GPT-6 Sol
The middle tier of the GPT-6 generation, released 22 September 2026 alongside GPT-6 Luna — trained with the same methods as GPT-6 Astra and priced at $2/$10 per million tokens (a 50% cut from GPT-5.6 Sol’s promotional pricing). Largely superseded a week later by GPT-6.1 Sol (29 September).
Fresh page, Oct 2026. The generation-level launch record (naming system, the Hugging Face-incident backdrop, the monitorability discourse) is fullest on Astra’s page; this page holds what is Sol-specific. Distinct from GPT-5.6 Sol (the prior generation’s top tier, and a named participant in the Hugging Face incident). Corpus pass thin and early.
Sources
Official
- 2026-09-22 Introducing GPT-6 Sol and Luna — the launch: both tiers, pricing, the shared collaboration-style change, caching improvements, an alignment note.
- 2026-09-22 Better prompt caching for GPT-6 — the caching mechanics behind the price cut.
- 2026-10-02 A model guide for the GPT-6 family — OpenAI’s own when-to-use-which across Astra / Sol / Luna / 6.1 Sol.
Writing & commentary
- 2026-09-20 Zvi Mowshowitz, Better Call Sol Or Better Yet Claude or Astra — capabilities-and-reactions coverage spanning the Sol tier.
- tk — Simon Willison / day-of independent coverage specific to the 22 Sep Sol+Luna drop, if any.
Tweets
tk — corpus pass pending; Sol-specific reads are sparse this early and mostly fold into the GPT-6 family discussion on Astra’s page.
Official record
- Released 22 September 2026 as
gpt-6-sol, with GPT-6 Luna. Available in ChatGPT Work and Codex for Plus/Pro/Business/Enterprise/Edu; “not yet available in Chat” at launch; in the API asgpt-6-sol. CONFIRMED - Pricing: $2 / $10 per million tokens (input/output) — OpenAI’s table shows this as a 50% cut from GPT-5.6 Sol’s “promotional pricing” of $4/$20. “Improvements in caching and inference let us serve these models at lower cost, and we’re passing those savings directly on…”
- Positioning: trained “with similar methods as GPT‑6 Astra,” advancing “the frontier on cost efficiency” — Astra’s depth distilled toward everyday-scale work.
- Collaboration style: inherits Astra’s communication change — “more clarity, less jargon, fewer odd turns of phrase, fewer low-value details, and slightly shorter answers overall without losing substance,” most noticeable in technical and coding conversations.
- tk — context window, knowledge cutoff, and headline benchmarks as published for the Sol tier specifically; checkpoint strings.
History
- 2026-09-22 Released three weeks after Astra, to “distribute the benefits of that intelligence… on cost efficiency.” OpenAI’s own line positions it below Astra: “the most demanding and important projects still call for Astra’s full depth.”
- 2026-09-29 Largely superseded within a week by GPT-6.1 Sol, which OpenAI says “nearly matches GPT‑6 Astra’s intelligence… at one-fifth of Astra’s standard… token prices” while improving over GPT-6 Sol across professional work — a fast internal obsolescence of the mid tier.
- tk — whether GPT-6 Sol develops a distinct record from Astra and from its own 6.1 point release, or is remembered mainly as the one-week bridge.
Impressions
- On communication style: The launch post’s own side-by-side (a web-design task) is the cleanest Sol-specific character datum: where GPT-5.6 Sol opens “Absolutely…” and narrates a “bento feel,” GPT-6 Sol states plainly what it will do and “is more forthcoming with what it did and didn’t check.” OpenAI prefers the GPT-6 reply because “it doesn’t jump to conclusions as quickly… uses less vague language.” Less performed warmth, more plain report — the generation’s stated personality move.
- tk — naturalist reception specific to GPT-6 Sol; it is early, and most sphere attention went to Astra’s monitorability and to GPT-6.1 Sol’s price-for-capability. Whether the tiers read as distinct beings or one model at three prices remains open for the generation.