Fable Power at Opus Prices: Anthropic's Claude Opus 5.5 Resets the Frontier Cost Curve

Just two days after OpenAI reset the frontier price curve with the half-price GPT-6 Sol and Luna, Anthropic has answered. On September 22, the lab shipped Claude Opus 5.5, the first model in its new Claude 5.5 family — and the headline number is not a benchmark. It is a cost line: Opus 5.5 delivers performance that Anthropic says “performs at the level of Claude Fable 5.1 on most work” while costing 40% less to run than Opus 5.

The timing is not a coincidence. Frontier labs are now locked in an unfamiliar war — not over who has the smartest model, but over who can deliver frontier intelligence at the price point where enterprises actually deploy it. The pressure comes from below: cheap, open-weight competitors from China and the West have been pulling the floor out of per-token pricing, and both OpenAI and Anthropic launched explicitly cheaper models this week, as CNBC noted, “as both frontier labs continue to face stiff competition from cheaper, open-weight competitors.”

What Anthropic actually shipped

Opus 5.5 is positioned as the new workhorse flagship of the Claude lineup, replacing Opus 5 for most workloads. On Anthropic’s own benchmarks, the model leads in agentic coding, computer use, and knowledge work, outperforming both Opus 5 and the larger Fable 5.1 on nearly every benchmark the company reports. It is aimed squarely at long-running, highly capable agents — the multi-hour coding sessions, tool orchestration, and document-heavy professional work that have become Claude’s franchise use case.

The API economics are aggressive:

• $4 per million input tokens and $20 per million output tokens — both down 20% from Opus 5
• Cache reads cut 60% to $0.20 per million tokens
• A 1M-token context window, per llm-stats’ model tracker
• 40% lower cost per task than Opus 5 in typical agentic workloads, per Anthropic’s own accounting

Availability is broad from day one: the model is available through the Claude API and on claude.ai for Pro, Max, Team, and Enterprise customers. Sonnet 5.5 and Haiku 5.5 are expected in the coming weeks, completing the family.

The benchmark picture

Third-party tracking puts Opus 5.5 near the top of the current stack. BenchLM ranks it #2 of 196 models at 88.45/100. On GDPval-AA v2.1, a benchmark of real work across 44 occupations, Anthropic reports 1846 Elo for Opus 5.5, compared with 1735 for Fable 5.1 and 1708 for the competition — and notably ahead of OpenAI’s GPT-6 Astra on most tasks despite being significantly cheaper, per The Decoder’s coverage.

The coding-agent story is particularly clean. Opus 5.5 leads Terminal-Bench 4.0 at 66.4%, and DataCamp’s head-to-head found that it leads GPT-6 Sol on both benchmarks the two models share, with the AutomationBench gap — 40.0% for Opus 5.5 — called “the cleanest number” in the comparison. Opus 5.5 also matches GPT-6 Astra on Terminal-Bench 4.0 for roughly 40% of the cost per task, with 5x cheaper cache reads.

“First release since we called for pacing the frontier”

The most interesting sentence in Anthropic’s announcement is not about performance at all. Opus 5.5 is, in the company’s words, “our first release since we called for pacing the frontier” — a reference to Anthropic’s recent public call for coordinated slowing of frontier development. It was tested before release by external evaluators, including Frontier Design and METR, and the announcement details the results of an automated behavioral audit run before launch.

That framing is a genuinely new pattern: a frontier lab releasing a flagship while explicitly branding it as evidence that capability pacing and shipping can coexist. Opus 5.5 is also described as Anthropic’s best-scoring model on safety evaluations to date — a claim that, if it holds up under external scrutiny, blunts the usual “safety vs. speed” framing that accompanies every frontier drop.

The “Claudish” problem

One candid admission came bundled with the launch: Anthropic confirmed it is working to fix what users call “Claudish” — the recognizable house writing style (sycophantic hedging, bulleted-list tic, em-dash-heavy cadence) that has made Claude output easy to fingerprint. The Decoder reports the company “promises to fix Claudish writing” in the 5.5 family. For enterprises rolling Claude into customer-facing copy, that is not a nicety; style fingerprinting is a real detection and brand-consistency problem.

Analysis: the frontier goes mid-market

The structural story of September 2026 is that “frontier model” and “expensive model” have finally decoupled. OpenAI’s Sol and Luna and now Anthropic’s Opus 5.5 all chase the same prize: Fable-5.1-class (or Astra-class) capability at a price where an agent can run all day without a procurement meeting. Anthropic’s move — matching its own premium tier’s performance at 40% lower task cost while beating GPT-6 Astra on most benchmarks at significantly lower prices — is the sharpest formulation of that bet yet.

For buyers, the practical takeaway is that the cost of frontier-grade agentic work just fell across the board, again. For the open-weight ecosystem, the squeeze tightens: the gap between “good enough and free” and “frontier and cheap” is now the only contested territory. And for Anthropic, Opus 5.5 is a statement that the lab can pace the frontier and still win a price war — at least until Sonnet 5.5, Haiku 5.5, and whatever OpenAI ships next.