Anthropic and OpenAI both released new frontier models on Tuesday. Claude Opus 5.5 and GPT-6 Sol and Luna are each cheaper than the models they replace, and they landed ten days after Anthropic’s chief executive published an essay arguing the industry should slow the rate at which it improves model capabilities.
What shipped
Opus 5.5 is $4 per million input tokens and $20 per million output, down from $5 and $25. Cache reads fall from $0.50 to $0.20, and cache reads are the line item that dominates agentic and coding bills. Anthropic says the model also uses fewer tokens per task, and puts the combined effect at 40% less to run on typical workloads. Output is more than 30% faster. Fast mode costs $8 and $40. Cybersecurity tasks get re-routed to Opus 4.8 when the safeguards intervene.
OpenAI cut prices one tier down. GPT-6 Sol goes from $4 to $2 in and $20 to $10 out. Luna, the small one, goes from $0.20 to $0.10 and $1.20 to $0.50, both 50% below the promotional GPT-5.6 rates they replace. GPT-6 Astra stays the top tier for the hardest work. OpenAI attributes the cut to caching and inference efficiency and says the new prices are permanent rather than introductory.
The call to pace, and what followed it
“We Must Pace the Frontier”, published on 12 September, proposes three steps: frontier companies give embedded third-party evaluators employee-like access to their pipelines, democratic countries coordinate on common safety standards, and governments eventually coordinate across blocs. Only the first is a unilateral commitment, and it is the one that shows up in this release. Anthropic describes Opus 5.5 as “our first release since we called for pacing the frontier” and says Frontier Design and METR evaluated it before launch. The essay is explicit that pacing does not mean halting training, and apart from the evaluators it sets no dates.
The cadence
The Register counted the releases, and the arithmetic is blunt. Anthropic’s cadence moved from roughly quarterly in 2025 to almost monthly this year: Fable 5.1 and Mythos 5.1 arrived 21 days before Opus 5.5, which came 39 days after Opus 5. At OpenAI, GPT-6 Astra landed on 3 September and Sol and Luna followed 19 days later. Both companies are preparing for initial public offerings.
What it means for the machine at home
Every release in this run is a price cut at the hosted frontier, which makes the cost argument for a local model weaker each time one lands. The arguments that are not about cost, that the data stays here and the thing keeps working when a vendor changes its mind, are untouched by any of it. One practical detail worth carrying: on agentic work, cache reads dominate the bill, which is why the 60% cut to that line matters more than the 20% headline on tokens.
My read
The race did not slow this month. The labs’ own release notes are the evidence, and the only part of the pacing plan with weight today is the smallest part: one company put outside evaluators in front of a launch and named them.
If you are deciding whether to buy a GPU for local inference, the hosted numbers keep moving the wrong way for that comparison. Buy the box because you want the model to keep working on your terms, and expect the per-token price of renting frontier intelligence to keep falling while you do it. Cost is the argument that will keep moving against you.
Sources: Anthropic’s Opus 5.5 announcement, OpenAI’s GPT-6 Sol and Luna announcement, Amodei’s “We Must Pace the Frontier”, and The Register’s cadence counts