GPT-6.2 Astra is OpenAI's new model for difficult coding, research and professional projects. It accepts text and images, returns text, and can work with tools through the Responses API. Its API model identifier is gpt-6.2-astra.
Quick verdict: evaluate GPT-6.2 Astra when a task combines difficult reasoning with a deliverable that must be usable, such as a tested software change, a financial workbook or a document built from several source files. Its lower token prices make it a meaningful replacement candidate for GPT-6 Astra. GPT-6.2 Sol remains the cheaper starting point for work that does not require Astra's additional capability.
What changed from GPT-6 Astra?
In the launch announcement, OpenAI emphasizes better professional outputs and more efficient completion of demanding tasks. It reports improvements in long-running workflows, scientific terminal work and documents that follow supplied templates. These are launch claims from the developer, rather than results independently reproduced for this guide.
The verified price change is substantial: Standard short-context input and output rates fall by 20%, and cached input is 60% cheaper. The documented maximum input and output limits remain unchanged. Teams therefore have a clear economic reason to test the successor, even before deciding how much its quality changes matter to their own work.
Documented specification | GPT-6.2 Astra | GPT-6 Astra |
|---|---|---|
Standard input per million tokens | $8 | $10 |
Standard output per million tokens | $40 | $50 |
Cached input per million tokens | $0.40 | $1 |
Context window | 1,050,000 tokens | 1,050,000 tokens |
Maximum input | 922,000 tokens | 922,000 tokens |
Maximum output | 128,000 tokens | 128,000 tokens |
Knowledge cutoff | May 18, 2026 | April 30, 2026 |
These values come from the 6.2 Astra reference and GPT-6 Astra reference. The shared capacity should not be mistaken for identical behavior. Integrations also need to account for the way 6.2 preserves or drops earlier reasoning.
GPT-6.2 Astra specifications
OpenAI documents text and image input with text output. Image generation is available as a tool in an appropriate workflow, but that is separate from a model directly returning images through an image-generation endpoint. Realtime audio and fine-tuning are not supported on the model page.
Feature | Documented support |
|---|---|
API model ID | gpt-6.2-astra |
Context and output | 1.05M context; 922K input; 128K output |
Modalities | Text and image input; text output |
APIs | Responses and Chat Completions |
Tool calling | Use Responses |
Structured Outputs | Supported |
Search and files | Web search and file search |
Execution and actions | Code execution, hosted shell, computer use, MCP and related tools |
Delegation | Responses Multi-agent beta |
Reasoning effort none | Rejected |
Last verified | November 17, 2026 |
For a new integration, the GPT-6.2 guide recommends starting at medium reasoning effort. Existing configurations using none or minimal should start with low instead. Effort settings affect the comparison: do not attribute a change in cost or latency entirely to the model when the reasoning configuration has also changed.
Benchmarks: what OpenAI reports
The announcement contains several different kinds of evidence. A professional-work assignment, a terminal science task and a human preference vote measure different outcomes. Their scores should remain attached to their evaluation names and conditions, rather than being collapsed into a universal accuracy claim.
Evaluation | Result reported in the announcement | Important condition |
|---|---|---|
Agents' Last Exam | 5.2 points above GPT-6 Astra at less than one-third the cost | Multistep professional workflows; OpenAI-reported comparison |
Terminal-Bench-Science 0.1 | 9 percentage points above GPT-6 Astra at less than one-third the cost | Scientific work performed in a terminal |
GDPval v2 | 6.2 percentage points above GPT-6 Astra with 22% fewer output tokens | OpenAI's internal multistep professional-work evaluation |
AA-AnalystAgent | 62.5% of questions answered correctly on all five independent attempts | A repeated-success measure, not ordinary single-attempt accuracy |
AA-3D | 1,318 Elo | Human preference between blinded interactive 3D experiences |
These are figures reported by OpenAI, including its summaries of named external evaluations. This guide has not rerun those tests or separately verified the underlying external leaderboards. Benchmark cost ratios are specific to the tested configurations; the API price change alone does not establish the same savings on every application.
OpenAI also compares Astra with Claude Fable 5.2 on Agents' Last Exam. Its announcement notes that the Fable leaderboard results exclude nine tasks using a fallback model. That qualification matters when reading the cross-provider comparison. A simple headline about which system wins would omit a material evaluation condition.
Pricing and long-context costs
The official model page lists $8 input and $40 output per million tokens at Standard short-context rates. Requests above 272,000 input tokens have doubled input and cache rates and a 1.5-times output rate for the full request. The threshold is based on input, not the advertised size of the context window.
