OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17, 2026. Astra's standard input and output rates fell by 20% from GPT-6 Astra, and its score on a multistep work benchmark rose. Sol improved on the work benchmark while keeping GPT-6.1 Sol's rates, so the same paid plan can cover more tasks.
Start with Astra for a company slide deck assembled from several documents or a complex code fix. Sol is the starting point for meeting notes, everyday coding, and recurring work.
API pricing and context across four models
The API lets software call a model directly. Prices are in US dollars per million tokens, the pieces of text that the API counts when it measures usage. Cached input gets a discounted rate when the model can reuse material from an earlier request.
| Model | Official positioning | Standard input / cached input / output | API context |
|---|---|---|---|
| GPT-6 Astra | Complex reasoning, coding, and document work | $10 / $1 / $50 | 1,050,000 tokens |
| GPT-6.2 Astra | Higher-tier model for complex coding and professional work | $8 / $0.40 / $40 | 1,050,000 tokens |
| GPT-6.1 Sol | Everyday coding and work that uses tools | $2 / $0.10 / $10 | 1,050,000 tokens |
| GPT-6.2 Sol | Everyday writing, coding, and recurring multistep work | $2 / $0.10 / $10 | 1,050,000 tokens |
Context is the amount of information a model can handle in one request. All four models have a context window of 1,050,000 tokens and a maximum output of 128,000 tokens. The two new models have a separate input limit of 922,000 tokens. They accept text and images and return text.
GPT-6.2 Astra pricing and performance on document and code tasks
Astra's standard input rate fell from $10 to $8, and its output rate went from $50 to $40. Each dropped 20%. The cached input rate dropped 60%, from $1 to $0.40. All three rates are now four times those of GPT-6.2 Sol.
OpenAI's GDPval v2 uses multistep tasks modeled on actual work. Astra scored 6.2 percentage points higher than its predecessor and used 22% fewer output tokens per task.
Think of turning source reports into a presentation that follows a company template or testing a code change across services. Astra follows requirements such as layout and tone more closely. For a presentation, the raw reports supply the material, and the company template supplies the rules for how that material should appear.
GPT-6.2 Sol pricing and benchmark results
GPT-6.2 Sol retains GPT-6.1 Sol's rates of $2 for standard input, $0.10 for cached input, and $10 for output. Its score rose 5.3 percentage points on OpenAI's GDPval v2. It fits recurring tasks that involve reading written material and images and calling a search tool along the way.
Artificial Analysis measured both new models at the Max reasoning setting.
| Artificial Analysis metric (Max) | GPT-6 Astra → GPT-6.2 Astra | GPT-6.1 Sol → GPT-6.2 Sol |
|---|---|---|
| Intelligence Index v5.0.1 overall score | 57 → 62 | 55 → 57 |
| Estimated average time per task on the same index | 20.4 minutes → 10.4 minutes | 23.1 minutes → 7.9 minutes |
| Coding Agent Index (separate Codex test) | Tied | −3 points |
To estimate time per task, Artificial Analysis divides the output token count by the output speed, then weights the result according to the benchmark. Sol lost three points on the separate test that measured coding agents through Codex. When you delegate a code change, compare the result with the requirements and check the test results for anything the model missed.
Starting models and reasoning levels by task
The reasoning level sets how deeply the model works through a problem before answering. A quick wording edit and a bug that spans services call for different starting settings.
| Task | Starting model | Starting reasoning level |
|---|---|---|
| Short questions or polishing text | GPT-6.2 Sol | Light |
| Summarizing meeting notes into decisions, action items, and owners, or reviewing proposed code changes | GPT-6.2 Sol | Medium |
| Combining several documents into slides that follow a company template | GPT-6.2 Astra | Medium |
| Tracing the cause of a code problem across several services | GPT-6.2 Astra | High |
For meeting notes, ask for a table with decisions, owners, and deadlines. When combining documents, ask the model to list which materials it used and include the sentences that support its work. That gives you specific passages to check against the finished result.
Model selection in ChatGPT Work, Codex, and the API
At launch, model selection was available in Work, Codex, and the API. The regular Chat menu was outside this rollout.
ChatGPT Work model menu
- Open Work in ChatGPT. The Work composer has a settings row below it.
- Open the model menu below the composer. On screens with the Power slider, open Advanced to see the model list.
- Choose GPT-6.2 Astra or GPT-6.2 Sol. The selected model appears below the composer.
Codex app and command line
- Open the model selector in the Codex app. On the command line, enter
/modelto see the available models. - Select GPT-6.2 Astra or GPT-6.2 Sol to apply it to the task.
API model and reasoning settings
- Set
modeltogpt-6.2-astraorgpt-6.2-solto route the request to the model you want. - For a new integration, start with
mediumreasoning. Both models rejectnoneas a reasoning setting. Use the recommended default when sending the request. - Use Responses for search tool calls or the beta feature that splits work among agents. These agents use tools to carry out steps of a task, and Responses connects tool calls with text replies. Chat Completions also supports basic conversations.
Work is available on Plus, Pro, Business, Enterprise, and Edu. Enterprise and Edu require an administrator to enable it. Work and Codex share usage, and connecting a personal API key incurs API charges.
Ultrafast generates tokens up to 4.5 times as fast as Standard. It applies to both models on Pro $500, to Sol in Codex on Pro $200, and to eligible models on Enterprise and Edu.
Long-input rates, reporting gaps, and security restrictions
- Long inputs. For all four models, an input above 272,000 tokens triggers higher rates for the entire request. Standard and cached input rates double, and the output rate rises to 1.5 times the standard rate.
- Reporting gaps. The system card says omissions increased in some tests when the model reported on blocked work, such as tasks with missing files or unavailable tools. Codex reduced those omissions in the tests by updating its default instructions. For a task with several files, ask for a list of the files the model read and a separate list of any it missed.
- Security restrictions. The new models apply stricter default restrictions to high-risk cybersecurity requests.
Sources
- OpenAI GPT-6.2 Astra and Sol launch announcement
- ChatGPT Learn guide to Work and Codex models and plans
- OpenAI API GPT-6.2 migration guide
- OpenAI API model comparison and pricing
- OpenAI GPT-6.2 system card
- Artificial Analysis scores and task times for OpenAI models
Translated from the Korean original: Naver Blog · Tistory
This post was researched and drafted with AI assistance.




Top comments (0)