The Frontier Race Is Now a Price Race
GPT-6.1 Sol, Haiku 5.5 and Argon all cut cost per token. Capability is table stakes.
October's model drops read like a pricing sheet. OpenAI's GPT-6.1 Sol launched at $2 in / $10 out per million tokens with near-Astra performance for coding and computer use, plus an Ultrafast variant on October 9. Claude Haiku 5.5 came in about 75% cheaper than its predecessor at $0.10 / $0.50. Gemini 4 Argon's intro price is $2 / $10.
The pattern: labs are competing on cost per call, not just leaderboard score. For anyone running agents that make thousands of calls, that shift matters more than a point of benchmark.
Reflection Beam (501B, open weights) and Mistral Large 4 (1.05T, weights late October) extend the open-weight pressure. When you can self-host a front-tier model, the closed lab's moat is service and safety, not raw ability.
Practical read: pick the model by the job and the bill, not the hype. The cheap routing tier (Haiku, GPT-6 Luna) now handles most sub-tasks the expensive model used to own.
Updated October 28, 2026: the price table below was verified at publication.
Frequently asked questions:
Q: How much do the new models cost?
A: GPT-6.1 Sol launched at $2 in / $10 out per million tokens; Claude Haiku 5.5 at $0.10 / $0.50, about 75 percent cheaper than its predecessor; Gemini 4 Argon at $2 / $10 intro.
Q: What's the pattern?
A: Labs are competing on cost per call, not just leaderboard score - for agents making thousands of calls, that matters more than a benchmark point.
Q: How do open weights fit in?
A: Reflection Beam (501B) and Mistral Large 4 (1.05T) extend the open-weight pressure - when you can self-host a front-tier model, the closed moat is service and safety, not raw ability.