Anthropic released Claude Opus 5.5 on Tuesday morning.

OpenAI released GPT-6 Sol and GPT-6 Luna minutes later.

Not the same week. Not the same afternoon. Minutes. Whatever release calendar these companies were running has collapsed into something closer to a reflex, and both of them cut prices on the way out the door.

Underneath the timing there is a real result, and it is not the one either press release is selling.

Opus 5.5, which Anthropic calls its new leading model and the strongest it has tested.

The price went down. $4 per million input tokens and $20 output, both 20 percent below Opus 5. Cache reads dropped 60 percent to $0.20. Anthropic puts the all-in saving at roughly 40 percent on typical workloads.

On Anthropic's own benchmarks, against Fable 5.1 and Opus 5:

Terminal-Bench 4.0: 66.4 percent, against 55.8 and 52.3.

CursorBench 4.0: 57.8 percent, against 51.8 and 46.6.

FrontierCode v1.1: 54.4 percent, against 50.3 and 48.0.

GDPval-AA v2.1: 1846 Elo, against 1735 and 1708.

Output is more than 30 percent faster. Anthropic also says it posted the best scores of any model to date on its automated behavioral audit and is much less likely to take hard-to-reverse actions. Sonnet 5.5 and Haiku 5.5 follow in the coming weeks.

Two models, split by job.

GPT-6 Sol handles complex work, coding and technical tasks. $2 input, $10 output per million tokens, down 50 percent from the 5.6 tier. OpenAI says it makes about half as many mistakes as its predecessor and reaches Astra-level reliability at much lower cost.

GPT-6 Luna is built for high volume clerical work, summarizing documents, pulling out information, answering straightforward questions. $0.10 input and $0.50 output. That is not a typo.

OpenAI credits the cuts to caching and inference improvements, and claims both models substantially outperform Anthropic's Fable and Opus models.

Sol is live in ChatGPT Work, Codex and the API. Luna is rolling out to the desktop app and to Free and Go users.

Here is where the press releases stop being useful, because OpenAI's comparison was against Opus 5, not the model Anthropic shipped that morning.

Run both new models on the same independent harness and the picture inverts.

Cost per task, measured on the Intelligence Index:

Quality, and this is the part OpenAI did not put in the announcement:

Sol's best possible result, 47.5 index points at max effort, still lands below Opus 5.5 running at its default medium setting, which scores 51.2.

Terminal-Bench 4.0: Opus 5.5 at medium hits 52.5 percent. Sol at max hits 43.9.

GDPval-AA: Opus 5.5 at medium scores 1576 Elo. Sol at max scores 1487.

AA-Briefcase: 1642 against 1483.

On business workflow automation they land in a dead heat, 61.2 against 61.7.

So Anthropic has the better model and OpenAI has the better deal, by a lot. Sol at medium costs about a fifth of what Opus 5.5 costs at medium. Whether that trade is worth it stops being a benchmark question and becomes a question about your specific workload.

One caveat worth holding onto. Anthropic's own Terminal-Bench figure is 66.4 percent while the independent run at medium effort reports 52.5. Both are real. They are measuring at different effort settings, and effort settings move these numbers more than most people realize.

On September 14, Dario Amodei published an essay called "We Must Pace the Frontier" calling for a global slowdown in AI development. He cited losing control of AI systems, cyberattacks and bioterrorism, and wrote that recursive self-improvement should be pursued very carefully, if at all.

Sam Altman, Demis Hassabis and Elon Musk all publicly agreed.

Eight days later, Anthropic shipped its most capable model ever, and OpenAI shipped two models within minutes of it.

There is a fair defense, and Anthropic makes it directly: Opus 5.5 posted the best behavioral audit scores of any model the company has tested, and is specifically less willing to take actions that cannot be undone. Pacing the frontier is not the same as stopping, and a safer strong model is arguably the point.

There is also the other reading, which is the one VC David Sacks offered when the essay landed. He argued it is simply good business for OpenAI and Anthropic to trade some raw power for reliability. Anthropic's annualized revenue has passed $100 billion, up from $65 billion in late July, with an IPO targeted for November at a valuation reported near $2 trillion.

Both readings can be true. But "we must pace the frontier" and "we shipped our strongest model in eight days" are going to sit in the same paragraph in a lot of articles this week, and the industry has not offered a version of that sentence that resolves cleanly.

The interesting thing about today is not who won. It is that the question changed shape.

For two years the frontier was a ladder. One model was best, you paid for it, and the rest were compromises. Today two labs shipped within minutes, both cut prices, and the honest answer to which one you should use is now a question about your workload rather than a leaderboard position.

If you are doing agentic coding or hard knowledge work, Opus 5.5 is measurably ahead and you will feel it. If you are running high volume tasks where the bar is competence rather than brilliance, Sol at a fifth of the cost, or Luna at fifty cents a million output tokens, will beat paying premium rates for headroom you never touch.

The practical move is to stop picking a house model. Price both against your own evals, on your own tasks, with effort settings you actually plan to run in production. That number is the only benchmark that pays your bill.

And keep the speed itself in view. Two frontier labs released four models between them in under an hour, nine days after their CEOs agreed publicly that this should slow down. Whatever is governing the pace here, it is not the essays.

The stories moving fast beyond today's headlines:

Anthropic shipped Opus 5.5 and OpenAI answered with Sol and Luna minutes later, both cutting prices hard. Opus 5.5 is the stronger model, beating Sol's maximum effort while running at its default. Sol costs roughly a fifth as much, and Luna is priced in cents.

Eight days earlier, the CEOs of both companies agreed publicly that AI development should slow down. Draw your own conclusion about which of those two facts describes the industry better.

Stay building. 🤖