# GPT-6 Sol GPT-6 Sol is the middle tier of [[OpenAI]]'s GPT-6 family: [[GPT-6 Luna]] below it, [[GPT-6 Astra]] above it. It was released on 22 September 2026 together with Luna, about an hour after [[Claude Opus 5.5]]. OpenAI says it was trained with methods similar to Astra's, to bring Astra's gains in professional work, factuality, coding and computer use to a faster, cheaper model. It replaces [[GPT-5.6]] Sol. There is no GPT-6 Terra, so Sol now covers the slot the old middle model used to fill (on [[Hacker News]], the common reading was that Terra never found a niche between Luna and Sol). The real story is the price. Sol costs half of GPT-5.6 Sol's promotional pricing, while independent tests put its overall intelligence roughly level with that predecessor. Think of it as the same brain at half the bill, with a few gains and a few regressions. ## Specifications and pricing | | GPT-6 Sol | |---|---| | Model ID | `gpt-6-sol` | | Input / output | $2.00 / $10.00 per million tokens (GPT-5.6 Sol promo: $4 / $20) | | Cached input / cache write | $0.20 / $2.50 | | Long prompts (> 272K input tokens) | 2× input and cache rates, 1.5× output, for the whole request | | Batch and Flex | 50% of standard rates. Fast mode: 2× | | [[Context Window\|Context window]] | 1,050,000 tokens (922,000 max input) | | Max output | 128,000 tokens | | Knowledge cutoff | 20 April 2026 | | Modalities | Text and image in, text out | | Reasoning effort | `none`, `low`, `medium` (default), `high`, `xhigh`, `max` | Built-in tools through the Responses API: web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, [[Model Context Protocol (MCP)|MCP]] and tool search. Chat Completions only supports function calling with reasoning effort set to `none`. The launch also improved prompt caching for GPT-6: higher cache hit rates by default, 90% discounts on cached reads, explicit breakpoints, and the ability to change reasoning effort or toggle tools mid-conversation without breaking the cache. In ChatGPT, Sol shipped in ChatGPT Work and [[OpenAI Codex|Codex]] for Plus, Pro, Business, Enterprise and Edu users, but not in the regular Chat surface. ## What OpenAI claims All numbers below come from OpenAI's launch post: - **AutomationBench** (business workflows across 47 tools): 33.2% at `xhigh`, $0.27 per task. [[Claude Opus 5]] at max scored 26.9% for 11.1× the cost; Astra at low effort scored 30.3% for 3.9× the cost - **Agents' Last Exam**: 56.4% at max, above Opus 5's best score at 60% lower cost per task - **DeepSWE v1.1** (real codebases): 68.8% at max, 1.1 points behind [[Claude Fable 5]] at `xhigh` (69.9%), for about 80% less per task - **OSWorld 2.0 offline** (computer use): 60.5% at `xhigh`, level with Opus 5 at medium (60.3%), about 80% cheaper per task - **Factuality** (OpenAI's internal eval on conversations where users flagged errors): about half as many mistakes as GPT-5.6 Sol - **Style**: the shorter, plainer communication style of Astra, with less jargon and fewer low-value details ## What independent testing found [[Artificial Analysis]] paints a more sober picture. Intelligence Index and Coding Agent Index scores stay level with GPT-5.6, with progress in some evals and regressions in others: - **Cost per task halves.** $1.06 at max on the Intelligence Index, against $1.99 for GPT-5.6 Sol. The price cut does all the work: Sol uses slightly MORE output tokens per task (31k vs 29k) - **Coding Agent Index: 57**, up 2 points, with Terminal-Bench 4.0 at 43% (vs 37%). At $2.99 per task it sits on the cost/performance Pareto frontier - **Fewer hallucinations, by refusing more.** The hallucination rate on AA-Omniscience drops from 92% to 60% because Sol only attempts 83% of questions (vs 99%). Accuracy drops from 59% to 54% - **Knowledge work regresses.** About 100 Elo points lost on GDPval-AA v2.1, which Artificial Analysis attributes to weaker presentation and deliverables that miss rubric elements [[Simon Willison]] switched his default model in Codex to GPT-6 Sol on launch day. ## Reception and a short life Developers on [[Hacker News]] were split. Some called it a cheaper Astra and used it as an orchestrator. Others said it was markedly worse than GPT-5.6 Sol at coding, and a week later the sentiment in the follow-up thread was mostly negative. One recurring (unverified) theory: GPT-6 Sol was the Terra-class model rebranded, and [[Claude Opus 5.5]] forced OpenAI to ship the "real" one. That real one came fast. On 29 September 2026 (DevDay), OpenAI released **[[GPT-6.1 Sol]]** (`gpt-6.1-sol`): same $2 / $10 price, cached input halved to $0.10, and near-Astra scores (it matches Astra on DeepSWE v1.1 at roughly a fifth of the cost). OpenAI's model docs now point to GPT-6.1 Sol as "the newer Sol model". GPT-6 Sol lasted one week as the current Sol. ## My take GPT-6 Sol was a pricing move wearing a new version number. Same intelligence as GPT-5.6 Sol, half the price, and a habit of saying "I don't know" more often. That last part is actually good for factual work; less good if you grade on deliverables. If you built on it, move to [[GPT-6.1 Sol]]. Same per-token price, better scores, cheaper cache. Just don't assume the swap is free: 6.1 Sol drops the `none` reasoning effort and requires the Responses API for tool calling, so re-run your evals before switching. ## References - OpenAI, "Introducing GPT-6 Sol and Luna" (2026-09-22): https://openai.com/index/introducing-gpt-6-sol-and-luna/ - OpenAI launch post on X: https://x.com/OpenAI/status/2102460975790137662 - OpenAI API model docs: https://developers.openai.com/api/docs/models/gpt-6-sol - OpenAI, "Using GPT-6" guide (migration notes): https://developers.openai.com/api/docs/guides/latest-model - OpenAI, "Introducing GPT-6.1 Sol" (2026-09-29): https://openai.com/index/introducing-gpt-6-1-sol/ - OpenAI API model docs, GPT-6.1 Sol: https://developers.openai.com/api/docs/models/gpt-6.1-sol - Artificial Analysis on X (benchmark results): https://x.com/ArtificialAnlys/status/2102462962758033624 - Simon Willison, "Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war": https://simonwillison.net/2026/Sep/22/opus-and-sol-and-luna/ - Simon Willison, OpenAI DevDay 2026 live blog: https://simonwillison.net/2026/Sep/29/openai-devday-2026-live-blog/ - Hacker News, "GPT-6 Sol and Luna": https://news.ycombinator.com/item?id=49805509 - Hacker News, "GPT 6.1 Sol: Near-Astra intelligence for a fifth of the price": https://news.ycombinator.com/item?id=49896586 ## Related - [[GPT-6.1 Sol]] - [[GPT-6 Luna]] - [[GPT-6 Astra]] - [[2026-09-22 GPT-6 Sol and Luna - same smarts as GPT-5.6, half the price]] - [[GPT-5.6]] - [[OpenAI]] - [[OpenAI Codex]] - [[ChatGPT]] - [[Claude Opus 5.5]] - [[Claude Opus 5]] - [[Claude Fable 5]] - [[Artificial Analysis]] - [[Large Language Models (LLMs)]] - [[Context Window]] - [[AI Frontier Model]] - [[Simon Willison]] - [[Hacker News]]