In the GPT-6.2 Astra vs GPT-6 Astra comparison, OpenAI reports a nine-percentage-point improvement on a science benchmark at less than one-third of the task cost. The result is a company-reported evaluation of scientific workflows, not a promise that every research task will improve by the same amount. The models are available across OpenAI products and Microsoft Foundry, giving teams several ways to assess the newer model. OpenAI’s announcement describes the benchmark comparison.
| Comparison point | GPT-6.2 Astra | GPT-6 Astra |
|---|---|---|
| Input price per million tokens | $8 | $10 |
| Cached input price per million tokens | $0.40 | $1 |
| Output price per million tokens | $40 | $50 |
| Context window | 1.05 million tokens | 1.05 million tokens |
| Maximum output | 128,000 tokens | 128,000 tokens |
These API figures come from the official GPT-6.2 Astra model page and GPT-6 Astra model page. Token prices and benchmark task costs measure different things.
What OpenAI changed on November 17, 2026
OpenAI’s API changelog records the GPT-6.2 API release on November 17, 2026, following the GPT-6 Astra release on September 3, 2026. Its GPT-6.2 product update, also dated November 17, describes availability in ChatGPT Work, Codex and the API.
Microsoft’s Foundry update, updated November 17, says GPT-6.2 Astra and Sol became generally available there as well. For a business planning an AI integration, project consultation can help define the workflow and scope before implementation.
What the nine-point science result actually measures
OpenAI reports that GPT-6.2 Astra scored nine percentage points higher than GPT-6 Astra on Terminal-Bench-Science 0.1 while using less than one-third of the task cost. The benchmark evaluates scientific research workflows performed with terminal tools. OpenAI also reports a new high on the Vals Science Index, a separate evaluation. These are results reported by OpenAI, rather than independently reproduced findings. See its model announcement for the comparison.
A percentage-point difference compares the scores directly; it does not mean the newer model completed nine percent more work or that its score rose by nine percent relative to the older model. The result is most relevant to tasks resembling the evaluated workflows, where a model must work through scientific problems using terminal tools.
The cost comparison describes the expense of completing benchmark tasks. It does not establish what a particular team will spend: an application’s total depends on the amount of input and output it processes, along with factors such as caching and tool use. A lower cost on the benchmark is therefore a reason to investigate a workflow, rather than a fixed saving to apply to a budget.
Benchmark outcomes can also depend on the evaluation setup, including prompts, tools and reasoning settings. OpenAI’s GPT-6.2 system card describes its evaluation methods and their limitations. For teams considering document assistants or other research workflows, the relevant question is how well a model handles their own tasks; AI development services cover workflow integration and evaluation.

How GPT-6.2 Astra changes developer workflows
For an API integration, the model identifier is the starting point: the GPT-6.2 Astra model page lists gpt-6.2-astra. A request’s model field can therefore identify the newer model as follows:
{"model":"gpt-6.2-astra"}That field is only part of a request, not a complete application. Developers replacing an existing model should also check how their application sends requests and handles tool calls. OpenAI’s GPT-6.2 migration guide says tool calling requires the Responses API and that reasoning effort cannot be set to none. An integration that depends on a different setting or endpoint for tool calls needs corresponding changes.
Tools and application design
The model supports function calling and structured outputs, alongside tools including web search, file search, code interpreter, hosted shell, computer use, MCP and tool search, according to its model documentation. These capabilities let an application combine generated responses with selected tools; the application’s design still determines which tools are connected and how their results are used.
For a customer-facing or internal mobile product, this distinction affects the integration plan: a chat screen is the interface, while request handling and tool connections form the workflow behind it. Techno Particles’ application development services describe work on mobile applications and integrations.
Compatibility and prompt caching
OpenAI’s migration guidance also calls for checking unsupported parameters and changes to prompt caching. Preserve stable prompt prefixes where practical, then examine the request structure when moving an existing integration. This matters because changing a model identifier alone does not address differences in supported settings or how repeated prompt content is handled.
Pricing, access and usage limits
OpenAI lists GPT-6.2 Astra API pricing at $8 per million input tokens, $0.40 per million cached input tokens and $40 per million output tokens. A cache write with 30-minute retention costs $10 per million tokens, according to the GPT-6.2 Astra model page.
The GPT-6 Astra model page lists $10 per million input tokens, $1 per million cached input tokens, $12.50 per million cache-write tokens and $50 per million output tokens. The newer model’s listed input and output rates are each 20% lower; its cached-input rate is 60% lower.
Both model pages list a 1.05-million-token context window and a maximum output of 128,000 tokens. The GPT-6.2 Astra page also lists a maximum input of 922,000 tokens. Context capacity and output capacity are separate limits: a large context window does not mean a single response can use the entire window.
Choosing an access route
For direct API use, the listed rate limits for GPT-6.2 Astra are 5,000 requests per minute and one million tokens per minute on the Build tier; 10,000 requests and four million tokens on Launch; and 15,000 requests and 40 million tokens on Grow. Both request and token limits matter when planning concurrent workloads.
OpenAI says Plus, Pro, Enterprise and Edu users can access the models in ChatGPT Work and Codex. Microsoft lists Global Standard, Data Zone Standard and Global Provisioned Throughput deployment types for Foundry; the Foundry announcement also describes platform-specific pricing and controls. Availability depends on the product, workspace settings and applicable usage tier.
For organisations budgeting a document assistant or another integrated workflow, generative AI development services describe implementation and workflow integration support. The appropriate access route depends on where the application runs and which deployment controls it needs.

Safety findings and operational limits
OpenAI’s GPT-6.2 announcement and its system card place GPT-6.2 Astra at the company’s High threshold for AI self-improvement and for biological and chemical capabilities. The model also meets OpenAI’s Critical threshold for cybersecurity. These labels describe OpenAI’s assessment under its safety framework; they are not measures of how often a particular deployment will encounter a harmful request.
The GPT-6.2 system card also reports greater evaluation awareness and stronger action-only monitor-evasion capability than GPT-6 Astra in some tests. Evaluation awareness concerns a model’s ability to recognize that it is being assessed. Monitor evasion concerns behavior that can make an action harder for an oversight mechanism to detect. These findings make the design of monitoring and human oversight relevant when a model can take actions through connected tools.
What the evaluation conditions mean
OpenAI notes that some capability evaluations use configurations without production safeguards, and the system card describes results as lower bounds. A finding under those conditions should therefore be read alongside the test setup, rather than treated as a prediction of behavior in every deployed application. The system card reports OpenAI’s own evaluations; it does not independently replicate their findings.
For an organisation considering a system that can act on information, the practical distinction is between generating an answer and carrying out an action. A review of where human approval belongs can be incorporated into the requirements and scope work described on the project consultation services page. The appropriate controls depend on what the application can access and what consequences its actions could have.
Choosing between the two models
The decision behind GPT-6.2 Astra vs GPT-6 Astra: 9-Point Science Gain at One-Third Cost depends on your own workload. Choose GPT-6.2 Astra if its quality and task-cost gains justify migration; retain GPT-6 Astra if compatibility or existing validation matters more.
Next, run both models on the same fixed workload and record quality, output tokens, latency, tool calls and interventions.




Leave a comment
Your email stays private. You can see your comment immediately on this browser. Others see it after approval.