Anthropic Releases Claude Opus 5.5, Cuts Costs 40% While Matching Rival Fable 5.1

Anthropic released Claude Opus 5.5 on September 22, 2026, the first model in a new family the company says matches Claude Fable 5.1 performance on most tasks while cutting operating costs by roughly 40 percent compared to Opus 5.

The model is available now via the Claude Platform (model ID claude-opus-5-5), Amazon Web Services, Google Cloud, and Microsoft Azure. Claude Sonnet 5.5 and Haiku 5.5 are expected in the coming weeks, according to Anthropic.

Pricing and efficiency

Anthropic cut token prices roughly 20 percent: Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5. Cache read costs dropped 60 percent, from $0.50 to $0.20 per million tokens, and cache writes fell from $6.25 to $5.

Accounting for both pricing and token usage, Anthropic claims total operating costs run about 40 percent lower than Opus 5, driven partly by the model using fewer tokens and generating output more than 30 percent faster. Five-hour usage limits for subscribers increase 20 percent, and Anthropic says lower costs stretch those limits 25 percent further overall.

Benchmark claims

According to Anthropic's own benchmarks, Opus 5.5 outperforms Fable 5.1 and OpenAI's GPT-6 Astra on most measured tasks:

Benchmark	Opus 5.5	Fable 5.1	Opus 5	GPT-6 Astra
Terminal-Bench 4.0	66.4%	55.8%	52.3%	57.9%
FrontierCode v1.1	54.4%	50.3%	48.0%	53.3%
CursorBench 4.0	57.8%	51.8%	46.6%	N/A
GDPval-AA v2.1	1,846	1,735	1,708	1,542
Humanity's Last Exam (with tools)	67.7%	65.6%	63.6%	57.2%
OSWorld 2.0	81.8%	80.7%	74.0%	N/A

Anthropic claims Opus 5.5 beats GPT-6 Astra on FrontierCode at about 20 percent of the per-task cost, matches Astra on Terminal-Bench 4.0 at 40 percent of the cost, and beats GPT-5.6 Sol on CursorBench by 11 points at one-third the cost. These figures come from Anthropic's own testing and have not been independently verified.

Addressing "Claudish" writing

Anthropic says Opus 5.5 is tuned to communicate more directly — leading with key information, reducing jargon, and following instructions more closely. The change responds to persistent user criticism that Claude's writing style, nicknamed "Claudish," is formulaic and convoluted. Anthropic says early testers found the new model's output clearer, though this remains a subjective, company-reported claim.

Safety and compliance changes

Opus 5.5 is the first Opus-family model with cybersecurity, biology, and frontier-LLM-development safeguards matching Fable 5.1. Flagged requests are routed to other models: most cybersecurity tasks go to Opus 4.8, while biology and frontier-development requests go to Opus 5. Verified organizations can apply for biological research access through Anthropic's Life Sciences Verification Program; a Cyber Verification Program extension is planned.

The model ships with "Preserved Thinking," an anti-distillation measure introduced with Fable 5.1 that blocks API users from editing Claude's prior reasoning context. Anthropic says this targets "distillation attacks," where attackers allegedly use large numbers of fake accounts to extract model capabilities at scale — the company cites a September 2026 threat report on activity it says it has detected and stopped. The measure applies to accounts created on or after August 31, 2026.

Opus 5.5 also adds watermarking for EU AI Act compliance, supports Zero Data Retention, and can no longer run with "Thinking" mode disabled. Anthropic says external groups Frontier Design and METR tested the model before release, and the company is calling for stricter screening of reinforcement learning training environments and higher public-policy-driven safety standards for future frontier models.

What this means

Anthropic's pricing move is a direct response to margin pressure from OpenAI and lower-cost Chinese models that trail on benchmarks but win on cost-per-task economics. A 40 percent claimed reduction in total operating cost, if it holds up in independent testing, would meaningfully shift the price-performance calculus for enterprise coding and agentic workloads currently split between Claude and GPT-6 Astra. The mandatory "Thinking" mode and Preserved Thinking anti-distillation measure also signal Anthropic hardening its API against reverse-engineering, a growing concern as rival labs race to match frontier capability at lower training cost. None of the benchmark figures here are independently verified — they come entirely from Anthropic's own reporting.