Updated
What is the GPT 6.2 Astra API?
The GPT 6.2 Astra API is live on UnificAlly as openai/gpt-6.2-astra, from $1.20/1M tokens of input and $6/1M tokens of output. GPT 6.2 Astra is OpenAI's most capable model, released on November 17, 2026 as the upgrade to GPT 6 Astra. OpenAI built it for long, hard projects: building and revising software, using a computer, and producing slides, spreadsheets, and documents that follow your template. It takes text and image input, returns text, and works across a 1,050,000-token context window with a 128,000-token max output and knowledge up to May 18, 2026. It scores higher than GPT 6 Astra on every capability benchmark in OpenAI's announcement and costs less per token. Reasoning effort runs from low up to max.
GPT 6.2 Astra API at a glance
| Model ID | openai/gpt-6.2-astra |
| Endpoint | POST https://direct-api.unifically.com/v1/chat/completions |
| Other formats | POST /v1/responses and POST /v1/messages |
| Request types | Chat, streaming, tool calls |
| Input | Text and images |
| Output | Text |
| Context window | 1,050,000 tokens (up to 922,000 input) |
| Max output | 128,000 tokens |
| Reasoning effort | low to max, set with reasoning_effort |
| Input rate | $1.20/1M tokens |
| Output rate | $6/1M tokens |
| Cache read rate | $0.06/1M tokens |
| Cache creation rate | $1.50/1M tokens |
The GPT 6.2 Astra API bills per token, with no subscription. The rates above apply to prompts up to 272,000 tokens, and longer prompts bill at the long-context rate. The OpenAI API pricing guide covers the rates for every GPT model.
How to use the GPT 6.2 Astra API
- Create an API key on the API keys page. The same key and balance work for every model on UnificAlly.
- Send a POST request to
/v1/chat/completionsonhttps://direct-api.unifically.comwith the model ID and amessagesarray. This is the OpenAI-compatible endpoint shared by every language model on UnificAlly. - Read the reply from
choices[0].message.contentin the same response. There is no task to check.
curl -X POST https://direct-api.unifically.com/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer YOUR_API_KEY" -d '{
"model": "openai/gpt-6.2-astra",
"reasoning_effort": "high",
"messages": [
{ "role": "user", "content": "Update last quarter's earnings deck with the figures from this release, keeping the same slide template." }
]
}'
The response comes back synchronously with the completion and the token usage:
{
"id": "3f8e2a1c-7b4d-4e9a-9c2f-5d1e8b7a6c40",
"object": "chat.completion",
"created": 1710000000,
"model": "openai/gpt-6.2-astra",
"choices": [
{
"index": 0,
"message": {
"role": "assistant",
"content": "Here is the updated deck outline, slide by slide..."
},
"finish_reason": "stop"
}
],
"usage": {
"prompt_tokens": 28,
"completion_tokens": 64,
"total_tokens": 92,
"prompt_tokens_details": {
"cached_tokens": 0
},
"completion_tokens_details": {
"reasoning_tokens": 0
}
}
}
Set "stream": true to receive tokens as they generate, and raise reasoning_effort to xhigh or max for the hardest multi-step tasks. Send requests to direct-api.unifically.com so long replies are not cut short by a proxy timeout.
GPT 6.2 Astra API parameters
| Parameter | Type | Default | What it does |
|---|---|---|---|
model | string | required | Model identifier, openai/gpt-6.2-astra |
messages | array | required | Non-empty array of message objects with role and content (a string or an array of content parts). Roles: system, developer, user, assistant, tool, function |
stream | boolean | false | If true, the response is streamed using server-sent events |
stream_options | object | none | Set { "include_usage": true } to get a final chunk with token usage. Only used when stream is true |
temperature | number | none | Sampling temperature between 0 and 2. Higher values produce more random output |
max_tokens | integer | none | Maximum number of tokens to generate in the completion |
max_completion_tokens | integer | none | Accepted as an alias for max_tokens |
reasoning_effort | string | none | How much the model reasons before answering, from low to max |
top_p | number | none | Nucleus sampling parameter. An alternative to temperature |
stop | string or array | none | Up to four sequences where the model stops generating further tokens |
tools | array | none | List of tools the model may call. Each tool requires a type and a function definition |
tool_choice | string or object | none | Controls which tool is called, if any: none, auto, required, or a specific tool |
response_format | object | none | Set { "type": "json_object" } to enable JSON mode |
The GPT 6.2 Astra API documentation has the model page, and the Chat Completions reference linked from it covers streaming, tool calls and error codes.