Standard rate per million tokens | Up to 272K input tokens | More than 272K input tokens |
|---|---|---|
Input | $8 | $16 |
Output | $40 | $60 |
Cached input | $0.40 | $0.80 |
30-minute cache write | $10 | $20 |
12-hour cache write | $20 | $40 |
A workload that passes 272K input tokens becomes more expensive across the whole request. For source-heavy research, count the actual input after adding system instructions, file contents, history and tool results. A PDF's page count is not a reliable substitute for measured token usage.
The examples below assume uncached input at Standard rates and the stated billed output volumes. They exclude cache writes, tools and retries. They illustrate arithmetic, not measured cost per completed task.
Request | Input cost | Output cost | Token subtotal |
|---|---|---|---|
100K input and 10K output | $0.80 | $0.40 | $1.20 |
250K input and 10K output | $2.00 | $0.40 | $2.40 |
300K input and 10K output | $4.80 | $0.60 | $5.40 |
Batch is priced at half of Standard, while Fast costs twice Standard at the same context length. Ultrafast has a time-limited rate of three times Standard through February 28, 2027, rising to six times Standard on March 1, Pacific time. At short context, that means promotional Ultrafast input/output rates of $24/$120, rather than a permanent $24/$120 price. Check the pricing page before budgeting.
Astra or GPT-6.2 Sol?
Sol's published Standard rates are $2 input and $10 output per million tokens, with $0.10 cached input. At the same short-context volumes for those components, Astra costs four times as much. Both models list the same maximum input and output capacities.
Workload | Practical starting point | Reason to consider the other model |
|---|---|---|
Difficult scientific or professional project | Astra | Sol may meet the acceptance standard at lower cost |
Repeated coding and agent steps | Sol | Escalate cases that fail validation or require more capable reasoning |
Template-sensitive deliverable | Compare both using the actual template | The announcement emphasizes Astra's output quality; verify the usable result |
Large document input | Compare on retrieval correctness | Context capacity is shared, so capacity alone does not decide |
Expensive failure or extensive rework | Test Astra against the current workflow | Review cost may outweigh token savings |
OpenAI reports that Sol comes within 0.5 points of Astra on FrontierCode 2.0 at about one-fifth the cost, with both at max effort. That is a specific coding evaluation result, not proof that Sol and Astra are interchangeable. It does make a strong case for including Sol in a coding procurement test instead of evaluating only the most expensive model.
The integration change to check before migrating
With 6.2, the API can drop reasoning items after changes to earlier conversation history, a model downgrade, or a change of end user or organization. OpenAI warns that losing those items can reduce answer quality or cause the model to repeat earlier reasoning. That can also raise cost and latency.
The practical recommendation in the prefix-preservation guide is to keep earlier input unchanged and append new instructions. Applications that rewrite old tool results, replace earlier instructions or summarize history themselves should test those paths. Supported OpenAI compaction has documented handling that differs from an arbitrary custom summary.
Integration check | What to verify |
|---|---|
Conversation history | Earlier messages and tool results are replayed consistently |
Tool calling | Requests use Responses, not a Chat Completions tool workflow |
Model fallback | Downgrades do not silently invalidate assumptions about preserved reasoning |
Request parameters | Unsupported sampling and logprob settings are removed |
Reasoning configuration | Medium for a new integration; supported settings for migrated requests |
Data handling | ZDR approval, endpoint eligibility and residency match the application |
Access and limitations
The launch announcement says the models are available through the API and in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, with eligible workspace access depending on administrator settings. Product usage limits and API billing are separate. Verify access on the route you plan to use.
The integration guide documents US and EU residency with Standard, Fast and Batch, while Ultrafast supports US residency only. Zero Data Retention is supported for approved organizations on eligible endpoints and features. These are conditional capabilities, not a statement that every account automatically has the same terms.
OpenAI's highlighted cybersecurity benchmark results use internal configurations with safeguards and classifiers disabled. They should not be presented as unrestricted capabilities available to ordinary users. The announcement says reduced-refusal Daybreak access currently covers earlier specialist configurations, with work planned to bring 6.2 models to the program.
Frequently asked questions
How much does GPT-6.2 Astra cost?
Standard short-context API rates are $8 input, $0.40 cached input and $40 output per million tokens. Long-context and faster processing modes have different rates.
Is it cheaper than GPT-6 Astra?
Published base input and output rates are 20% lower, and cached input is 60% lower. Actual completed-task cost also depends on token usage, tool calls and failed attempts.
Does it generate images directly?
The base model returns text and can accept images. A supported image-generation tool is a separate capability in the surrounding workflow.
Should I migrate immediately?
Test the actual deliverables and conversation-history behavior first. The price reduction is verified; the workload-specific quality and cost improvement must be measured in the application.
Sources and updates
Checked November 17, 2026 against the Astra model reference, pricing, integration guide and launch announcement. For the previous version, see our GPT-6 Astra guide. Explore other options in the AI model directory. Provider benchmark claims are attributed; this guide does not claim new independent testing.