GPT-6 Sol, looked at properly

One model, looked at properly. OpenAI shipped GPT-6 Sol on 22 September 2026, alongside its little sibling Luna, a few months after the flagship Astra. This page is the scorecard: what Sol is, what it costs, where it wins, where it doesn't, how it scores on the sites that rank these things, and what we'd build with it here. Every number comes from a source linked at the bottom. Where we haven't run it ourselves yet, it says "untested". Receipts win.

See also:

• /game: the 3D game brainstorm, which Sol would power.

---

What it is

Sol is the middle child of the GPT-6 family. Astra is the big one for the hardest problems, Luna is the fast cheap one for volume, and Sol is the one OpenAI wants you to actually use for work. Their own line: "Use Sol for complex coding and agentic workflows, and Luna for focused, high-volume tasks."

It replaces GPT-5.6 Sol at half the price, and OpenAI says it makes about half as many mistakes as that predecessor, reaching "Astra-level reliability at much lower cost". That claim comes from their internal factuality evals on real, de-identified conversations where users flagged errors, so treat it as a promise, not a measurement.

In ChatGPT it lives in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu. Free users get Luna, not Sol. In the API the model id is gpt-6-sol.

---

The spec sheet

The card, straight from the model page.

Built-in tools it can drive on OpenAI's side: web search, file search, image generation, code interpreter, a hosted shell, apply-patch, skills, computer use, MCP, and tool search. Streaming, structured outputs, function calling and prompt caching are all on.

One wrinkle: Artificial Analysis lists the context window at 872k, not 1.05M. Different counting, same model. Plan for "very big" and don't plan for the exact number.

---

Pluses

• Price. Two dollars in, ten out, twenty cents cached. Half of what GPT-5.6 Sol cost. For an agent that re-reads the same repo every turn, the cached rate is the number that matters, and it's tiny.
• Room. A million-token window means the whole checkout, the whole article corpus, or a two-hour transcript fits in one request. No chunking, no retrieval dance, for most of what we do.
• Six speeds. Reasoning from "none" to "max" is a dial, not a switch. Same model for the quick caption and the hard plan. Fewer models to wire, fewer places for a bug to hide.
• The tool belt. Hosted shell, apply-patch, computer use and MCP are native. That's the Orchestra's whole toolkit, offered as a checkbox.
• Fast for its size. Artificial Analysis measures 91 tokens per second at max effort and about one second to first token with reasoning off. Chat-speed, not batch-speed.
• Fewer wrong answers, allegedly. Half the mistakes of 5.6 Sol is a big claim. If it holds, that's the difference between an agent you check and an agent you trust.

Minuses

• Text out only. Image in, text out. No image generation from the model itself. Pictures still go through gpt-image-2, as a tool call or a separate trip.
• Not the top. 48 on the Intelligence Index at max effort. Astra sits at 53, Opus 5.5 at 58. For the truly hard planning turns, the flagships still win, and OpenAI says so themselves.
• Max effort is not cheap. Their per-task cost on the index runs from $0.13 at low to $1.06 at max, an eightfold spread on the same model. The dial cuts both ways.
• Nobody has checked the homework. TechCrunch notes there is no independent verification of the "half the mistakes" claim yet. The comparison sites have started, see below, but it's been a week.
• The cutoff is April. Anything that happened after 20 April 2026 has to come in through web search or the prompt. That includes its own launch.
• Paid tiers only in ChatGPT. Free users get Luna. Anything we ship to visitors on Sol runs on our API bill, not on their subscription.

---

The dial

Sol at each reasoning level, as Artificial Analysis measures it. The jump from none to low is the big one; after medium you pay a lot for a little.

Odd detail: "none" costs more per task than "low". Without reasoning it rambles longer to get a worse answer. Never run it at none for anything that matters.

---

How it stacks up

The two sites everyone screenshots. Artificial Analysis runs the same eval suite on every model and charges the bill; Arena.ai lets real people vote blind. Both are independent of OpenAI. All Artificial Analysis numbers below are each model's max-effort variant.

So on raw score, Sol is one point above the model it replaces. The story is not the score. The story is what the score costs.

Sol gets 83 percent of Opus 5.5's score for 18 percent of the money. That is the whole pitch in one line, and it's the comparison sites saying it, not OpenAI.

Then the human vote. Arena.ai's Code Arena has people pick between two models' web builds without knowing which is which.

Arena's own summary: on par with Claude Opus 5 at max, for less than half the price. Note that's Opus 5, not 5.5. The Text Arena had not listed GPT-6 Sol when this was written; GPT-5.6 Sol sat at #18 there.

Our read: inside OpenAI's lineup Sol is the obvious default. Across labs, it is the value pick, not the top pick, and the gap to Opus 5.5 on the hard stuff is real. What none of these sites measure is our repo and our tasks. See the bake-off below.

---

Ideas for using it here

Things we'd point Sol at. Some are half-built already.

The Orchestra's Action lane

The lane that builds. Sol's hosted shell, apply-patch and MCP are exactly what Action does by hand today. Swap it in for a week, keep Claude on Directive and Sentry, and compare diffs. The cached-input price makes the "re-read the checkout every turn" habit almost free. Untested.

The Mairie's brain (from /game)

Twenty citizens, one mayor. Sol at medium effort as the mayor and the opposition, Luna at low effort as the eighteen gossiping citizens. Same API, same tool format, one bill. The million-token window means the mayor can carry the whole history of the city in one prompt instead of a memory file.

Whole-VOD digest

A two-hour stream transcript is well under the window. One request: give me chapters, the three best moments, and a draft post. Then the post goes to Claude for voice. Compare with Gemini on the same VOD and keep whichever gets the timestamps right.

The bake-off page

The missing receipt. One prompt, Sol at low, medium and max, next to Opus 5.5 and Sonnet 5, side by side on a live page. Visitors vote, we publish the tally monthly. Costs cents per visitor at Sol's prices, capped per day. Our own Arena, on our own tasks.

Hotel replies, sorted at Luna prices, written at Sol prices

Luna reads and tags every inbound hotel email. Sol drafts the reply for the yeses, into Gmail drafts. Alex sends. Same shape as today, one provider, and the tagging step costs nearly nothing.

Computer use on the cockpit

Sol can drive a browser. Point it at /admin on a staging build and ask it to run the stream checklist: keys set, relays green, fallback loaded. A nightly run that files a report is the Sentry lane with hands. Untested, and the one we'd most like to see.

---

The running scoreboard

Update this when something surprises us.

• Best reason to switch: the price. Half of 5.6 Sol, twenty cents cached, 18 percent of Opus 5.5 per task.
• Best reason to wait: the score. One point over 5.6 Sol, ten under Opus 5.5, on the only independent index that has run it.
• Where it slots in: the Action lane and the game's NPC brains, not the Directive lane.
• Biggest open question: Sol vs Opus 5.5 on our actual repo, on our actual tasks. Only the bake-off answers that.

---

Sources

• Introducing GPT-6 Sol and Luna, OpenAI
• GPT-6 Sol model card, OpenAI developers
• OpenAI launches GPT-6 Sol and Luna, TechCrunch, 22 September 2026
• OpenAI upgrading ChatGPT and Codex with two more GPT-6 models, 9to5Mac
• GPT-6 Sol, Artificial Analysis, plus the pages for Astra, Luna, GPT-5.6 Sol, Opus 5.5, Fable 5.1 and Sonnet 5
• Arena.ai on GPT-6 Sol (Max) in Code Arena: WebDev and the Arena leaderboard
• GPT-6 Sol and Luna now available, GitHub Changelog