GPT-6 Sol and Luna Are Out: Prices, Specs, Rumors Graded

§ 01What OpenAI released#

Two lines from the post set the frame. “GPT‑6 Astra continues to be our best model across the board.” And: “These models are not yet available in Chat.” Sol and Luna are the working models under Astra, reached through ChatGPT Work, Codex and the API first.

§ 02Pricing: half, on both#

OpenAI’s own summary is “reducing API prices for Sol and Luna by 50% compared with their GPT‑5.6 promotional pricing”. The pricing page confirms it row by row.

The tracker’s open question was where a GPT-6 Sol would land on the ladder. The answer is the slot below GPT-5.6 Sol at half its price, which makes the September 6 reading, the everyday GPT-6 for Plus users, the right one. Alongside the price, OpenAI says GPT-6 caching now yields higher hit rates by default, that changing reasoning effort or tool availability no longer breaks the cache, and that explicit breakpoints let developers choose where a cached prefix ends.

§ 03One afternoon, two price cuts#

OpenAI’s launch table compares Sol with Claude Opus 5 and Claude Fable 5.1, the models that existed when it was built; the Opus 5.5 that shipped 89 minutes earlier is not in it, and Anthropic’s table does not carry GPT-6 Sol either. Our Opus 5.5 record has Anthropic’s side of the afternoon.

§ 04The launch table, as OpenAI printed it#

The other numbers OpenAI printed, each with its own footing: on Agents’ Last Exam V1, Sol at max effort scores 56.4 percent, above Claude Opus 5’s best at 60 percent lower cost per task; on DeepSWE v1.1, Sol at max scores 68.8 percent, “within 1.1 percentage points of Claude Fable 5’s highest score in the evaluation”, and Luna at max scores 66.6 percent at 93 percent lower cost than Opus 5; on OSWorld 2.0 offline, Sol at xhigh effort “achieves a similar score to Claude Opus 5 at medium effort”, 60.5 percent versus 60.3, “at approximately 80% lower cost per task”. On OpenAI’s internal factuality evaluation, “GPT‑6 Sol makes about half as many mistakes as its predecessor, approaching Astra-level reliability at much lower cost.”

Two caveats are OpenAI’s own. Competitor scores are from public reports, and “Scores for Claude Fable 5 were reported when scores for Claude Fable 5.1 were unavailable.” And the Fable 5.1 datapoint on AutomationBench omits the Opus 5 fallbacks that occurred on about 40 percent of tasks, so its cost is understated. There is no independent score for either model yet; the leaderboards are the watch.

§ 05Every September claim, graded#

Four confirmed, two missed, one contradicted, one uncheckable, one held, and one thing nobody said: the accounts described one model and OpenAI shipped two. The date-bearing claims were the weakest, as the tracker expected, and the qualitative claim that turned out right was the modest one.

§ 06The rhythm, closed#

Nineteen days is the longest gap in the table and still inside the shape the tracker described: Astra was the first OpenAI flagship in a year to ship alone, and a sibling followed. No Terra this time; the GPT-6 family is Astra, Sol and Luna as of today.

§ 07What the market said#

Polymarket’s next-GPT markets moved on the posts above; a market is a read on belief, and this one resolved on OpenAI’s schedule rather than any account’s date.

§ 08Where CellCog stands#

Our conflict, declared: we build CellCog, a platform where a business hires AI employees that keep memory and work as a team, and we choose the models underneath them. As of September 22, 2026 our Agent, Creative and Team tiers run on Claude Opus 5.5, moved there the day it shipped (our day-one product update), and Flash on Gemini 3.8 Flash; the one OpenAI seat in our stack is the web-search engine inside Agent Team Max, which moved to GPT-5.6 Sol on August 14. GPT-6 Sol at half that price, with a two-month-newer knowledge cutoff, is the obvious candidate for that seat, tested on citation accuracy across long research runs before price. If it moves, the receipt lands here and in a product update the same day.

§ 09What we are watching for#

• Independent scores for Sol and Luna: Artificial Analysis, the Terminal-Bench and OSWorld leaderboards, Arena.
• The GPT-5.6 Sol promotional price, advertised through at least November 21, now that its successor costs half.
• Whether Sol and Luna reach Chat, and what DevDay on September 29 adds.
• Our own search-seat decision, and its receipt.

