than its predecessor. Input and output token rates dropped by 20%, while prompt cache reads fell by 60%. Because agentic loops reuse large repository contexts across consecutive steps, lower cache read prices drive substantial cost reductions in production environments.

Pricing Tier (per 1M tokens)	Claude Opus 5.5	Claude Opus 5	Change
Prompt Cache Reads	$0.20	$0.50	-60%
Input Tokens	$4.00	$5.00	-20%
Output Tokens	$20.00	$25.00	-20%
Cache Writes	$5.00	$6.25	-20%
Fast Mode (up to 2.5x speed)	$8.00 in / $40.00 out	Not available	New option

Opus 5.5 generates text more than 30% faster than Opus 5 at standard settings. Anthropic also expanded the rolling five-hour usage limits for subscribers on Pro, Max, Team, and Enterprise tiers, adding a rate limit reset feature that users can activate on demand.

Agentic coding and software engineering benchmarks

In software engineering, Opus 5.5 establishes top scores across major coding benchmarks while maintaining lower token overhead. On Terminal-Bench 4.0, which evaluates command-line task execution in containerized environments, the model reached 66.4% accuracy, outpacing GPT-6 Astra (57.9%), Fable 5.1 (55.8%), and Opus 5 (52.3%).

On FrontierCode v1.1, which measures whether an agent’s pull request would be accepted into production codebases, Opus 5.5 scored 54.4% at default effort, beating GPT-6 Astra’s highest score (53.3%) at roughly 20% of the cost per task. On CursorBench 4.0, which tests multi-file refactoring on ambiguous developer tasks, Opus 5.5 scored 57.8%, leading Fable 5.1 (51.8%) and Opus 5 (46.6%).

Evaluation Benchmark	Claude Opus 5.5	Claude Fable 5.1	Claude Opus 5	GPT-6 Astra	GPT-5.6 Sol
Terminal-Bench 4.0	66.4%	55.8%	52.3%	57.9%	37.3%
FrontierCode v1.1 (Main)	54.4%	50.3%	48.0%	53.3%	47.5%
CursorBench 4.0	57.8%	51.8%	46.6%	(not reported)	41.7%
GDPval-AA v2.1 (Elo)	1,846	1,735	1,708	1,542	1,588
AutomationBench	40.0%	31.4%	26.9%	41.4%	28.8%
Humanity’s Last Exam (HLE)	67.7%	65.6%	63.6%	57.2%	(not reported)
Terminal-Bench-Science 0.1	58.7%	52.6%	29.0%	64.6%	22.4%
OSWorld 2.0 (Computer Use)	81.8%	80.7%	74.0%	(not reported)	(not reported)
Chartography	89.0%	88.4%	83.4%	(not reported)	(not reported)

Professional knowledge work and multimodal reasoning

On knowledge-intensive evaluations, Opus 5.5 reached 1,846 Elo points on GDPval-AA v2.1, which measures complex work across 44 professional occupations. The score places Opus 5.5 ahead of Fable 5.1 (1,735 Elo) and Opus 5 (1,708 Elo). In multidisciplinary reasoning, the model scored 67.7% on Humanity’s Last Exam when equipped with external tools.

In factual research trials with obscured source material, Opus 5.5 cleared quality thresholds in 16 of 18 attempts without introducing unverified figures or quotes. In contrast, neither Fable 5.1 nor Opus 5 cleared the same threshold in any trial. On computer use tasks measured by OSWorld 2.0, partial task completion reached 81.8%.

Communication style and concise outputs

Anthropic updated the model’s communication style to address feedback regarding verbose responses in earlier Opus versions. Opus 5.5 organizes answers with key findings up front, reduces specialized jargon, and adheres strictly to formatting guidelines. Early testers reported that code summaries and pull request descriptions required fewer manual revisions during long collaboration sessions.

Safety, alignment audits, and specialized verification programs

Opus 5.5 scored higher than any previous Claude model on Anthropic’s automated behavioral audit, an alignment suite covering roughly 2,000 simulated scenarios. In evaluations measuring containment boundary violations, Opus 5.5 attempted to circumvent sandbox constraints 85% less often than Opus 5 or Claude Mythos 5.1, with every attempt categorized as low severity and self-reported. Because the model exhibits advanced biological and cyber reasoning, Anthropic deployed safeguards derived from Claude Fable 5.1. Routine code debugging remains accessible to all users, while specialized cybersecurity tasks redirect to Opus 4.8. Organizations conducting advanced biological research can access unconstrained safeguards through the Life Sciences Verification Program, while defensive security researchers can apply to the Cyber Verification Program. To prevent automated model scraping, Opus 5.5 incorporates preserved thinking, an anti-distillation control that prevents API users from modifying past context to extract underlying reasoning steps. The system also complies with European Union AI Act watermarking rules and supports zero data retention configurations.

Platform availability and deployment

Claude Opus 5.5 is available immediately on Claude web applications, the Claude Code command-line tool, and the Anthropic API. Enterprise access is provided through Amazon Web Services Bedrock, Google Cloud Vertex AI, and Microsoft Azure. Anthropic announced that Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks.