OpenAI began releasing GPT-6 in September 2026. It is a family of models, not a single system. Each model is built for a different mix of power, speed and price.
This article looks at the numbers behind GPT-6. It covers release dates, benchmark scores, factual accuracy, prices and access. Figures come from OpenAI’s published results unless noted.
Benchmark scores show how a model does on set tests. They do not predict how it will do on every real-world job. Test settings also change the results.
What Is GPT-6?
GPT-6 is the sixth major generation of OpenAI’s GPT models. GPT stands for Generative Pre-trained Transformer. These models learn patterns from large amounts of data and use them to respond to instructions.
Modern GPT models do more than write text. They can reason through problems, read images and documents, write code and use outside tools. OpenAI says GPT-6 improves reasoning, tool use and alignment.
OpenAI is also putting more weight on AI agents. An agent is a system that uses tools and apps to reach a goal with less human help. This shift is a main theme of the GPT-6 release.
The flagship model, GPT-6 Astra, launched on 3 September 2026. It was built for science, cybersecurity, programming and professional analysis. Later releases added cheaper and faster options.
This reflects a trend in the AI industry. The most powerful model is not the best choice for every task. A firm sending thousands of routine replies may want a fast, cheap model, while a researcher may need deep reasoning.
GPT-6 Release Timeline
GPT-6 arrived in stages instead of one launch. The name now covers several models and product features. The table lists each release and how many days it came after Astra.
| Date | Release | Main focus | Days after Astra |
|---|---|---|---|
| 3 September 2026 | GPT-6 Astra | Flagship model for advanced reasoning | 0 |
| 22 September 2026 | GPT-6 Sol and Luna | Faster, lower-cost models | 19 |
| 29 September 2026 | GPT-6.1 Sol | Better balance of performance and cost | 26 |
| 7 October 2026 | GPT-6 in ChatGPT | Wider rollout and Intelligent UI | 34 |
| 26 October 2026 | GPT-6.1 Terra | Visual reasoning and computer use | 53 |
| 30 October 2026 | GPT-6 Rosalind Research | Life sciences research, approved organizations only | 57 |
Source: OpenAI announcements. Day counts calculated by Coincentral.com.
The first release aimed for top capability. Later releases put more weight on speed, cost and access. In the 53 days after Astra, OpenAI made four more announcements.
GPT-6.1 Sol came seven days after Sol and Luna. The wider ChatGPT rollout followed 34 days after Astra. By November, GPT-6 was a group of specialized models.
OpenAI also released GPT-6 Rosalind Research on 30 October. It is for selected life sciences research customers. Access is limited to approved organizations.
It is the only release in the table with restricted access. The other releases reach users through ChatGPT or the API, subject to account eligibility.
The GPT-6 Model Family and API Prices
OpenAI sells four main GPT-6 models. They share the same core advances but suit different jobs. Prices below are per one million tokens, and a token is a small piece of text.
GPT-6 Astra is the most capable model. It costs $10 per million input tokens and $50 per million output tokens. OpenAI points to hard reasoning, complex software work, science and multi-step professional tasks.
GPT-6.1 Sol is the balanced option. OpenAI says its standard prices are one-fifth of Astra’s, at $2 for input and $10 for output. It improves on the original Sol in coding, professional work and computer use.
GPT-6.1 Terra focuses on visual reasoning and computer use. It costs $1 for input and $5 for output. OpenAI suggests it for reading technical drawings, matching product photos to catalog entries and moving through app screens.
GPT-6 Luna is the lowest-cost model. It costs $0.10 for input and $0.50 for output. It suits simple, frequent jobs such as sorting data or labeling information, though it is less capable than Astra on the hardest problems.
| Model | Best suited for | Input per 1M tokens | Output per 1M tokens |
|---|---|---|---|
| GPT-6 Astra | Hard reasoning, science, complex software | $10 | $50 |
| GPT-6.1 Sol | Strong results at a lower price | $2 | $10 |
| GPT-6.1 Terra | Visual reasoning and computer use | $1 | $5 |
| GPT-6 Luna | Simple, high-volume tasks | $0.10 | $0.50 |
Source: OpenAI standard API token rates. Extra charges may apply for tools, caching and processing modes.
These are standard rates. Charges for tools, caching and faster processing modes may be added. Developers should check the full price list before they plan a budget.
The price gaps are wide. Astra’s output price is five times Sol’s and 100 times Luna’s. That spread is why OpenAI offers four models instead of one.
GPT-6 Astra vs GPT-5.6 Sol: Benchmark Scores
GPT-5 and its updates made AI assistants useful for writing, coding and research. GPT-5.6 launched in July 2026 and focused on getting more useful work from each token. GPT-6 keeps that thread while putting more weight on tasks the AI carries out itself.
OpenAI published scores for GPT-5.6 Sol and GPT-6 Astra on six tests. Astra scored higher on every one. The table shows the results.
| Benchmark | What it measures | GPT-5.6 Sol | GPT-6 Astra | Change |
|---|---|---|---|---|
| Terminal-Bench 4.0 | Tasks in terminal environments and developer tools | 37.3% | 57.9% | +20.6 pts |
| AutomationBench | Multi-step business workflows across tools | 18.1% | 41.4% | +23.3 pts |
| DeepSWE v1.1 | Software engineering | 72.7% | 74.1% | +1.4 pts |
| FrontierMath Tier 4 | Difficult math reasoning | 83.0% | 97.6% | +14.6 pts |
| GPQA Diamond | Science reasoning | 94.6% | 96.0% | +1.4 pts |
| OSWorld 2.0 (offline subset) | Computer interaction | 65.7% | 72.6% | +6.9 pts |
Source: OpenAI published GPT-6 Astra evaluations. Results may depend on reasoning settings, tools and test method.
A simple average across the six tests rose from 61.9% to 73.3%. That is a gain of 11.4 percentage points. The average is our own calculation, not an OpenAI figure.
The gains are uneven. AutomationBench rose 23.3 points and Terminal-Bench rose 20.6 points. DeepSWE and GPQA Diamond each rose 1.4 points.
Some tests may be near their upper limits. Astra scored 96.0% on GPQA Diamond, which leaves little room to show more progress. This is one reason a single benchmark can mislead.
FrontierMath Tier 4 tests difficult math reasoning. Astra scored 97.6% against 83.0% for GPT-5.6 Sol, a gain of 14.6 points. Strong scores do not make a model error-free, and it can still miss in unfamiliar real-world cases.
The size of the change depends on the job. A user who asks AI to rewrite a short email may see little difference. A developer fixing a complex app may see a larger one.
Coding, Agents and Computer Use
The largest gains came in work that needs many steps. Coding, agent tasks and computer use all fall into this group. The chart ranks the gains in percentage points.
Coding
On Terminal-Bench 4.0, Astra scored 57.9% against 37.3% for GPT-5.6 Sol. That is a 20.6 point gain, or about 55% in relative terms. The test covers tasks in terminal environments and with developer tools.
Real programming is more than writing short snippets. Developers must read existing systems, check dependencies, edit several files and test changes. AI-written code can still add bugs or security flaws, so human review remains needed.
Model choice also matters. Astra suits demanding engineering work. Sol may offer better value when strong results are needed at a lower cost, and Terra may fit narrower tasks that involve screens or simple edits.
AI agents
On AutomationBench, Astra scored 41.4% against 18.1% for GPT-5.6 Sol. The score is more than twice as high. The test checks whether agents can finish multi-step business workflows across different tools.
An example shows what an agent does. Instead of explaining how to organize financial data, it might pull documents, extract figures, fill a spreadsheet and write an analysis. The 41.4% score shows that even the stronger model failed many such tasks.
Computer use
On the offline subset of OSWorld 2.0, Astra scored 72.6% against 65.7%. The test checks whether AI can read screens and pick the right actions. The results apply to one test setup, not every computer task.
These skills could make automation easier to set up. Companies might not need a custom link for every software tool.
Computer use also carries risk. An agent that clicks the wrong button could take an action no one wanted. Permissions and approval steps still matter for sensitive work.
| Area | Benchmark | GPT-5.6 Sol | GPT-6 Astra | Relative change |
|---|---|---|---|---|
| Coding | Terminal-Bench 4.0 | 37.3% | 57.9% | +55% |
| Coding | DeepSWE v1.1 | 72.7% | 74.1% | +2% |
| AI agents | AutomationBench | 18.1% | 41.4% | +129% |
| Computer use | OSWorld 2.0 (offline subset) | 65.7% | 72.6% | +11% |
Relative change calculated by Coincentral.com from OpenAI’s published scores and rounded to the nearest whole percent.
Factual Accuracy and Hallucinations
A hallucination is a response that sounds credible but is wrong or unsupported. It remains one of the main weaknesses of generative AI. OpenAI says GPT-6 reduces these errors.
OpenAI reports that GPT-6 Sol makes about half as many mistakes as GPT-5.6 on one internal factuality test. It reports further gains with GPT-6.1 Sol. At low reasoning effort, the share of responses with at least one factual error fell from 11.4% to 7.7%.
That is a drop of 3.7 percentage points, or about 32% in relative terms. The table and chart show the figures.
| Model | Responses with at least one factual error | Setting |
|---|---|---|
| GPT-6 Sol | 11.4% | Low reasoning effort |
| GPT-6.1 Sol | 7.7% | Low reasoning effort |
Source: OpenAI internal factuality evaluation. The test uses difficult conversations.
These numbers need context. The test uses hard conversations in which users had already found factual mistakes. It does not sample every ChatGPT chat.
The 7.7% figure does not mean GPT-6.1 Sol is 92.3% accurate overall. Users should still check important facts, especially in news, finance, health and legal work.
Cost, Speed and Value Compared With Astra
Token prices differ widely across the family. Sol costs 80% less than Astra at standard input and output rates. Terra costs 90% less and Luna costs 99% less.
The table shows what one million input tokens plus one million output tokens would cost for each model. Astra costs $60 for that volume. Luna costs $0.60.
| Model | Combined cost | Share of Astra price | Saving vs Astra |
|---|---|---|---|
| GPT-6 Astra | $60.00 | 100% | None |
| GPT-6.1 Sol | $12.00 | 20% | 80% |
| GPT-6.1 Terra | $6.00 | 10% | 90% |
| GPT-6 Luna | $0.60 | 1% | 99% |
Combined cost equals one million input tokens plus one million output tokens at standard rates. Calculated by Coincentral.com.
The cheapest model is not always the best value. A stronger model may finish a hard task in fewer tries. Businesses can track the cost per completed task, not just the price per token.
A simple example shows the math. Sol costs one-fifth as much as Astra per token. If Sol needs more than five times as many tokens to finish the same job, Astra becomes the cheaper choice.
Speed also depends on the model. Astra handles hard reasoning and may need more processing than smaller models. Luna is built for fast, low-cost jobs, and OpenAI offers processing modes that let developers pay more for quicker replies.
Total task time can matter more than reply time. A stronger model may take longer to answer yet finish a hard job in fewer attempts. That can cut repeated failures and manual fixes.
GPT-6 Access in ChatGPT and the API
OpenAI expanded GPT-6 in ChatGPT on 7 October 2026. Free and Go users get GPT-6 experiences powered by Luna. Higher plans add more features, reasoning settings and models.
Astra powers the premium Pro reasoning experience. GPT-6.1 Sol and Terra are available in supported Work and Codex environments. Terra is also available through the API.
The chart shows the access named in OpenAI’s announcements. Developers who use the API pay by the token, at the rates listed above. Users should check which models their own account includes.
Using Astra through the API is not the same as using ChatGPT’s Pro reasoning mode. Plans, tools and usage limits vary by product. OpenAI also keeps separate usage terms for developers on the API.
The 7 October rollout also introduced Intelligent UI. It lets ChatGPT build charts, graphics, buttons, forms and other interactive items inside a chat. The system picks the format based on the request.
Availability of Intelligent UI varies by plan, platform and reasoning setting. It changes how users receive information. It is separate from the tools AI agents use to operate other apps.
What the Numbers Do Not Show
Higher scores do not mean tasks are solved. Flip the numbers and the limits are clearer. On AutomationBench, Astra did not complete 58.6% of tasks.
| Benchmark | GPT-5.6 Sol not completed | GPT-6 Astra not completed |
|---|---|---|
| Terminal-Bench 4.0 | 62.7% | 42.1% |
| AutomationBench | 81.9% | 58.6% |
Calculated as 100 minus the reported score. Source scores: OpenAI published evaluations.
Astra also did not complete 42.1% of Terminal-Bench 4.0 tasks. These figures show that autonomous AI still needs supervision for many workflows.
Published scores also depend on settings. Reasoning effort, available tools and testing method can all change a result. Comparing numbers from different setups can mislead.
Safety
In September, OpenAI said Astra reached the Critical cybersecurity capability threshold under its Preparedness Framework. That label reflects skill at advanced cybersecurity tasks. It also creates risk if the skill is misused.
OpenAI says it has added safeguards, monitoring and limits around sensitive capabilities. Prompt injection, wrong tool use and ignoring task limits remain concerns. Careful testing and oversight stay important.
Business use
More capable agents open the door to automating whole workflows. A company could use AI to pull data from several teams, analyze it and write a report. Customer service tools could look into routine questions using approved apps.
The benefits could include faster work and lower admin costs. Yet the more an AI system does alone, the more safeguards matter. Companies must think about security, access permissions, data rules and the cost of wrong actions.
A practical approach is to start with narrow, measurable use cases. Teams can track results before they expand automation.
Jobs
GPT-6 strengthens the case for automating some digital tasks. Research, data processing, programming and admin work are likely candidates. Automating tasks is not the same as removing whole jobs.
Most jobs mix judgment, communication and responsibility. An analyst might use GPT-6 to gather information, but investment choices still need human judgment. The long-term effect on jobs remains uncertain.
Rival models
OpenAI faces strong competition from Anthropic and Google. Claude and Gemini models also post strong results in reasoning, coding and professional tests. The top benchmark score does not make a model the best choice for every user.
Some people prefer Claude for writing or software work. Others like Gemini for its links to Google services. ChatGPT’s strengths include its range of tools, reasoning options and growing interactive features.
Key Takeaways
- GPT-6 Astra launched on 3 September 2026, followed by Sol, Luna, Terra and a wider ChatGPT rollout.
- Astra beat GPT-5.6 Sol on all six published benchmarks, with gains from 1.4 to 23.3 points.
- Astra’s AutomationBench score rose from 18.1% to 41.4%, yet 58.6% of tasks were still not completed.
- On a hard internal test, GPT-6.1 Sol had errors in 7.7% of evaluated responses, down from 11.4% for GPT-6 Sol.
- Standard prices run from $10 input and $50 output for Astra to $0.10 and $0.50 for Luna.
- Free and Go users in ChatGPT can reach GPT-6 experiences through Luna.
For most users, GPT-6 is a useful upgrade on complex projects. For simple tasks, the change may be small. Developers building AI apps may gain the most from the agent and coding results.
GPT-6 shows steady progress, with the biggest gains in multi-step work. Results from controlled tests may differ from daily use. Reliability, security and cost will decide how useful these gains prove to be.
Readers who want to compare models should test their own tasks. A short trial with real work, tracked for cost and errors, says more than a headline score.
Sources
- OpenAI: GPT-6 Astra launch announcement and published evaluations (3 September 2026)
- OpenAI: GPT-6 Sol and Luna release (22 September 2026)
- OpenAI: GPT-6.1 Sol release and factuality results (29 September 2026)
- OpenAI: GPT-6 rollout in ChatGPT and Intelligent UI (7 October 2026)
- OpenAI: GPT-6.1 Terra release (26 October 2026)
- OpenAI: GPT-6 Rosalind Research (30 October 2026)
- OpenAI: Standard API token pricing for GPT-6 models
- GPT-6 Explained: What’s New and How It Compares to Previous Models (updated 2 November 2026)
