Claude Opus 5.5 Launches With 40% Cost Cut Ahead of Anthropic IPO

Anthropic has shipped Claude Opus 5.5, the first model in a new Claude 5.5 generation, in a launch that reads less like a routine upgrade and more like a company repositioning itself for the most scrutinized balance sheet in software history. The model went live on Tuesday across Anthropic's own apps and API as well as Amazon Web Services, Google Cloud, and Microsoft Azure, and the pitch is unusually simple: Fable-class performance at a mid-tier price2210.

What Anthropic actually shipped

The company's own framing is that Opus 5.5 performs at the level of Claude Fable 5.1, its premium frontier model, on most work, while costing roughly 40% less to run than its predecessor Opus 52130. That 40% figure combines a sticker price cut with reductions in token usage — the model is reported to need fewer tokens to complete equivalent tasks, which matters enormously for agent-style workloads where output volume, not headline price, dominates the bill241.

The per-token pricing is concrete. Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, a 20% cut from Opus 5's $5/$25, while cache reads fall from $0.50 to $0.20 per million tokens — the number that quietly reshapes economics for long-running coding agents that re-read context constantly1910. Anthropic is also adjusting Claude Code around the launch: five-hour session limits rise 20%, and because the new model is cheaper, it effectively goes 25% further within those limits, with Pro, Max, and Team users getting a usage reset to spend anytime9.

Benchmark claims from Anthropic put the new Opus ahead of both Fable 5.1 and OpenAI's more expensive GPT-6 Astra on most tasks, with third-party analysis from Artificial Analysis folded into some coverage24. Early testers reported genuinely large gains rather than the usual few-point margins: one completed a 680,000-line code migration in under a day, work that would have taken an engineering team weeks, and on a task asking the model to cut load times across a web app's pages, Opus 5.5 succeeded 39 out of 40 times where Opus 5 made smaller improvements that sometimes altered behavior in the process21. Anthropic is also promising to fix the recognizable, formulaic "Claudish" writing style that has become a running complaint among heavy users24.

Notably, the version number itself is a signal. Observers expected a modest Opus 5.2; Anthropic jumped to 5.5, and the model was spotted in stealth testing under the codename claude-wafer-eap before launch, with a leak that predicted the $4/$20 pricing exactly69. Sonnet 5.5 and Haiku 5.5 are expected in the coming weeks with similar performance, efficiency, and safety gains, which means today's release is the top of a new family rather than a one-off2224.

The safety layer is the more interesting story

What separates this launch from a generic price war is the safety architecture — and the fact that it landed in the middle of an ongoing argument about Anthropic's own practices. The company had external evaluators, including Frontier Design and METR, test Opus 5.5 before release, and on Anthropic's automated behavioral audit — described as its most comprehensive alignment test — the model is the strongest performer to date2110.

But the model also ships with a routing mechanism that is drawing skeptical coverage. Safety classifiers can silently reroute flagged requests to older models mid-workflow: cybersecurity requests that trip safeguards go to Opus 4.8, and flagged biology requests go to Opus 5101. The New Stack's take is pointed — your agent calls might secretly get downgraded without you knowing, which is a developer-experience problem as much as a safety feature1. That concern has recent history behind it: coverage from late August described an incident where Claude deleted a developer's 700 GB home directory during a safeguard test, with reporting suggesting an automatic model downgrade to Opus 4.8 may have contributed to a fatal variable collision15. The New York Times framed the whole release as arriving amid a safety debate34, and this launch is Anthropic's first since CEO Dario Amodei called for pacing the frontier of AI development — a stance it is now balancing against an IPO clock21.

Why the IPO makes this launch different

The financial backdrop explains the urgency. Anthropic is preparing a public listing that could value it above $1 trillion, with some market coverage floating figures approaching $2 trillion, and IPO marketing is expected to begin in mid-October, with a potential November listing in play22232928. For a company about to make its debut, a model that delivers frontier capability at a steep discount is not just a product — it is the gross-margin story investors will be asked to underwrite.

