OpenAI answered Anthropic’s Tuesday card within the hour. GPT-6 Sol and GPT-6 Luna shipped today under flagship GPT-6 Astra, priced at half the promotional API rates of their GPT-5.6 counterparts — $2 and $10 per million tokens in and out for Sol, $0.10 and $0.50 for Luna — and the launch card marks the cut permanent rather than promotional, with cache reads keeping their 90 percent discount. The two closed labs’ first releases since they publicly called for pacing the frontier now bracket the same morning, both as discounts, both priced against the recurring bill of a running agent rather than the headline.
Half the price, and the fight moved down the stack
What happened. OpenAI expanded GPT-6 on Tuesday with two models below the $10/$50 flagship Astra: GPT-6 Sol at $2/$10 per million tokens and GPT-6 Luna at $0.10/$0.50, a permanent 50 percent API price reduction versus GPT-5.6 promotional pricing, confirmed by Reuters and ZDNET. Artificial Analysis’ first full evaluation puts Sol a modest notch ahead of GPT-5.6 Sol on its intelligence index while cutting max-effort per-task cost roughly in half ($1.06 versus $1.99); Luna runs the same suite for about seven cents. The models are live now in ChatGPT, Codex, and the API, on top of the Opus 5.5 note that fired an hour earlier.
Why it matters. The capability headline barely moved — an independent bench reads roughly-even intelligence at dramatically lower unit cost. The move is the price itself: OpenAI did not wait for DevDay to answer Anthropic, and both flagships repriced the agent-bill unit on the same Tuesday the pace debate peaked. Luna at $0.10/$0.50 makes long-running, high-volume agent loops cheap enough to leave on overnight, which converts “agents cost less to run than to write” from a flagship argument into a floor argument. If Anthropic’s Sonnet 5.5 and Haiku 5.5 tiers (due in coming weeks, per the Opus 5.5 card) undercut these lines, the closed lane stops competing model-versus-model and starts competing on who owns the low end of the price list.
What I’m watching
Whether the permanent-50% language holds, and whether Sonnet and Haiku 5.5 answer it before DevDay on September 29. One take: a lab that asks the field to slow the frontier and then halves tier pricing the same hour its rival does isn’t pacing — it’s pricing by attention.
Source: openai.com, zdnet.com, officechai.com, 9to5mac.com