What GPT 6.2 Astra can do
Top score on Agents' Last Exam
Agents' Last Exam v1.1 tests long, multi-step professional work. GPT 6.2 Astra scores 78.3% at xhigh effort, 5.2 points above GPT 6 Astra at less than a third of its cost per task, and 5.5 points above Claude Fable 5.2, which spends $60.05 a task there against $2.89. Gemini 4 Carbon is the closest rival at 77.4%.
First on the Vals Science Index
On Terminal-Bench-Science 0.1, where agents analyze data and run simulations from a terminal, GPT 6.2 Astra reaches 77.1% in OpenAI's run, 9 points above GPT 6 Astra for $7.15 a task against $23.80. Vals AI's independent re-run puts it first of 47 models on the same benchmark at 71.43% and first of 27 on the Vals Science Index.
Legal and finance deliverables that pass review
Harvey LAB-AA v1.1 only credits a legal deliverable that meets every requirement with no material hallucination. GPT 6.2 Astra passes 15.0%, more than double Claude Fable 5.2 at 7.2%, at about a sixth of the cost. On OpenAI's GDPval v2 set of real consulting, finance, and design projects it scores 50.4%, up from 44.2% for GPT 6 Astra.
First on AA-3D for interactive 3D and web builds
On Artificial Analysis's AA-3D board, human judges pick between two blinded 3D experiences built from the same prompt. GPT 6.2 Astra leads at 1,318 Elo, ahead of Claude Fable 5.2 at 1,271 and GPT 6.2 Sol at 1,259. OpenAI also trained it to pick layouts and imagery that suit the goal of a website.
What GPT 6.2 Astra is best for
Professional agent work
Long multi-step jobs across real software. 78.3% on Agents' Last Exam v1.1.
Decks, workbooks, and memos
Slides that keep your template down to the font size, and valuation workbooks built from source files.
Legal and finance drafting
15.0% on Harvey LAB-AA v1.1, more than double Claude Fable 5.2.
Scientific computing
Data analysis and simulation from the command line. 77.1% on Terminal-Bench-Science 0.1.
Websites and 3D builds
Interactive pages, games, and 3D scenes from a written brief. First on AA-3D at 1,318 Elo.
Engineering design
Part-by-part CAD models, assembly guides, and simulation setups through computer use.
GPT 6.2 Astra use cases
Hand GPT 6.2 Astra last quarter's deck and this quarter's filings and have it rebuild the slides in the same template, with headlines that state the takeaway. Give it financial statements and assumptions and ask for an editable valuation workbook whose three statements reconcile. Legal teams can run contract review and research memos where a single invented fact fails the work. Research groups can point it at a terminal to analyze data, run simulations, and fit models. Product teams can turn a written brief into an interactive website or a playable 3D scene. Keep the prompt prefix the same between turns and append new instructions, which OpenAI recommends for better cache reuse and lower cost.
GPT 6.2 Astra limitations
GPT 6.2 Astra does not lead every board. The independent Artificial Analysis Intelligence Index, pulled November 18, 2026, scores it 62 at max effort, third of 91 models behind Claude Fable 5.2 at 65. Fable 5.2 also leads on HLE-Diamond, and in OpenAI's own FrontierCode 2.0 chart it scores 49.1% against 44.2% for Astra. On Terminal-Bench 4.0, Gemini 4 Carbon edges it on Artificial Analysis, 65% against 64%.
It is slow at max effort. Artificial Analysis measures 66 output tokens per second and about 135 seconds before the first answer token. Drop to medium or high for interactive work.
Low effort falls off sharply. On OpenAI's LifeVal test it scores 46.3% at low against 78.9% at medium, so keep low for simple, high-volume calls. none is rejected.