The pressure came from a specific direction. Reuters reported that Anthropic was weighing a new model release ahead of the IPO as OpenAI's GPT-6 Astra, launched September 3, captured roughly 13% of enterprise AI spending versus about 8% for Anthropic's Claude Fable on Ramp's transaction data252826. OpenAI also surpassed Anthropic on OpenRouter weekly spend for the first time in over two and a half years, per the same reporting25. Anthropic's fundamentals remain formidable — an annualized revenue run rate above $65 billion as of July 2026, up from roughly $9 billion at the end of 2025, and a Reuters-cited 2028 forecast of $190 billion to $200 billion — but the trend line, not the level, is what a pre-IPO company has to answer25. One additional wrinkle: Meta, a major Anthropic customer, is reportedly working to reduce its reliance on Claude by building more capability in-house28.

Against that, the Reuters-sourced table of capability is instructive. On Terminal-Bench 4.0, Claude Fable 5 scored 42.0, Fable 5.1 hit 55.8, and OpenAI's prior-generation GPT-5.6 Sol managed roughly 37 — Anthropic has been winning on capability while losing on deployment momentum, and Opus 5.5 is the obvious fix for the second problem25.

Gemini, the September pileup, and the open-weight squeeze

This launch did not happen in isolation. September 2026 has been a barrage: a tracker of new releases counts twelve models across seven labs in the first ten days of the month alone41. Anthropic opened on September 1 with Claude Fable 5.1 and its invitation-only twin Mythos 5.1; Google and Meta both shipped on September 2, with Google releasing Gemini 3.8 Flash plus a restricted Cyber variant and Meta putting out Muse Spark 1.341.

Google's position in this race is peculiar. Gemini 3.5 Pro has been announced and listed as "coming soon" with no date and no published price, and after GPT-6 Astra shipped it is reportedly the only major model still pending in one release tracker's upcoming table41. In a month where Anthropic and OpenAI are trading blows on price-performance, Google's silence is itself a data point — the Gemini line is competing on Flash-tier speed and cost while its Pro flagship waits in the wings.

The open-weight picture is more nuanced than the closed-lab drama suggests. Open-weight models now handle a majority of tokens passing through Vercel's AI Gateway, yet Anthropic still captures 64% of the spend on that platform — volume has shifted to free and downloadable weights, but revenue has not1. That is precisely the environment in which a 40% cost cut makes strategic sense: it narrows the gap that pushes high-volume workloads toward open alternatives while defending the premium spend that closed models still command. If the rumored successor models from DeepSeek, Qwen, or Llama arrive with frontier-adjacent coding capability, the price floor moves again, and Anthropic's bet is that Opus 5.5 keeps it ahead of that floor long enough to survive as a public company.

The reading

The most defensible interpretation of Tuesday's launch is that it is an IPO document disguised as a model release. Every element — the version-number jump past 5.2, the leak-accurate pricing, the Fable-parity claims benchmarked directly against GPT-6 Astra, the timing weeks before mid-October investor marketing — is arranged to answer the one question a pre-listing Anthropic cannot afford to leave open: whether it can hold capability leadership while its enterprise share slips. The coverage broadly agrees on the facts of price, performance, and safety review; it diverges on the meaning, with the financial press framing the model as valuation support and the technical press zeroing in on silent rerouting as a trust risk for developers.

My read: the cost story is real and the safety-first framing is genuine, but the rerouting mechanism will be the thing enterprise buyers actually negotiate over, because a model that silently downgrades mid-task is a model you cannot fully put in a production pipeline without monitoring. Watch two things — whether Sonnet 5.5 and Haiku 5.5 land before the IPO roadshow, and whether Google's still-unreleased Gemini 3.5 Pro forces a second price move before Anthropic's bankers finish the book. A cheaper frontier model helps an IPO story; a price war does not.