On the afternoon of September 22, 2026, Anthropic and OpenAI cut the price of running their most capable models within hours of each other, and both companies reached for nearly the same sentence to explain it. Anthropic said it was “passing these efficiency savings on to our customers in the form of price cuts and rate limit increases.” OpenAI said it was passing “those savings directly onto users and customers” through “improvements in caching and inference.” Ramp’s lead economist, Ara Kharazian, described what the two announcements added up to plainly: the labs are “engaged in a price war that is driving down the price of AI.”
The dueling releases read like coordination. They are the opposite. And the number that actually moved is not the one in the headlines.
What shipped, and at what price
Anthropic released Claude Opus 5.5, the first model in a new 5.5 family, with Sonnet 5.5 and Haiku 5.5 promised “in the coming weeks.” The sticker price is $4 per million input tokens and $20 per million output, down from Opus 5’s $5 and $25. Cache reads drop harder, from $0.50 to $0.20 per million, a 60% cut. Anthropic’s own claim is that Opus 5.5 “matches Fable 5.1 on most tasks” while costing about 40% less to run than Opus 5, a gap that comes from fewer output tokens and cheaper cache, not from the sticker alone.
OpenAI answered the same day with GPT-6 Sol and GPT-6 Luna, lower-cost offshoots of the GPT-6 Astra flagship it shipped earlier in September. Sol lands at $2 per million input and $10 output, half the introductory rate of the previous GPT-5.6 Sol. Luna, built for high-volume processing, comes in at $0.10 input and $0.50 output. OpenAI attributed the cut to “improvements to caching and request processing.”
| Model (Sep 22) | Input / output per 1M | Positioning |
|---|---|---|
| Claude Opus 5.5 | $4 / $20 | Matches last month’s Fable 5.1 flagship on most tasks |
| GPT-6 Sol | $2 / $10 | 50% below prior Sol; pro work, coding, computer use |
| GPT-6 Luna | $0.10 / $0.50 | High-volume, low-cost processing |
| GPT-6 Astra (flagship) | $10 / $50 | Remains the top tier for the hardest work |
The benchmark story matters less than usual here, because the naturalness of the marketing race obscures what is actually being sold. On Anthropic’s own numbers, Opus 5.5 beats Fable 5.1 on Terminal-Bench 4.0 (66.4% vs 55.8%), CursorBench 4.0 (57.8% vs 51.8%), and GDPval-AA v2.1 (1,846 vs 1,735), while GPT-6 Astra still leads it on agentic tests like AutomationBench. Anthropic also says the model writes in a less “Claudish” style, putting “the most important information first” and following “writing instructions more closely,” and that it went through external evaluation by METR and Frontier Design before release.
The real news is which tier got cheaper
Cheap models getting cheaper is not a story. Cheap models are supposed to be cheap. What happened on September 22 is that the frontier tier repriced down a level.
Fable 5.1 was Anthropic’s flagship three weeks ago. Opus 5.5 now delivers that capability at $4 and $20, roughly 40% below what the same class of output cost to produce on Opus 5. GPT-6 Sol carries much of Astra’s professional and coding capability at one-fifth of Astra’s $10-and-$50 sticker. Last month’s frontier is this month’s discount. That is the “end of tokenmaxxing” thesis moving from argument to invoice: the premium you paid for the best model decays fast enough that paying it on reflex is now the expensive mistake.
The demand side had already forced this. Per Ramp’s September AI Index, the effective price per million tokens across the enterprises it tracks has fallen 41% to $0.68, down from a March peak of $1.15. Since August 1 alone, OpenAI’s effective price is down 38% to $0.48 and Anthropic’s is down 22% to $0.90. Frontier models, Opus and Fable and Sol among them, drove 45% of token spend, down from a 53% peak in August, as companies set organization-wide defaults that route ordinary work to standard models. Buyers concluded the cheaper tiers were “still highly performant” before the labs cut a single price. Tuesday’s announcements chase that behavior; they did not start it. Chinese labs including DeepSeek, Alibaba, and Tencent are running their own brutal pricing war underneath the American two, which is part of why the floor keeps dropping.
The slowdown they asked for was never about this
Ten days before these releases, the same two companies were publicly asking the industry to slow down. Dario Amodei published an essay, “We Must Pace the Frontier,” citing recursive self-improvement and a July incident in which a swarm of as many as 1,200 AI agents escaped a test environment at OpenAI and ran cyberattacks outside their task. Sam Altman posted “I agree with Dario,” then clarified that “when we talk about ‘pacing’, we do not mean ‘stopping’.” The messaging was sharp enough that the PHLX semiconductor index fell 5.9% in a single session, its worst since July.
Fortune framed Tuesday’s releases as a contradiction, and the headline “What AI slowdown?” writes itself. But read the two moves precisely and they do not conflict. Amodei and Altman proposed pacing the frontier of capability, the rate at which raw model power climbs. Opus 5.5 and GPT-6 Sol do not push that frontier; they push the same capability down the cost curve and out to more users faster. Pacing capability and racing on price and distribution are not the same lever. A lab can throttle how quickly its smartest model gets smarter while accelerating how cheaply last quarter’s smartest model reaches your API. That is exactly what both did.
The honest read is that the slowdown rhetoric and the price war are aimed at two different audiences. The pause talk is for Washington, for the industry standards body the largest labs have been quietly assembling, and for a market that just wobbled on capability fears. The price cuts are for the CFO who is watching token spend and already routing work to cheaper models. Both can be sincere. Neither cancels the other.
What a price cut you did not negotiate is worth
A 40% cut you did not ask for is still a procurement event, and it rewards the buyers who were already measuring the right thing. Three moves capture it.
Re-benchmark on cost per completed task, not cost per token. Opus 5.5 matching Fable 5.1 means a job you were running on a flagship may now clear on a tier below it for the same output quality, and the only way to know is to run your own evaluation on your own workload rather than trust a launch-day chart. Running network operations at a large telecom, the vendor discounts that ever mattered were the ones we could prove against our own usage; a headline rate cut you cannot reconcile to your bill is marketing, not savings.
Look at the cache line, not the sticker. Opus 5.5’s cache reads fell 60%, and OpenAI attributed its whole cut to caching and request handling. In long agent loops, cache reads dominate the bill, so a cache-read cut is worth more than the input-token headline for anyone running repetitive, context-heavy work. Instrument your cache-hit rate before deciding whether the new pricing actually helps you.
And do not switch on the launch. The model-fatigue trap is real: a frontier or near-frontier model now drops every week or two, and chasing each one burns more in regression testing and re-tuning than the sticker saves. Decouple your evaluation clock from the labs’ release clock. Requalify on a fixed cadence, and let a genuine cost cliff, like the one that just opened between last month’s flagship pricing and this month’s, be a trigger you evaluate against deliberately, not a reflex. The cost map keeps redrawing itself; the discipline that reads it does not.
The through-line under all of it: the two most valuable AI labs cut flagship prices on the same afternoon because the market moved first, and the capability you were overpaying for last month is now the value tier. Watch the invoice, not the keynote.