Opus 5.5 cuts 40%, GPT-6 Sol and Luna halve: the AI price war

The same morning, minutes apart

On the morning of 23 September, Anthropic introduced Claude Opus 5.5 as its new flagship. Almost before the page had finished loading, OpenAI shipped GPT-6 Sol and GPT-6 Luna. Two launches on the same morning is the kind of coincidence that has a marketing department behind it.

Both announcements spend most of their word count on scores. It’s the price list that tells you what is actually going on. One note on method: every figure below comes from Next’s write-up published that same morning, the only source for this article.

Anthropic cuts the price of its top of the range

Opus 5.5 lands at $4 per million input tokens and $20 per million output, with cache reads at $0.20 per million. Anthropic claims roughly 40% less than Opus 5 on typical workloads, and answers generated 30% faster.

That’s the solid part of the announcement, because the rest defuses itself. The company puts its model ahead of Fable 5.1 and GPT-6 Astra on its own tests — except on agentic search (Terminal-Bench-Science 0.1) and on AutomationBench, which measures agents under real working conditions — then concedes that the gap with Fable 5.1 is “smaller” than the scores would have you believe. It even writes that “the gaps between benchmark scores increasingly fail to reflect the differences observed in real-world situations”. Rare, and honest: the leaderboard matters less than the invoice.

One detail worth knowing before you wire anything up: Opus 5.5 inherits Fable 5.1’s safeguards on cybersecurity, biology and distillation, and most cybersecurity-related tasks get rerouted to Opus 4.8. Fixing a bug in your own code still works, but in that territory the model you call isn’t necessarily the one that answers. Going further means going through the Cyber Verification Program.

OpenAI isn’t fighting the same fight

Sol and Luna aren’t there to take on Opus 5.5 — that job still belongs to GPT-6 Astra. Sol targets everyday tasks, Luna speed and tight budgets on high-frequency work. Both arrive at half the price of GPT-5.6: $2 in and $10 out for Sol, $0.10 and $0.50 for Luna. Long-context caching improves too, with a 90% discount on input tokens already in cache.

The number that stings is somewhere else. On AutomationBench, Sol scores 33.2% at $0.27 per task, where Claude Opus 5 — the predecessor, not the new one — tops out at 26.9% for a stated cost about eleven times higher. The model explicitly sold as the weaker one does better, for a fraction of the money.

A range stretching out at the bottom

Laid end to end, these prices sketch a market that no longer segments by performance but by price per token. From Luna at $0.10 in to Opus 5.5 at $4, that’s a factor of forty inside a single day of announcements. A useful yardstick for the scale: on 7 September, Next noted that one single prompt on GPT-6 Astra could run to $33.

So the real architecture question is no longer “which model is best”, but “at what volume do I drop from the top of the range to the one that’s good enough”. Both vendors have just handed you the numbers to settle it yourself — an elegant way of admitting that the benchmark race stops convincing anyone the moment the bill lands.