GPT-6 is OpenAI’s September 2026 model generation — and unlike GPT-5.x, it arrived as a staggered family. GPT-6 Astra (September 3) is the flagship; GPT-6 Sol and GPT-6 Luna (September 22) bring similar training to cheaper, faster tiers; GPT-6.1 Sol (September 29) upgrades Sol to near-Astra intelligence at one-fifth of Astra’s price; and on October 7 GPT-6 reached the ChatGPT Chat tab with a new Intelligent UI.
This page covers the whole family. It succeeds GPT-5.6 and competes head-on with Claude Opus 5.5 and Claude Fable 5.1.
The GPT-6 lineup
| Model | Released | API ID | Input / Output (per 1M) | Cached input | Role |
|---|---|---|---|---|---|
| GPT-6 Astra | Sep 3, 2026 | gpt-6-astra | $10 / $50 | $1.00 | Best results, no compromise |
| GPT-6.1 Sol | Sep 29, 2026 | gpt-6.1-sol | $2 / $10 | $0.10 | Near-Astra at a fifth of the price |
| GPT-6 Sol | Sep 22, 2026 | gpt-6-sol | $2 / $10 | $0.20 | Previous Sol (superseded by 6.1) |
| GPT-6 Luna | Sep 22, 2026 | gpt-6-luna | $0.10 / $0.50 | $0.01 | Fast, efficient everyday work at scale |
All three tiers share a 1.05M-token context window and 128K max output, support text and image input with text output, and accept reasoning.effort of low, medium, high, xhigh and max (Luna also offers none). Knowledge cutoff is April 30, 2026 (Astra, Sol) or May 18, 2026 (Luna).
GPT-6 Astra (September 3)
OpenAI’s launch claim: “the most intelligent and aligned model in the world”, leading in computer use, web browsing, software engineering, cybersecurity, science and professional work.
| Benchmark (OpenAI-reported) | GPT-6 Astra | Comparison cited by OpenAI |
|---|---|---|
| ARC-AGI-3 | 99.9% | Beat the human-action-efficiency baseline in 96% of levels (ARC Prize Foundation) |
| FrontierMath Tier 4 (v2) | 98% | Ceiling of that tier |
| Terminal-Bench 4.0 | 57.9% | Fable 5.1 55.8%, GPT-5.6 Sol 37.3% |
| Terminal-Bench Science 0.1 | 64.6% | Fable 5.1 52.6%, GPT-5.6 Sol 22.4% |
| Agents’ Last Exam | 59.3% | Opus 5 55.5%, GPT-5.6 Sol 53.6% |
| OSWorld 2.0 (offline, partial) | 72.6% | GPT-5.6 Sol 65.7%, with ~47% less time per task |
| GPQA Diamond | 96.0% | GPT-5.6 Sol best 94.9% |
| ExploitBench | 100% | GPT-5.6 Sol 78.5% (without production safeguards) |
Astra also changes how Codex handles long sessions: instead of repeatedly compacting context into a summary, it can keep notes across context windows and search earlier windows (experimental, to become the default within weeks). On alignment, an eval built from a Hugging Face incident found GPT-5.6 Sol went beyond its authorized goal in 48% of cases, versus 0% for Astra.
Cybersecurity. Astra reached the critical threshold in OpenAI’s Preparedness Framework. It will refuse advanced tasks such as building proof-of-concept exploits, with a less restrictive tier planned through OpenAI Daybreak. Extra safety checks can pause or stop some legitimate work, including defensive security.
GPT-6 Sol and Luna (September 22)
Sol and Luna carry Astra’s methods down-market, and OpenAI cut API prices 50% versus the GPT-5.6 promotional rates ($4 / $20 → $2 / $10 for Sol; $0.20 / $1.20 → $0.10 / $0.50 for Luna). Prompt caching improved too, with a 90% discount on cached reads and higher default hit rates.
- On AutomationBench, GPT-6 Sol at xhigh scored 33.2% at $0.27 per task, ahead of Claude Opus 5 at max (26.9%) at roughly 9% of Opus 5’s cost per task.
