OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17, 2026, in ChatGPT Work, Codex and its developer platform. The company cut Astra’s standard input and output prices by 20% from the previous Astra and kept Sol’s prices steady.
The choice depends on the job. In a coding test that asks whether a change is ready to merge, OpenAI says Sol finished within half a point of Astra at about one-fifth the cost. A test of interactive 3D design put more distance between them.
The cost of a higher Intelligence Index score
Artificial Analysis combines coding, science, knowledge work and general knowledge in its Intelligence Index. Its weighted average cost per task makes it possible to compare two models even when they reach the same score at different reasoning settings. A point farther left costs less. A higher point earns a better score.
The horizontal axis uses a logarithmic scale, so the distance from $1 to $2 is the same as from $5 to $10. The labels L, M, H, XH and Max mean low, medium, high, extra high and maximum effort. The lines follow each provider’s settings in that order. Artificial Analysis tested Claude Fable 5.2 with its default fallback.
Intelligence Index score vs. cost per task
Source: Artificial Analysis model profiles
Data as of: 2026-11-17
Each point pairs a score with its cost at one reasoning setting. Only maximum has a reported cost for GPT-6.2 Sol.
| Model | Reasoning setting (full name) | Reasoning setting | Cost per task (USD) | Intelligence Index (points) | Label placement |
|---|---|---|---|---|---|
| GPT-6.2 Astra | Low | L | 0.66 | 45.7 points | below |
| GPT-6.2 Astra | Medium | M | 1.62 | 57 points | above |
| GPT-6.2 Astra | High | H | 2.07 | 60.1 points | below |
| GPT-6.2 Astra | Extra high | XH | 2.52 | 61.6 points | above |
| GPT-6.2 Astra | Maximum | Max | 4 | 62.1 points | above |
| GPT-6.2 Sol | Maximum | Max | 1.01 | 57.3 points | above |
| Claude Fable 5.2 | Low | L | 3.1 | 53.9 points | below |
| Claude Fable 5.2 | Medium | M | 3.8 | 56.4 points | above |
| Claude Fable 5.2 | High | H | 9.08 | 61.8 points | above |
| Claude Fable 5.2 | Extra high | XH | 14.83 | 62.2 points | above |
| Claude Fable 5.2 | Maximum | Max | 26.54 | 64.6 points | below |
| Gemini 4 Carbon | Low | L | 5.22 | 54.2 points | below |
| Gemini 4 Carbon | Medium | M | 4.68 | 58.3 points | below |
| Gemini 4 Carbon | High | H | 5.48 | 60 points | above |
| GPT-6 Astra | Low | L | 1.42 | 43.3 points | above |
| GPT-6 Astra | Medium | M | 2.73 | 51.6 points | below |
| GPT-6 Astra | High | H | 3.24 | 54.8 points | above |
| GPT-6 Astra | Extra high | XH | 3.66 | 55.3 points | below |
| GPT-6 Astra | Maximum | Max | 6.24 | 56.9 points | above |
| GPT-6.1 Sol | Low | L | 0.16 | 34.8 points | below |
| GPT-6.1 Sol | Medium | M | 0.36 | 46.3 points | below |
| GPT-6.1 Sol | High | H | 0.59 | 50.3 points | below |
| GPT-6.1 Sol | Extra high | XH | 0.76 | 52.5 points | below |
| GPT-6.1 Sol | Maximum | Max | 1.27 | 54.6 points | below |
Two prices for a 57-point score
Astra at medium and Sol at maximum both scored about 57 points. The average task cost $1.62 with Astra and $1.01 with Sol. Sol appears as one dot because the published profile gives a cost only for maximum effort. Its scores and token counts at the other settings appear in the next chart.
At high effort, Astra and Gemini both scored about 60. Astra cost $2.07 per task and Gemini cost $5.48. Claude at maximum reached 64.6 points for $26.54, compared with Astra’s 62.1 points for $4.00. These prices describe the mix of tasks in the index.
The earlier Astra scored 56.9 at maximum effort for $6.24 per task. Its successor reached 62.1 for $4.00. At the same setting, Sol went from 54.6 points for $1.27 to 57.3 for $1.01.
