Anthropic and OpenAI Cut Prices on the Same Day, Definitely Did Not Coordinate About It
Claude Opus 5.5 launched with a 20% price cut. An hour later, OpenAI answered with GPT-6 Sol and Luna at up to half off, because nothing says 'independent business strategy' like matching your rival's move within sixty minutes.
At some point on September 22, Anthropic shipped Claude Opus 5.5, cut the price by 20 percent to four dollars per million input tokens and twenty per million output, and took a victory lap on benchmarks: 66.4 percent on Terminal-Bench 4.0, up from 55.8 for the prior generation, and — Anthropic’s phrase, not ours — “the strongest performer to date” on the company’s own behavioral safety audit. Cache reads got the steepest cut of all, down 60 percent to twenty cents. A user reportedly ran a 680,000-line code migration on it in under a day. It was, by the standards of a Tuesday model release, a genuinely good one.
About an hour later, OpenAI released two models nobody had been told to expect: GPT-6 Sol and GPT-6 Luna, priced at exactly half of what their GPT-5.6 predecessors cost. Sol landed at two dollars in, ten out — which is to say, precisely half of what Anthropic had just charged for Opus 5.5, on both ends, to the dollar. Luna undercut everyone by an order of magnitude, at ten cents in and fifty cents out, instantly becoming one of the cheapest frontier-adjacent models OpenAI has ever shipped. OpenAI said Sol makes “half as many mistakes” as its predecessor, which is a specific, confident number for a category of error the entire industry still mostly describes using the word “hallucination.”
Nobody at either company has said the word “coordination,” and nobody needs to, because the timeline does the talking. A twenty-percent Anthropic cut, followed within the hour by an OpenAI cut precisely twice as aggressive on the model priced to compete directly with it, is not a coincidence in the sense of two people independently discovering fire. It’s a coincidence in the sense of two people standing in the same room, watching each other’s hands, and pricing accordingly. The most likely story isn’t a leak or a backchannel — it’s simpler and dumber than that: both labs have been watching the same market get more price-sensitive for months, both had models ready, and both were waiting to see who would blink first. OpenAI just made sure it wasn’t going to be the one still standing there an hour later looking expensive.
The benchmarks tell a story that’s harder to spin into pure victory for either side. On SiliconANGLE’s rundown, Opus 5.5 clears GPT-6 Astra — OpenAI’s actual flagship, not the two mid-tier models launched that day — “in virtually every category,” including agentic coding and knowledge work, while generating output 30 percent faster than its own predecessor. Sol and Luna, meanwhile, aren’t trying to beat Astra at all; they’re positioned a full tier below it, at 33.2 percent on AutomationBench and 68.8 on DeepSWE v1.1 for Sol, competent but not category-leading numbers for a model whose entire pitch is that it’s cheap. Which means the actual headline isn’t “OpenAI beat Anthropic” or the reverse — it’s that OpenAI decided the fight worth winning on launch day wasn’t capability, it was the invoice.
Anthropic, for its part, built in a guardrail that undercuts its own speed narrative: Opus 5.5’s cybersecurity and high-risk biology capabilities are deliberately restricted or routed to older, presumably more supervised models. The company shipped its fastest, cheapest, most capable Opus yet and simultaneously decided parts of it needed a chaperone. That’s either responsible product design or a tacit admission that “40 percent cheaper and faster than ever” and “safe enough to run unsupervised” are, for the moment, mutually exclusive claims — possibly both.
What nobody launched on September 22 was Sonnet 5.5 or Haiku 5.5 — Anthropic’s actual mid-tier and budget models, the ones that would compete head-to-head with Sol and Luna rather than with Astra. Commentators noted the gap immediately: Anthropic now has a flagship priced into a corner by a rival’s mid-tier release, and no announced answer at the tier where that rival is currently winning on price by a factor of up to forty. The price war, in other words, isn’t over. Tuesday was just the opening round, fought entirely at the top of the market, while the cheap seats — the ones actual developers building high-volume products will use — remain unsettled and, as of this writing, very much still up for grabs.
For customers, the practical upshot is a genuinely better deal than existed on Monday: more capability per dollar across two labs, on the same day, without either company having to be asked. For anyone tracking whether the AI industry runs on ruthless competitive pressure or on watching your rival’s calendar and pressing “publish” at the right moment — the evidence, this week, points to both, simultaneously, and nobody involved seems bothered by which one it actually is.
Sources: