Anthropic Launches Claude Opus 5.5 at 20% Lower List Price, Claims Parity with Claude Fable 5.1

Anthropic released Claude Opus 5.5 on September 22, 2026, cutting list token pricing 20% to $4 per 1M input tokens and $20 per 1M output tokens, down from $5/$25 for Opus 5. Anthropic claims the model performs at the level of Claude Fable 5.1 "for most tasks" while costing 40% less to run than its predecessor at default settings.

The launch came roughly an hour before OpenAI shipped GPT-6 Sol and Luna, priced 50% below GPT-5.6. Opus 5.5 drew 17.2 million views on its announcement post within hours, according to engagement data from X, overshadowing the OpenAI release in social attention.

Pricing: the fine print

The headline 40% cost reduction applies only at Opus 5.5's default "medium" effort setting. Independent tracker Artificial Analysis found that at maximum effort, Opus 5.5 costs $5.98 per Intelligence Index task versus $5.86 for Opus 5 — effectively erasing the savings. The firm's breakdown shows why: higher token usage alone would push per-task cost up roughly 80% to $10.51; the 20% base price cut brings that to $8.41; cheaper cache reads at $0.20 bring the final figure down to $5.98. Vals AI separately confirmed the model frequently uses more tokens than Opus 5, particularly on coding tasks, partially offsetting the lower sticker price.

Measured against Fable 5.1, coding assistant Cline reports Opus 5.5 beats it on the Artificial Analysis Intelligence Index at roughly 2.5x lower cost — a favorable comparison, since Fable 5.1 is a larger, pricier model.

Benchmarks

According to Anthropic's own comparison table, Opus 5.5 beats Fable 5.1 on every listed benchmark and outperforms GPT-6 Astra on most. Third-party results include:

• CursorBench (Max effort): 57.8%, described by Cursor as the new top score, at 40% lower cost per task than Opus 5
• FrontierSWE (Proximal): 62.3%, second place behind GPT-6 Astra's 65.5%, ahead of Fable 5.1 (56.3%) and Opus 5 (52.0%)
• FrontierCode 1.1 (Cognition): 65.3% on the Extended benchmark, claimed as the top score
• ParseBench (table parsing, LlamaIndex): 93.9%, a 7-point gain over Opus 5, though LlamaIndex flagged the $0.058-per-page cost as too high for production OCR
• Vals Index: ranked #1, with Anthropic holding the top three spots pending GPT-6 Sol's evaluation

Anthropic's own account, via researcher Shayne Redford, claims stronger results than Astra on CursorBench, KWBench, and OSWorld, along with better instruction-following and stronger science and health capabilities.

Model changes

Anthropic says Opus 5.5 addresses the most common complaint about Opus 5: poor writing structure. The model is tuned to put key information first and adhere more closely to user-specified writing rules. Speed is up an estimated 30% over Opus 5, according to developer accounts. Claude Code, the Claude app, and Cowork now default to Opus 5.5 at medium effort.

The accompanying system card describes multi-agent scaling tests up to 100 parallel agents — a first for a publicly reported frontier system card, according to independent commentators. Opus 5.5 is also the first Opus-tier model to carry Fable 5.1-class safeguards on cyber, biological, and frontier-AI-development misuse; flagged requests fall back to another model, and Anthropic says it is working to cut false positives.

One unconfirmed claim: developer Theo (t3.gg) suggested Opus 5.5 may be a smaller model than Opus 5 based on observed behavior, crediting post-training improvements. Anthropic has not confirmed parameter count or architecture changes.

Subscription changes accompany the release: five-hour session limits rise 20%, and lower per-token pricing means existing limits stretch roughly 25% further. Pro, Max, and Team subscribers also get a one-time banked rate-limit reset. Anthropic says Sonnet 5.5 and Haiku 5.5 will follow within weeks, contradicting earlier speculation that the Haiku line had been discontinued.

What this means

The 20% list-price cut is real, but Anthropic's 40% total-cost claim only holds at default settings — a distinction easy to miss in a headline. Buyers running Opus 5.5 at maximum effort for complex agentic or coding tasks should expect costs comparable to, or above, Opus 5, not below it. The genuine competitive edge appears to be efficiency at matched capability: Opus 5.5 reportedly matches or beats the larger Fable 5.1 at a fraction of its cost, which is the more meaningful economic story than the list-price cut itself. With OpenAI cutting GPT-6 pricing 50% the same day, token pricing across frontier labs is compressing fast — the real competition is shifting to token efficiency per unit of capability, not sticker price.