Output tokens at each reasoning setting
The next chart plots the score against the output each model generated across the suite. The horizontal axis shows millions of output tokens on a logarithmic scale. Doubling the token count moves a point the same distance anywhere on that axis. A longer step to the right means the model produced more output. A step up means it earned a higher score.
Intelligence Index score against output across the suite
Source: Artificial Analysis model profiles
Data as of: 2026-11-17
Each dot shows one reasoning setting. Equal horizontal distances indicate equal ratios of output tokens.
| Model | Reasoning setting (full name) | Reasoning setting | Output tokens (millions) | Index score (points) | Label placement |
|---|---|---|---|---|---|
| GPT-6.2 Astra | Low | L | 6.31 | 46 points | above |
| GPT-6.2 Astra | Medium | M | 12.4 | 57 points | above |
| GPT-6.2 Astra | High | H | 20 | 60 points | above |
| GPT-6.2 Astra | Extra high | XH | 27.1 | 62 points | below |
| GPT-6.2 Astra | Maximum | Max | 65.8 | 62 points | above |
| GPT-6.2 Sol | Low | L | 7.21 | 41 points | above |
| GPT-6.2 Sol | Medium | M | 14.6 | 53 points | below |
| GPT-6.2 Sol | High | H | 22.7 | 55 points | below |
| GPT-6.2 Sol | Extra high | XH | 33.3 | 56 points | above |
| GPT-6.2 Sol | Maximum | Max | 60.9 | 57 points | above |
| Claude Fable 5.2 | Low | L | 34.9 | 54 points | above |
| Claude Fable 5.2 | Medium | M | 41 | 56 points | below |
| Claude Fable 5.2 | High | H | 85.3 | 62 points | below |
| Claude Fable 5.2 | Extra high | XH | 144 | 62 points | above |
| Claude Fable 5.2 | Maximum | Max | 275 | 65 points | above |
| Gemini 4 Carbon | Low | L | 83.9 | 54 points | above |
| Gemini 4 Carbon | Medium | M | 124 | 58 points | above |
| Gemini 4 Carbon | High | H | 205 | 60 points | above |
| GPT-6 Astra | Low | L | 10.27 | 43 points | below |
| GPT-6 Astra | Medium | M | 18.17 | 52 points | above |
| GPT-6 Astra | High | H | 25.58 | 55 points | above |
| GPT-6 Astra | Extra high | XH | 32.28 | 55 points | below |
| GPT-6 Astra | Maximum | Max | 54.33 | 57 points | below |
| GPT-6.1 Sol | Low | L | 8.18 | 35 points | above |
| GPT-6.1 Sol | Medium | M | 15.12 | 46 points | above |
| GPT-6.1 Sol | High | H | 24.65 | 50 points | below |
| GPT-6.1 Sol | Extra high | XH | 33.29 | 52 points | below |
| GPT-6.1 Sol | Maximum | Max | 63.66 | 55 points | below |
Astra used 27.1 million output tokens at extra high and 65.8 million at maximum. Both scores round to 62. Sol went from 33.3 million to 60.9 million over the same step and gained one rounded point.
Claude reached 65 at maximum with 275 million output tokens. Gemini reached 60 at high with 205 million. Astra also scored 60 at high, using 20 million. These counts cover the full suite. Each provider sets its own price per token.
GPT-6.2 Astra and Sol access and input limits
OpenAI lists Plus, Pro, Enterprise and Edu among the plans with access to both models in ChatGPT Work and Codex. Developers can call either model through the application programming interface (API), where the bill depends on how much material the model reads and generates.
Both models accept text and images and return text, according to their technical guides. Each can take up to 922,000 input tokens and generate up to 128,000 output tokens. A token is a small unit of content, often part of a word. The two limits add up to a context window of 1.05 million tokens.
What a document review costs with GPT-6.2 Astra and Sol
OpenAI’s price sheets put Astra at four times the price of Sol for the same number of tokens at standard rates. Per million tokens, Astra costs $8 for fresh input, $0.40 for cached input and $40 for output. The corresponding Sol rates are $2, $0.10 and $10. Cached input is material the system has already processed and can reuse, such as instructions shared across a series of requests.
Standard API prices per million tokens
Source: OpenAI
Data as of: 2026-11-17
Astra’s output rate fell from $50 to $40 per million tokens. Both Sol generations cost $10.
| Model | Fresh input (USD) | Cached input (USD) | Output (USD) |
|---|---|---|---|
| GPT-6 Astra | 10 | 1 | 50 |
| GPT-6.2 Astra | 8 | 0.4 | 40 |
| GPT-6.1 Sol | 2 | 0.1 | 10 |
| GPT-6.2 Sol | 2 | 0.1 | 10 |
For example, consider a review that takes 100,000 fresh input tokens and generates 10,000 billable output tokens. At standard rates, Astra costs $0.80 to read the material and $0.40 to generate the output, for a total of $1.20. The same counts cost $0.20 and $0.10 with Sol, or $0.30 in total. However, an actual run can use a different number of tokens on each model.
The size of a single request also matters. OpenAI doubles the input and cache rates once a request exceeds 272,000 input tokens. Output then costs 1.5 times the base rate. Those rates apply to the whole request, including the portion below the threshold.
How coding scores change with the price of a task
Cognition’s FrontierCode 2.0 tests whether a maintainer would merge a model’s changes to an existing codebase. OpenAI says that at maximum effort, Sol finished within half a point of Astra at about one-fifth the cost on Cognition’s test.
Artificial Analysis runs its own implementation of that benchmark, FrontierCode-AA 2.0, using 132 tasks. The chart uses only its measurements. Each dot pairs the average cost of one task with the score for the same model and effort. As in the index chart, better scores are higher and lower prices are farther left.
FrontierCode-AA 2.0 score vs. cost per task
Source: Artificial Analysis, FrontierCode-AA 2.0
Data as of: 2026-11-17
Each point pairs a score with its cost at one reasoning setting. Only maximum has a reported cost for GPT-6.2 Sol.
| Model | Reasoning setting (full name) | Reasoning setting | Cost per task (USD) | FrontierCode-AA score (%) | Label placement |
|---|---|---|---|---|---|
| GPT-6.2 Astra | Low | L | 0.81 | 32.66% | below |
| GPT-6.2 Astra | Medium | M | 1.91 | 46.56% | below |
| GPT-6.2 Astra | High | H | 3.08 | 47.13% | above |
| GPT-6.2 Astra | Extra high | XH | 3.75 | 48.95% | above |
| GPT-6.2 Astra | Maximum | Max | 5.87 | 45.38% | above |
| GPT-6.2 Sol | Maximum | Max | 1.3 | 44.38% | below |
| Claude Fable 5.2 | Low | L | 2.12 | 38.44% | above |
| Claude Fable 5.2 | Medium | M | 2.69 | 41.3% | below |
| Claude Fable 5.2 | High | H | 9.64 | 49.13% | below |
| Claude Fable 5.2 | Extra high | XH | 21.41 | 47.58% | below |
| Claude Fable 5.2 | Maximum | Max | 43.37 | 48.76% | above |
| Gemini 4 Carbon | Low | L | 10.12 | 29.62% | above |
| Gemini 4 Carbon | Medium | M | 7.44 | 39.45% | below |
| Gemini 4 Carbon | High | H | 8.88 | 40.56% | below |
| GPT-6 Astra | Low | L | 1.34 | 26.22% | below |
| GPT-6 Astra | Medium | M | 2.8 | 36.58% | below |
| GPT-6 Astra | High | H | 3.77 | 40.14% | below |
| GPT-6 Astra | Extra high | XH | 4.04 | 41.4% | above |
| GPT-6 Astra | Maximum | Max | 6.54 | 42.37% | below |
| GPT-6.1 Sol | Low | L | 0.2 | 20.16% | below |
| GPT-6.1 Sol | Medium | M | 0.41 | 33.39% | below |
| GPT-6.1 Sol | High | H | 0.63 | 36.74% | below |
| GPT-6.1 Sol | Extra high | XH | 0.7 | 38.52% | above |
| GPT-6.1 Sol | Maximum | Max | 1.13 | 40.63% | below |
In the Artificial Analysis run, Astra scored 48.95% at extra high for $3.75 per task. At maximum it spent $5.87 and scored 45.38%. Sol at maximum scored 44.38% for $1.30. The one-point gap belongs to this run. OpenAI’s half-point comparison comes from a separate evaluation.
