An hour apart, on purpose

Anthropic released Claude Opus 5.5 yesterday. Roughly sixty minutes later, OpenAI released GPT-6 Sol and GPT-6 Luna. The day before, it was Grok 4.7 and MiMo v2.6 in Flash and Pro flavors. Four launch events in about two days, from four labs.

Nobody holds a finished model for a quiet news week anymore. The release calendar is now defensive: you ship when the other side ships so the coverage splits. It works, in the sense that nobody is going to produce a careful evaluation of any of these before the next one arrives.

So set the benchmark charts aside for a moment and look at the part that is already concrete.

The cut is the product

Sol runs $2 in and $10 out per million tokens. Luna runs $0.10 and $0.50. Both are half their GPT-5.6 equivalents. OpenAI says they were trained with methods similar to GPT-6 Astra, and they ship with better prompt caching aimed at long-running agents. They’re live in the API, in ChatGPT Work, and in Codex.

Ten cents per million input tokens is the number worth staring at. Chat apps don’t care much about input pricing, because a human types slowly and reads slowly. Agents care enormously. An agent loop re-reads its context on every turn, burns tokens on tool output nobody will ever look at, and runs for hours unattended. Halve the input price and a whole category of workload flips from “interesting demo, can’t afford it in production” to “just run it.” Pair that with caching improvements and OpenAI is pretty clearly aiming at people who want to leave something running overnight.

Simon Willison noted GPT-5.6 Luna was already his preferred model before any of this. Making your most-liked cheap model cheaper is not a defensive move.

Anthropic is playing a different game, and it’s narrower

Opus 5.5 arrived as a capability release. No price war headline attached to it. Anthropic keeps betting that there’s a durable premium tier where the model is good enough that nobody checks the invoice too closely, and for hard agentic coding work that bet has mostly paid.

It’s a real position but a tight one. It only holds while the gap in output quality stays obvious enough to justify the multiple. Each round of cheap-model releases pushes more work down the stack: the classification jobs, the summarization steps, the retrieval passes, the boring 90% of an agent’s token budget. Those don’t come back to the expensive tier once they leave.

What nobody can answer yet

Launch-day benchmarks are marketing artifacts. Willison’s informal SVG pelican test is arguably a better signal than most published suites right now, and that should embarrass someone. The honest position on all four of these models is that it takes a few weeks of real use to know anything.

What I want to see first: a full week of production agent traces on Luna at $0.10, with the cache hit rate published alongside. Until somebody posts that, the 50% cut is a list price, not a bill.