LLM Price War Hits Enterprise: Opus 5.5 Cuts 20% Ahead of IPO

Every platform lead with a Claude or GPT line item on their FY26 budget just got handed a negotiation lever, and most of them don't know it yet. In a 13-day window bracketing China's Mid-Autumn and National Day holidays, at least six frontier labs are pushing next-generation models into gray testing, teaser posts, or outright API release. The pricing curve is bending faster than the capability curve, and that changes the build-versus-buy math for anyone shipping agentic workloads into production.

What Happened

The headline move belongs to Anthropic. As BigGo Finance reported, developer Lyra surfaced (via Chinese outlet 36Kr) that Claude Opus 5.5 is already in internal testing under the code name "claude-wafer-eap," with a launch potentially as early as Tuesday. Anthropic is skipping the widely expected 5.2 release entirely. Expected API pricing lands at $4 per million input tokens and $20 per million output, roughly a 20% cut against the current Opus 5 at $5 and $25. Cache reads collapse to $0.20 per million tokens, with cache writes at $5.

Alongside Opus 5.5, Anthropic is quietly gray-rolling Claude Fable 5.2, with some Fable 5.1 requests being silently redirected. Testers report Fable 5.2's 3D generation matches or exceeds GPT-6 Astra, with one developer producing a complete 3D interactive system in vanilla JavaScript from a cold prompt.

Google is running its own reconnaissance. A model labeled "gemini-3.8-flash" is showing up on LMArena with capabilities that vastly exceed anything a Flash-tier release should deliver. The community consensus, backed by leaked code name "Argon," is that this is Gemini 4 Pro in stealth. Head of developer relations Logan Kilpatrick has publicly stated Gemini 4 will be Google's largest pre-training project to date, framed as "putting Google back on top." Notably, Gemini 3.5 Pro was internally canceled for insufficient evolutionary progress.

China's side is just as loud. Stepfun released Step 5 Preview on September 20, a 600B-parameter sparse MoE with 27B active parameters, million-token context, and native vision. Moonshot teased Kimi K3.1 with a string of pi digits. DeepSeek confirmed V4.1 Pro in documentation. Alibaba's Qwen3.8-Flash-Next is functioning as a Qwen4 preview. Musk says Grok 4.7 "needs a few more days to cook."

Technical Anatomy

Strip the marketing and two structural stories emerge. The first is that cache economics, not sticker price, are where the real cost curve is bending. Opus 5.5 at $0.20 per million cache-read tokens is a genuine unlock for long-horizon agent architectures. If you're running a retrieval-augmented agent that hits the same context repeatedly across a session, your effective per-inference cost is now dominated by the tokens you generate, not the tokens you re-read. That flips the design constraint for anyone building coding copilots, compliance review agents, or multi-step research tools. The Anthropic docs already reward aggressive prompt caching, and Opus 5.5 makes that pattern close to free.

The second story is the MoE cost floor being set in China. Step 5 Preview scored 44 on the Artificial Analysis Intelligence Index, landing in the global top three among open-source models, against a peer median of 24. Its pricing is $1 in and $2.70 out per million tokens, roughly half the peer median of $2 and $10. Full weights are scheduled to open-source on October 15. That combination, top-three quality plus half-price hosted API plus open weights in three weeks, is a serious problem for any Western vendor selling mid-tier proprietary inference into cost-sensitive verticals.

The capability demos tell a related story. Opus 5.5 rendering a Waymo model down to the Jaguar badge and rooftop lidar, Gemini 4 Pro producing an interactive SVG of a pelican pedaling a bicycle, Fable 5.2 generating a 3D interactive system from a cold prompt, these aren't parlor tricks. They're evidence that the frontier is now cross-modal spatial reasoning, not just chat quality. Teams building on Gemini APIs for anything involving 3D, layout, or diagram generation should expect the ground to shift under them within the quarter.

Who Gets Burned

Three groups are exposed. First, any startup whose entire product moat is "we wrap Opus 5 and charge a markup." A 20% price cut from the upstream vendor compresses your margin the day it ships, and your enterprise buyers will notice within one billing cycle. If you can't articulate what you do beyond token arbitrage, your next board meeting gets uncomfortable.

Second, mid-tier proprietary model vendors sitting between the frontier labs and the open-weights floor. When Step 5 Preview lands at 44 on the intelligence index with open weights and $1 input pricing, the argument for a closed, mid-quality API at $3-5 input pricing evaporates. Enterprise procurement teams that ran a bake-off six months ago will re-run it, and the results won't be kind to anyone in the squeezed middle.

Third, and this is the one nobody wants to say out loud, OpenAI. GPT-6 Astra has 13% of enterprise AI spending against Claude Fable's 8%, per the numbers circulating in the reporting. Anthropic is pricing Opus 5.5 aggressively into an IPO window, which means the pricing is a market-share weapon, not a margin decision. If Fable 5.2 genuinely matches Astra on 3D and Opus 5.5 opens a real gap on code refactoring and long-horizon planning, Anthropic converts pre-IPO narrative into contracted revenue, and OpenAI has to decide whether to match the cut or defend margin. Either answer costs them.

The CFO at any Series B or later company running six-figure monthly LLM spend should be asking their VP Eng this week a very specific question: what is our contractual exit window on the current model provider, and what does a 60-day migration to Opus 5.5 or a self-hosted Step 5 actually cost in engineering hours versus inference savings? If nobody has modeled that, you're leaving money on the table by default.

Playbook for AI Development

Three concrete actions for the next 30 days. One, instrument your current LLM spend by workload, not by aggregate. You need to know which specific features are burning tokens, because the cache-read pricing on Opus 5.5 only matters if you know where your redundant context lives. Most teams don't have this telemetry and can't make an informed switching decision.

Two, run a parallel eval harness against at least one open-weights candidate before October 15. When Step 5 Preview weights drop, the teams that already have a benchmarking pipeline will make a self-hosting decision in a week. The teams that don't will spend a quarter arguing about it. If your workloads are latency-tolerant and privacy-sensitive, and this describes most fintech and iGaming compliance work, the economics of self-hosting a top-three open model on your own inference stack via Hugging Face tooling are about to look very different.

Three, renegotiate. If you have a committed spend deal with any frontier vendor signed before September, the ground has moved. Your account executive knows it. Ask for either a price adjustment matching the new tier or a contractual right to migrate workloads to the cheaper SKU without penalty. Vendors staring at a competitive pricing war will grant concessions they wouldn't have entertained six weeks ago.

Key Takeaways

• Opus 5.5's 20% price cut and $0.20 cache-read tier make long-context agent architectures materially cheaper to operate, changing the unit economics for coding and compliance agents.
• Anthropic is using pricing as a pre-IPO market-share weapon against GPT-6 Astra's 13% enterprise share, which forces OpenAI into a margin-versus-share decision.
• Step 5 Preview at $1 input pricing plus open weights on October 15 sets a new floor that squeezes every mid-tier proprietary vendor.
• Teams evaluating LLM vendor lock-in should now be asking themselves how fast they can migrate a production workload, not which model scores highest on a leaderboard.
• The 13-day launch window signals commoditization is accelerating; procurement use sits with buyers for the first time in 18 months.