At maximum effort on FrontierCode-AA, GPT-6 Astra scored 42.37% for $6.54 per task, compared with 45.38% for $5.87 for its successor. The Sol comparison moved the other way on price. GPT-6.1 Sol scored 40.63% for $1.13, while GPT-6.2 Sol scored 44.38% for $1.30.
Claude scored 49.13% at high for $9.64 per task. At maximum its average cost reached $43.37 and its score was 48.76%. Gemini’s high setting reached 40.56% for $8.88. Its low setting cost more than its high setting in this run. A setting’s name alone does not tell a developer what the task will cost.
How judges compared interactive 3D designs
For interactive 3D design, Artificial Analysis gives two models the same brief. Judges try the resulting experiences without seeing the models’ names and vote for the one they prefer. The chart uses Artificial Analysis ratings at maximum effort for both generations and Claude Fable 5.2, rounded to the nearest Elo point. A higher Elo rating indicates a stronger record in the judges’ paired comparisons. The gap between the Elo ratings does not give the percentage of judges who chose Astra.
Blind ratings for interactive 3D designs
Source: Artificial Analysis, AA-3D
Data as of: 2026-11-17
Astra rose from 1,168 to 1,318 Elo points. Sol rose from 1,121 to 1,259.
| Category | Rating |
|---|---|
| GPT-6.2 Astra | 1,318 Elo |
| Claude Fable 5.2 | 1,271 Elo |
| GPT-6.2 Sol | 1,259 Elo |
| GPT-6 Astra | 1,168 Elo |
| GPT-6.1 Sol | 1,121 Elo |
OpenAI also tested work that people usually hand over as a file. On its GDPval v2 benchmark, models receive the files and messages available at the start of a consulting, finance or design assignment. Astra improved by 6.2 percentage points over GPT-6 Astra while generating 22% fewer output tokens. Sol gained 5.3 points over GPT-6.1 Sol.
One spreadsheet test asks a question a financial analyst might face when reading public accounts. A model has to find how much California Medicaid spent on medical transport. Across the wider AA-AnalystAgent test, OpenAI says Astra answered 62.5% of questions correctly on all five attempts. A question had to yield the right answer every time to count toward that figure.
How GPT-6.2 Astra and Sol diagnose research failures
OpenAI’s system card describes a test built from 41 bugs that had taken experienced researchers hours or days to diagnose. Six further tasks asked the model to find a problem without being told what to look for. The graders gave partial credit. A suggestion to restart a stuck program could earn some points, while an explanation of the underlying fault and a repair could earn full credit.
A second test asked each model to write a small program for an OpenAI chip. It had to produce the correct result and run faster than a reference program. On chip programming, Astra scored 76.92% and Sol 74.71%. On debugging, their scores were 81.85% and 74.66%. GPT-6 Astra earned 66.72% on chip programming and 78.05% on debugging. The system card gives no GPT-6.1 Sol result for either test. All bars show the average share of available grading credit.
Research engineering scores on two OpenAI tests
Source: OpenAI system card
Data as of: 2026-11-17
The Astra–Sol gap was 7.19 points on debugging and 2.21 points on chip programming.
| Evaluation | GPT-6.2 Astra | GPT-6.2 Sol | GPT-6 Astra |
|---|---|---|---|
| Research debugging | 81.85% | 74.66% | 78.05% |
| KernelGen 1P | 76.92% | 74.71% | 66.72% |
A separate experiment gave each model two hours and one Nvidia H100 graphics processor to improve a training script. Success meant reaching a fixed prediction target in less time. On version five of that test, Astra saved 42.87% of the baseline training time on average. Sol saved 14.59%. OpenAI tested the previous Astra on version four, which limits comparisons across generations.
GPT-6.2 Astra and Sol in mental health conversations
OpenAI also reports a test on which both new models lost ground. MentalHealthBench uses 1,215 synthetic conversations and criteria written by clinicians. Researchers sampled four replies per conversation at maximum reasoning effort. They averaged those replies within each task before calculating the overall scores.
The chart shows the overall scores and their standard errors. Of the 1,215 cases, 650 were non-acute, 221 were high-acuity and 344 were emergent. Astra scored above Sol in the first two groups. In the most urgent group, Sol scored 52.6% and Astra 51.0%. That small gap is within one standard error for either estimate.
MentalHealthBench scores at maximum reasoning effort
Source: OpenAI system card
Data as of: 2026-11-17
Bars show overall scores across 1,215 conversations. Error bars show one standard error.
| Model | Overall score | Standard error (percentage points) |
|---|---|---|
| GPT-6 Astra | 58.7% | 1 |
| GPT-6.2 Astra | 53.3% | 1 |
| GPT-6.1 Sol | 57.9% | 1 |
| GPT-6.2 Sol | 50.9% | 1 |
The breakdown by urgency shows where the results diverge. Each row covers a different set of conversations.
| Conversation group | GPT-6.2 Astra | GPT-6.2 Sol | GPT-6 Astra | GPT-6.1 Sol |
|---|---|---|---|---|
| Non-acute | 53.4% | 49.4% | 58.5% | 57.3% |
| High-acuity | 56.4% | 52.9% | 59.8% | 59.1% |
| Emergent | 51% | 52.6% | 58.5% | 58% |
Source: OpenAI system card
Data as of: 2026-11-17
Each row describes a different group of conversations. Percentages are average rubric scores.
OpenAI traced about half of each model’s overall decline to asking for needed context less often. For instance, a concise reply can score poorly if the situation calls for a follow-up question before the model can offer useful guidance. The study graded generated replies against criteria set for each conversation. It did not measure patient outcomes.
How safety scores changed over repeated attempts
OpenAI ran a second test in which evaluators could change their next request after seeing a model’s reply. To count as a success, the model had to withhold harmful assistance in every attempted conversation. Giving the evaluators more attempts made that condition harder to meet.
The bars compare all four models at the largest attempt budget, where OpenAI reports an exact figure for each one. At the smallest budget, GPT-6.2 Astra scored 96.1% and GPT-6.2 Sol scored 95.4%. Both ended at 80.9%. The full figure in the system card includes intermediate estimates and 95% confidence intervals.
Multiturn safety at the largest tested attempt budget
Source: OpenAI system card
Data as of: 2026-11-17
At the largest budget, both new models reached 80.9%.
| Model | Success rate |
|---|---|
| GPT-6 Astra | 67.2% |
| GPT-6.2 Astra | 80.9% |
| GPT-6.1 Sol | 63.6% |
| GPT-6.2 Sol | 80.9% |
At the largest budget, the previous Astra scored 67.2% and GPT-6.1 Sol scored 63.6%. This particular test rewards consistency across repeated attempts. A model can give many safe replies and still fail a case if one conversation yields harmful assistance.
What happens when a file is missing
OpenAI reports a weakness that matters when users hand off a job and return later to inspect the result. In two tests, the new models more often failed to tell users that an obstacle had blocked their work. In one scenario, the researchers removed the very file the model was supposed to edit. OpenAI says it added instructions to Codex that reduced such omissions.
The next pricing change
OpenAI offers its Ultrafast mode for both models through the API. In ChatGPT Work and Codex, Pro 500 subscribers can use it with either model, while Pro 200 subscribers get it with Sol. Eligible Enterprise and Edu users need a workspace administrator to enable the mode.
Developers budgeting for Ultrafast have a date to watch. Both model guides list the API rate at three times the standard price through February 28, 2027. OpenAI lists six times the standard price starting March 1, 2027, Pacific time.
Sources: OpenAI announcement, OpenAI API changelog, OpenAI GPT-6 Astra guide, OpenAI GPT-6.1 Sol guide, OpenAI Astra model guide, OpenAI Sol model guide, Cognition FrontierCode methodology, Artificial Analysis FrontierCode-AA results, OpenAI system card, Artificial Analysis Astra results, Artificial Analysis Sol results, Artificial Analysis Claude results, Artificial Analysis Gemini results, Artificial Analysis GPT-6 Astra results, Artificial Analysis GPT-6.1 Sol results, Artificial Analysis AA-3D