Security work is restricted. OpenAI launched it with tighter cyber safeguards that also block some legitimate vulnerability research and security testing. In two of OpenAI's alignment tests it was also more likely to leave out an obstacle that blocked its work, such as a missing file, so check agent reports.
Long prompts cost more. Past 272,000 tokens of input, the whole request bills at the long-context rate: double on input and cache, and 1.5x on output.
GPT 6.2 Astra vs GPT 6 Astra
GPT 6.2 Astra beats GPT 6 Astra on every capability benchmark OpenAI charted for both, and it does so with fewer tokens. Agents' Last Exam v1.1 rises from 73.1% to 78.3%, Terminal-Bench-Science 0.1 from 68.1% to 77.1%, Harvey LAB-AA v1.1 from 8.6% to 15.0%, and SecureSWE, which tests writing secure code, from 46.5% to 65.7% at max effort. On GDPval v2 it gains 6.2 points while using 22% fewer output tokens. OpenAI also set its token prices 20% below GPT 6 Astra's, with a 60% cut on cached input, so the cost per finished task drops further than the per-token price. Both models meet OpenAI's Critical cybersecurity threshold, and GPT 6.2 Astra is the first OpenAI model to meet the High threshold for AI self-improvement. There is little reason to stay on GPT 6 Astra.
When to use GPT 6.2 Astra
Use GPT 6.2 Astra for the hardest long jobs: professional documents that must follow a template, legal and finance work where errors fail the task, scientific computing, and agent runs across real software. xhigh is often enough; save max for the toughest problems. For everyday coding and chat, GPT 6.2 Sol answers faster for much less per token.
FAQs
People also ask about GPT 6.2 Astra
openai/gpt-6.2-astra, called through the OpenAI-compatible POST /v1/chat/completions endpoint with your UnificAlly API key. The same ID also works on /v1/responses and /v1/messages, and tool calls work on all three.
Sign up on UnificAlly and create a key on the API keys page. One key and one balance cover GPT 6.2 Astra and every other model, with no subscription.
Send a POST request to https://direct-api.unifically.com/v1/chat/completions with the model ID openai/gpt-6.2-astra and a messages array. The reply comes back in the same response, in the OpenAI chat completion format. Set "stream": true to receive it as it is written.
1,050,000 tokens in total, with up to 922,000 tokens of input and 128,000 tokens of output. Knowledge runs up to May 18, 2026. It takes text and image input and returns text.
low, medium, high, xhigh, and max, set with reasoning_effort. OpenAI rejects none for this model, so every request does some reasoning. On Agents' Last Exam, xhigh scored slightly higher than max (78.3% against 77.6%) at a lower cost per task.
OpenAI reports 78.3% on Agents' Last Exam v1.1, 77.1% on Terminal-Bench-Science 0.1, and 15.0% on Harvey LAB-AA v1.1, the top score in each of its charts. Independently, it is first of 27 models on the Vals Science Index at 68.44% and first on Artificial Analysis's AA-3D board at 1,318 Elo.
Yes, on every capability benchmark OpenAI published for both. Agents' Last Exam goes from 73.1% to 78.3%, Terminal-Bench-Science 0.1 from 68.1% to 77.1%, and GDPval v2 from 44.2% to 50.4% with 22% fewer output tokens. It also costs less per token.
On professional work and science, yes. It leads Fable 5.2 on Agents' Last Exam (78.3% against 72.8%) and Terminal-Bench-Science 0.1 (77.1% against 73.3% in OpenAI's run). Fable 5.2 still scores higher on the Artificial Analysis Intelligence Index (65 against 62), HLE-Diamond, and FrontierCode 2.0.
Once a request's prompt passes 272,000 tokens, the whole request bills at the long-context rate. Input and cache rates double and output bills at 1.5x. Both rates are listed on the pricing page.
OpenAI rates it at the Critical level for cybersecurity and launched it with tighter safeguards for higher-risk security help. Those safeguards also block some legitimate vulnerability research and security testing.
OpenAI released GPT 6.2 Astra on November 17, 2026, together with GPT 6.2 Sol, as the upgrade to GPT 6 Astra. It is live on UnificAlly as openai/gpt-6.2-astra.
GPT 6.2 Astra guides and comparisons
Review pricing, limits, and tested outputs before running this model.