- On DeepSWE v1.1, Sol at max scored 68.8%, within 1.1 points of Claude Fable 5’s best (69.9%) at about 80% lower cost; Luna at max scored 66.6%.
- On OSWorld 2.0 offline, Sol at xhigh reached 60.5% vs Opus 5 at medium’s 60.3%, at roughly 80% lower cost.
- Availability: ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu; Luna in the desktop app for Free and Go users.
GPT-6.1 Sol (September 29)
“Near-Astra intelligence for a fifth of the price.” OpenAI reports 6.1 Sol:
- matches Astra on DeepSWE v1.1 at about one-fifth of the cost, and beats GPT-6 Sol’s best score by 6.4 points at lower effort;
- scores 2.2 points above Opus 5.5 (medium effort) on AutomationBench at roughly a third of the cost, and 4.8 points above GPT-6 Sol;
- gains 7 points on OSWorld 2.0 offline over GPT-6 Sol at max effort, coming within 2.1 points of Astra at about one-seventh of the cost;
- more than doubles GPT-6 Sol on Terminal-Bench Science 0.1 at $5.47 per task (vs $23.21 for Opus 5.5 and $23.80 for Astra), though Astra still leads at 68.1%;
- cuts the factual-error share at low effort from 11.4% to 7.7%.
A faster GPT-6.1 Sol Ultrafast (up to 8x token generation speed in Codex) was announced as coming days later. Standard pricing is $2 / $10, with cached input at $0.10.
GPT-6 in ChatGPT (October 7)
OpenAI rolled GPT-6 into ChatGPT’s Chat tab for the 1.2 billion-plus weekly users, built around Intelligent UI: answers can contain interactive diagrams, buttons, forms, charts and on-the-fly tools such as calculators. Plus, Pro, Business and Enterprise get Sol under the hood and Free and Go get Luna. GPT-6 Instant can start answering while still thinking, cutting time-to-first-token on web-search questions by 44% versus GPT-5.6 Instant on average. Work mode and Codex models were not changed by that release.
How GPT-6 compares
| GPT-6 Astra | GPT-6.1 Sol | Claude Opus 5.5 | Claude Fable 5.1 | |
|---|---|---|---|---|
| Price (in / out, per 1M) | $10 / $50 | $2 / $10 | $4 / $20 | $10 / $50 |
| Context | 1.05M | 1.05M | 1M | 1M |
| Terminal-Bench 4.0 | 57.9% | n/a | 66.4% (Anthropic) | 55.8% |
| Terminal-Bench-Science 0.1 | 64.6% | n/a | 58.7% | 52.6% |
- vs Claude Opus 5.5: the closest rival. Opus is cheaper than Astra and wins on terminal coding and knowledge-work Elo; Astra wins science and business workflows. GPT-6.1 Sol is the closer price match.
- vs Claude Fable 5.1: the same $10 / $50 price; OpenAI’s pages claim Astra leads on most shared benchmarks.
- vs Gemini 4 Argon: Argon launched at $2 / $10 intro pricing but is limited to trusted defenders at the moment; GPT-6.1 Sol is generally available at the same price.
- vs GPT-5.6: GPT-6 Sol and Luna are half the price of GPT-5.6’s promotional rates with materially better scores.
Limitations
- Self-reported benchmarks. Most numbers are OpenAI’s own, often at xhigh/max effort; independent replication will matter, especially for ARC-AGI-3 and the exploit evals.
- Safety friction. Astra’s safeguards may slow or stop legitimate work, particularly in security. The least restricted cyber access runs through a separate program.
- Confusing ladder. Sol exists as both GPT-6 Sol and GPT-6.1 Sol; new projects should target 6.1 Sol.
- Surface differences. Intelligent UI and GPT-6 Instant are Chat-tab features; Work mode and Codex use the earlier launches.
For the wider picture, see the 2026 LLM guide.
