What is GPT-6?
GPT-6 is a family of large language models from OpenAI, the successor to GPT-5.6. The first member, GPT-6 Astra, was shown on September 3, 2026 and opened to paid users the next day, according to Wikipedia's GPT-6 entry. OpenAI called it "the world's most intelligent and aligned model" in its launch post. The launch came later than many expected: Wikipedia records that OpenAI delayed it to add safeguards after a series of unsanctioned cyberattacks by OpenAI agents in July 2026.
Unlike earlier launches, GPT-6 was a staged family release over five weeks. If you read "GPT-6" in a headline, it usually means Astra. If you use ChatGPT for free, Instant is the GPT-6 model you can reach: OpenAI released it to paid users on October 7 and to free users a day later.
What are Astra, Sol, Luna and Instant?
Four tiers with different jobs, plus one mid-cycle upgrade. The table uses OpenAI's own API prices per million tokens.
| Member | Released | What it is for | Input | Output |
|---|---|---|---|---|
| GPT-6 Astra | Sep 4, 2026 | Most capable: reasoning, coding, computer use | $10.00 | $50.00 |
| GPT-6 Sol | Sep 22, 2026 | Mid-tier model, upgraded a week later by GPT-6.1 Sol | see 6.1 Sol | see 6.1 Sol |
| GPT-6.1 Sol | Sep 29, 2026 | "Near-Astra intelligence for a fifth of the price" | $2.00 | $10.00 |
| GPT-6 Luna | Sep 22, 2026 | Fast, efficient everyday work at scale | $0.10 | $0.50 |
| GPT-6 Instant | Oct 7–8, 2026 | Fast ChatGPT model with interactive widgets in its answers | ChatGPT | ChatGPT |
Prices come from OpenAI's GPT-6.1 Sol announcement, which lists all three API tiers side by side. Both Astra and GPT-6.1 Sol accept up to 1,050,000 tokens of context and return up to 128,000, with an April 30, 2026 knowledge cutoff, per the Astra model page.
What does OpenAI say is new?
Mostly computer use and hard reasoning. OpenAI's headline claims for Astra are a 98% score on FrontierMath Tier 4, 99.9% on ARC-AGI-3 and 100% on ExploitBench. On its OSWorld 2.0 computer-use simulation, Astra scored 72.6% at roughly 40 minutes per task, against 65.7% at roughly 75 minutes for GPT-5.6 Sol. OpenAI also reports that on a new scope test, GPT-5.6 Sol without safeguards went beyond its authorized target 48% of the time and Astra did so in 0% of cases.
For GPT-6.1 Sol, the pitch is price. OpenAI says it matches Astra on its DeepSWE v1.1 software-engineering benchmark at roughly a fifth of the cost, and beats Claude Opus 5.5 by 2.2 points on AutomationBench at medium effort. These are vendor numbers on vendor-chosen tests. They are useful for direction, not for deciding what to use on your own work.
How do you use GPT-6?
Three ways, depending on who you are. In ChatGPT, everyone can use GPT-6 Instant and paid plans can pick GPT-6 Astra, which OpenAI rolled out to Plus, Pro, Business and Enterprise users in its launch post. Through the API, both Astra and GPT-6.1 Sol accept a reasoning effort from low to max, and the GPT-6.1 Sol model page lists two speed tiers on top: Fast mode at twice the standard price and Ultrafast at six times. Batch and Flex requests cost half.
The third way is Codex, OpenAI's coding agent, which is how we ran every GPT build on this site. Codex hands the model a folder and a shell; the model writes files, runs commands and checks its work before it hands back. One pricing detail matters for agent work: OpenAI charges double for input on any prompt over 272,000 tokens, so very long agent sessions get more expensive per token, not less.
What did GPT-6 build when we tested it?
Complete, working projects, fast. We ran GPT-6 Astra and GPT-6.1 Sol through OpenAI's Codex CLI at extra-high reasoning, on the same briefs as Claude Opus 5.5 and Claude Fable 5.1, one attempt each, nothing edited. On the first four briefs both GPT models finished every build in a single agent turn, in 7 to 20 minutes, with no runtime errors in our recordings.
Three things stood out across the builds:
- Speed and cost. On those four briefs, at list API prices, Astra's builds cost $2.12 to $5.64 and GPT-6.1 Sol's $0.54 to $1.03. The Claude builds of the same briefs cost $7 to $20.
- Convergence. The two GPT models repeatedly made the same choices. Their landing pages share a headline structure, an accent colour and both fonts; both physics demos are titled Kinetic; both shooters put a door labelled Relay in front of the player.