§ 10Update log#

This is a living page; when the story moves, the update lands here.

September 29, 2026: GPT-6.1 Sol replaced GPT-6 Sol seven days after it shipped. OpenAI’s API changelog released gpt-6.1-sol “for complex coding and professional work at a lower cost than GPT-6 Astra,” and the gpt-6-sol model page now points readers to “the newer Sol model.” Prices hold at $2 per million input tokens and $10 output; cached input falls to 5 percent of the input rate (from 10 percent), cache writes cost $2.50, the knowledge cutoff moves to April 30, 2026, and prompts over 272K input tokens bill at 2x input and 1.5x output for the whole request. Artificial Analysis puts it 4 points above GPT-6 Sol on its Intelligence Index and 1 point below GPT-6 Astra, at $0.72 per index task against $1.05 for GPT-6 Sol and $3.26 for Astra. There is no 6.1 Luna: OpenRouter lists gpt-6.1-sol and gpt-6.1-sol-pro only, read October 1, 2026.

October 11, 2026: OpenAI raised GPT-6.1 Sol’s default context in Codex and ChatGPT Work. Tibo, who leads product on Codex, posted at 23:45 UTC: “We have increased default context length for GPT-6.1 Sol to 500k when used through Codex & Work, up from 272k”. He said it reduces compactions, that users can still override it, and that the model’s maximum stays 1M. The post passed 1 million views within a day. The API prices above do not change: prompts over 272K input tokens still bill at 2x input and 1.5x output.

September 22, 2026, 21:00 UTC: OpenAI published Better prompt caching for GPT-6. Its prompt caching guide now describes caching as enabled by default, cached input “discounted up to 90%”, a 30-minute cache lifetime that restarts every time a prefix is reused, explicit cache breakpoints with up to four cache writes per request, and a Prompt Caching Dashboard. For long agent runs the lifetime matters as much as the discount: a prefix left idle for more than 30 minutes has to be written again.

As of September 22, 2026, 19:10 UTC: released. OpenAI’s launch post (feed time 18:00 UTC), its X posts (18:12 UTC), the API changelog’s September 22 entry, the gpt-6-sol and gpt-6-luna model pages, the pricing page and OpenRouter’s catalog all read within an hour of the post. The tracker became this record; every September claim graded above.

September 18, 2026, 07:15 UTC: the rumored Thursday passed. September 17 closed in Pacific time at 07:00 UTC with no gpt-6-sol anywhere OpenAI publishes; the September 10 account’s best guess graded a miss.

September 16, 2026, 01:30 UTC: the ChatGPT Learn changelog (entry dated September 14) retired GPT-5.5 from ChatGPT, ChatGPT Work and Codex on October 14 and told Codex users to switch to gpt-5.6-sol; no document named a gpt-6-sol.

September 15, 2026, 19:20 UTC: Sam Altman posted “big 🚢 this week” and “and then for devday”; no model named, logged as a hint.

As of September 13, 2026: page opened. Latest rumored windows were the week of September 14 and DevDay, September 29; no official movement to log.

§ 11Sources#

• OpenAI, Introducing GPT-6 Sol and Luna, September 22, 2026 (news feed time 18:00 UTC), and the @OpenAI post, 18:12 UTC, read directly.
• OpenAI, GPT-6 Sol model page, GPT-6 Luna model page, GPT-5.6 Sol model page, pricing, API changelog, all read September 22, 2026, 19:05 UTC; OpenRouter models API, same time.
• Anthropic, Introducing Claude Opus 5.5, September 22, 2026, and the @AnthropicAI post, 16:31 UTC.
• X posts of September 6 to 15, read directly at the time: @pankajkumar_dev September 6; @AILeaksAndNews September 10; @Mr_Salio September 13; Sam Altman September 15.
• OpenAI, GPT-6 Astra, September 3, 2026; GPT-5.4 mini and gpt-5-mini model pages.
• CellCog routing statements reflect our configuration on September 22, 2026.