Best for
- Astra: the most demanding professional work — computer use, scientific research, large codebases
- GPT-6.1 Sol: agentic coding and business workflows when you want near-Astra quality at a fifth of the price
- Luna: high-volume, cost-sensitive tasks and the default for free ChatGPT users
- GPT-6 in ChatGPT Chat with Intelligent UI (interactive answers, tools generated on the fly)
Pros & cons
Strengths
- Astra posts the top scores on OpenAI's own science and computer-use tests — ARC-AGI-3 99.9%, FrontierMath Tier 4 98%, Terminal-Bench-Science 0.1 64.6%, GPQA Diamond 96.0%
- GPT-6.1 Sol delivers near-Astra results on agentic coding, computer use and professional work at one-fifth of Astra's price ($2 / $10, cached input $0.10)
- A full price ladder from $0.10 / $0.50 (Luna) to $10 / $50 (Astra), all with a 1.05M context window and 128K max output
- Strong alignment story — on a Hugging Face–incident eval, without production safeguards, GPT-6 Astra exceeded its authorized scope in 0% of cases vs 48% for GPT-5.6 Sol (OpenAI-reported)
Limitations
- Astra's cyber capability reached OpenAI's critical Preparedness threshold, so it refuses advanced offensive tasks (such as proof-of-concept exploits) until the Daybreak access program widens
- Astra costs $10 / $50, 2.5x Claude Opus 5.5, which beat it on Terminal-Bench 4.0 and GDPval-AA in Anthropic's comparison
- Prompts above 272K input tokens cost 2x input and 1.5x output on Astra
- Headline numbers are OpenAI-reported, often at xhigh or max effort and with safety checks that can pause or stop legitimate work
Frequently asked questions
What are GPT-6 Astra, Sol and Luna?
Three tiers of OpenAI's GPT-6 family. Astra (Sept 3, 2026) is the flagship, billed as OpenAI's most intelligent and aligned model. Sol and Luna (Sept 22) are cheaper, faster models trained with similar methods. GPT-6.1 Sol (Sept 29) upgrades Sol to near-Astra intelligence. API IDs are `gpt-6-astra`, `gpt-6.1-sol`, `gpt-6-sol` and `gpt-6-luna`.
How much does GPT-6 cost?
Per 1M input / output tokens: Astra $10 / $50, GPT-6.1 Sol and GPT-6 Sol $2 / $10, Luna $0.10 / $0.50. Cached input is $1 for Astra, $0.10 for 6.1 Sol and $0.01 for Luna. Sol and Luna are 50% cheaper than GPT-5.6 Sol and Luna promotional pricing ($4 / $20 and $0.20 / $1.20).
Should I use GPT-6.1 Sol or GPT-6 Astra?
OpenAI says 6.1 Sol matches Astra on DeepSWE v1.1 at roughly a fifth of the cost and comes within 2.1 points of Astra on OSWorld 2.0 at about one-seventh of the cost per task. Astra still leads on the hardest science work (Terminal-Bench Science 0.1: 68.1%) and is recommended when you want the best result without compromise.
Is GPT-6 available in the free version of ChatGPT?
Partly. Since October 7, 2026, ChatGPT's Chat tab uses GPT-6 with Intelligent UI — Sol for Plus, Pro, Business and Enterprise, Luna for Free and Go users, rolling out to free users from the next day. Work mode and Codex models were not changed by that release.
What is the context window of GPT-6?
1.05 million tokens for Astra, Sol and Luna, with up to 128,000 output tokens. Knowledge cutoff is April 30, 2026 for Astra and Sol and May 18, 2026 for Luna.
How does GPT-6 compare with Claude Opus 5.5?
Anthropic's table shows Opus 5.5 ahead on Terminal-Bench 4.0 (66.4% vs 57.9%) and GDPval-AA, and Astra ahead on Terminal-Bench-Science (64.6% vs 58.7%) and AutomationBench (41.4% vs 40.0%). Opus 5.5 costs $4 / $20 against Astra's $10 / $50.