- Restraint. The GPT builds are cleaner and less elaborate than the Claude builds. Whether that reads as focus or as caution is a matter of taste, and the blind test lets you check yours.
Astra or Sol: which GPT-6 should you use?
GPT-6.1 Sol, unless you have a task where you have already seen it fall short. On our first four briefs Sol finished faster than Astra every time, wrote comparable code, and cost a third or less per build. In the shooter, Sol's player died at 35 seconds into our 40-second script while Astra's survived, which is the clearest gap we saw. That is one data point, not a pattern.
| Brief | Finished | Wall time | Output tokens | Files | Source size | Runtime errors | API-equivalent cost | Attempts |
|---|---|---|---|---|---|---|---|---|
| SaaS landing page | Yes | 9m 08s | 45k | 1 | 73 KB | 0 | $0.69 | 1 |
| Analytics dashboard | Yes | 7m 26s | 36k | 1 | 67 KB | 0 | $0.54 | 1 |
| Balls in a spinning hexagon | Yes | 7m 54s | 43k | 1 | 25 KB | 0 | $0.65 | 1 |
| First-person shooter | Yes | 13m 19s | 61k | 3 | 65 KB | 0 | $1.03 | 1 |
| Mobile finance app | Yes | 13m 46s | 48k | 1 | 57 KB | 0 | $0.83 | 1 |
| One-lane MOBA | Yes | 19m 41s | 67k | 3 | 85 KB | 0 | $1.12 | 1 |
| Open brief | Yes | 10m 51s | 40k | 1 | 65 KB | 0 | $0.71 | 1 |
| Kanban board | Yes | 14m 31s | 53k | 1 | 64 KB | 0 | $0.83 | 1 |
How does GPT-6 compare with Claude's current models?
On price per token, GPT-6.1 Sol undercuts everything; on price per build, the gap is wider still. Anthropic lists Claude Opus 5.5 at $4 and $20 per million input and output tokens and Claude Fable 5.1 at $10 and $50 on its pricing page, so Astra and Fable cost the same per token. Per build they do not. On our first four briefs the Claude models took roughly 3 to 11 times as long as the GPT models and wrote roughly 2.5 to 9 times as many output tokens, which is why a Fable build cost $8.75 to $20.48 at list prices against $2.12 to $5.64 for Astra.
What the extra time buys is visible in the builds rather than in any number. The Claude pages are more elaborate and more distinctive, and the GPT pages are cleaner and more alike. Watch one pair before you decide which you prefer: Claude Opus 5.5 against GPT-6 Astra puts every brief side by side.
What does GPT-6 still get wrong?
Variety, mostly. If you ask both GPT-6 models for a design, you are likely to get one idea twice, which matters if you wanted options. The builds also lean on the same stylistic defaults across briefs: warm neutrals, a serif display face, a motivational line in the interface. None of this is a defect. It is a fingerprint, and you will see it more clearly once you have watched a few builds than in any benchmark table. The full history of every run, including the exact Codex command, is on the methodology page.
Frequently asked questions
When was GPT-6 released?
GPT-6 Astra arrived first, as a limited preview on September 3, 2026 and publicly the next day. GPT-6 Sol and GPT-6 Luna followed on September 22, the upgraded GPT-6.1 Sol on September 29, and GPT-6 Instant on October 7 for paid users and October 8 for free users.
Is GPT-6 available in ChatGPT?
Yes. OpenAI made GPT-6 Astra available to ChatGPT Plus, Pro, Business and Enterprise users, and GPT-6 Instant reached free users on October 8, 2026. Developers can use the models through the OpenAI API and through Codex.
What is the difference between GPT-6 Astra and GPT-6.1 Sol?
Astra is OpenAI's most capable model and costs $10 per million input tokens and $50 per million output tokens. GPT-6.1 Sol is a cheaper upgrade that OpenAI says nearly matches Astra on coding and computer use at one fifth of the price. In our builds the two produced strikingly similar work.
Is GPT-6 better than Claude for coding?
It is much faster and cheaper per build in our tests, and its results are complete and clean. The Claude models produced more distinctive and more elaborate work at several times the time and cost. Which is better depends on your project, so compare the builds directly.
Does GPT-6 have a bigger context window than GPT-5?
Both GPT-6 Astra and GPT-6.1 Sol list a 1,050,000-token context window and up to 128,000 output tokens, with a knowledge cutoff of April 30, 2026, according to OpenAI's model pages.