OpenAI's GPT-6 Sol launch charts were out of date before they published them

Here is the least surprising fact about a model launch: the comparison chart is rigged, gently, in the vendor's favour. Here is the more interesting one: on 22 September 2026, OpenAI's comparison chart for GPT-6 Sol was rigged by accident, against a model that had already been superseded by the time anyone read it.

OpenAI's own launch page benchmarks GPT-6 Sol against Claude Opus 5 and Claude Fable 5.1, model for model, across AutomationBench, Agents' Last Exam, DeepSWE and OSWorld. Every one of those rows was already stale. Anthropic released Claude Opus 5.5 that same morning — by TechCrunch's account, about 90 minutes before OpenAI's own announcement — at the same $4-input, $20-output price Anthropic had used for Opus 5, but claiming roughly 40% lower cost on typical workloads and a real jump in coding and computer-use scores. OpenAI's chart doesn't know that happened. It can't; it was written before Opus 5.5 existed as a public model, and OpenAI published it anyway.

VentureBeat, to its credit, said this plainly rather than letting the chart stand: OpenAI's benchmarking compares Sol to the now-superseded Opus 5, and nobody — not OpenAI, not Anthropic, not an independent evaluator — has yet run both models in the same harness to say which one actually wins on cost per finished task.

The number OpenAI wants you to remember is 9%: GPT-6 Sol at maximum reasoning effort beats Claude Opus 5 at its own maximum effort on AutomationBench, a 47-tool workflow test, for 9% of Opus 5's cost per task. That's a strong claim, and it's OpenAI's own harness, own effort settings, own definition of a 'task.' It also compares against a model Anthropic had, by the time you could read the sentence, already replaced with Opus 5.5: 20% cheaper per token and, on Anthropic's own figures, 40% cheaper to run. Run the same maths against Opus 5.5 rather than Opus 5 and the '9% of the cost' figure gets a lot less dramatic — nobody has published what it actually becomes, OpenAI included.

None of this makes Sol a bad model, or the price cut fake. The 50% reduction against GPT-5.6 is real, verifiable, and — per OpenAI's own statement to VentureBeat — permanent rather than a limited-time promotion. What's worth noticing is narrower and more useful: a vendor benchmark chart is a photograph of a market that was still moving when the shutter closed. Every lab does this, in both directions — Anthropic's own Opus 5.5 comparison table benchmarks against 'GPT-5.6 Sol,' not the GPT-6 Sol OpenAI would release 90 minutes later, for exactly the same reason. Nobody can benchmark against a model that hasn't shipped yet. The honest response is to discount every comparison chart by however many hours separate the two launches, not to treat either one as the final word.

Strip out the stale comparisons and a few numbers survive contact with the same day's news. GPT-6 Sol at $2 input / $10 output lands exactly on Claude Sonnet 5's permanent pricing — not a coincidence, given Sonnet 5 is the model Sol competes with day to day — and 50% below Opus 5.5 on both input and output, whichever version of Claude you're actually comparing it to. GPT-6 Luna's $0.10/$0.50 undercuts almost everything else in its weight class bar open-weight models you'd have to self-host. And OpenAI's claim that Sol makes roughly half as many mistakes as GPT-5.6 Sol on its internal factuality test is a same-model, same-harness, before-and-after comparison — the one kind of number in this whole launch that isn't vulnerable to a rival shipping something newer an hour later.

The lesson isn't that OpenAI cheated. It's that 'launch day' benchmarking has a structural flaw the whole industry has agreed to ignore: three frontier labs now ship inside a 48-hour window often enough that any comparison chart is a bet on nobody else moving before the blog post goes live. This week, that bet lost, in public, on OpenAI's own page. Read the price. Distrust the chart until someone runs Sol against Opus 5.5 in the same room.