gpt-6-astraGPT-6 Astra is OpenAI's September 2026 flagship, and on APIMODELS it costs $2.40 per 1M input tokens and $12.00 per 1M output — 76% below OpenAI's own $10 / $50 list, with cached input at $0.24. Where it clearly wins is computer use and long-horizon agents: 72.6% on OSWorld 2.0 against GPT-5.6 Sol's 65.7% (and OpenAI reports a 47% cut in task completion time), 57.7% on Terminal-Bench 4.0 against Sol's 37.3%, 88.0% on SRE-Bench against 55.9%, and 100% on ExploitBench against 78.5%. Long-context retrieval is the other real gain: 96.3% on OpenAI MRCR v2 8-needle in the 512K–1M band, where Sol scores 73.8%. On maths and science it leads too — 97.6% on FrontierMath Tier 4 v2 and 96.0% on GPQA Diamond. Be aware of what does NOT change: on general coding the gains are marginal, with FrontierCode 1.1 at 53.3% against Fable 5.1's 53.5% and Opus 5's 53.4%, and DeepSWE v1.1 at 74.1%; on Humanity's Last Exam with tools it trails Claude at 57.2% against Fable 5.1's 65.0%. The widely-quoted 99.9% on ARC-AGI-3 needs OpenAI's stateful adapter harness — plain API callers should not expect it. Context is 1,050,000 tokens with 128,000 max output, input is text + vision, and requests above 272K input tokens bill at 2× input / 1.5× output, the same step as GPT-5.6. Call it on /v1/chat/completions or /v1/responses with one API key, reachable from China with no organization verification, and only successful requests are billed. ⚠️ What this API does and does not give you (measured on our own endpoint on 2026-09-06, not copied from a spec sheet): function calling works on both /v1/chat/completions and /v1/responses; the built-in web_search tool works; code_interpreter works. The computer_use_preview tool and file_search are NOT available — both of our upstream channels reject them. So if you saw the ChatGPT app drive a browser or a desktop on its own and you need that exact behaviour, this API cannot give it to you. Worth separating: much of that experience is the app, not the model — the design tooling, deep research and sandboxed computer environment are scaffolding OpenAI built around it, and a first-party API key does not hand you those either. What transfers to any API is the model plus function calling, and that part we serve in full.
View complete API reference with all parameters and examples.
Enable real-time streaming responses with Server-Sent Events.
{
"model": "gpt-6-astra",
"stream": true,
"messages": [...]
}Enable the model to use tools and call functions.
{
"model": "gpt-6-astra",
"tools": [{
"type": "function",
"function": {
"name": "get_weather",
"description": "Get current weather for a location",
"parameters": {
"type": "object",
"properties": {
"location": {"type": "string", "description": "City name"}
},
"required": ["location"]
}
}
}],
"messages": [{"role": "user", "content": "What's the weather in Tokyo?"}]
}Get structured JSON responses from the model.
{
"model": "gpt-6-astra",
"response_format": {"type": "json_object"},
"messages": [{"role": "user", "content": "Extract info as JSON: John is 30 years old"}]
}| Parameter | Type | Required | Description |
|---|---|---|---|
| model | string | Yes | Model identifier (e.g., gpt-6-astra) |
| messages | array | Yes | Array of message objects with role and content |
| max_tokens | integer | No | Maximum tokens in the response |
| stream | boolean | No | Enable streaming responses (SSE) |
| temperature | number | No | Sampling temperature (0.0 - 2.0) |
| top_p | number | No | Nucleus sampling threshold (0.0 - 1.0) |
| tools | array | No | Function calling tools definition |
| response_format | object | No | Output format (e.g., json_object) |
View complete API reference with streaming, thinking, and more.
Billing: Cost = (input_tokens * input_price + output_tokens * output_price) / 1,000,000
$2.40 / $12.00 against the official $10 / $50
Verified on our own endpoint on 2026-09-06: function calling (both /chat/completions and /responses), the built-in web_search tool, and code_interpreter all work and return real tool-call items. NOT available through this API: the computer_use_preview tool and file_search — both of our upstream channels reject them. If autonomous computer control is a hard requirement, this API cannot give it to you, and we would rather say so before you integrate
The model itself scores 72.6% on OSWorld 2.0 against Sol's 65.7%, and OpenAI reports 47% less time per task. That is the model's capability, not a tool we can hand you: those runs use OpenAI's own harness. Through any API — ours or first-party — you build the agent loop yourself with function calling
The design tooling, deep research and sandboxed computer environment you see in the ChatGPT app are product scaffolding OpenAI built around the model. A first-party API key does not hand you that experience either. What transfers to any API is the model plus function calling — and that part we serve in full
Terminal-Bench 4.0 57.7% vs 37.3%; SRE-Bench 88.0% vs 55.9%
MRCR v2 8-needle 96.3% in the 512K–1M band (Sol: 73.8%)
GPT-6 Astra is a Large Language Model API provided by OpenAI. GPT-6 Astra is OpenAI's September 2026 flagship, and on APIMODELS it costs $2.40 per 1M input tokens and $12.00 per 1M output — 76% below OpenAI's own $10 / $50 list, with cached input at $0.24. Where it clearly wins is computer use and long-horizon agents: 72.6% on OSWorld 2.0 against GPT-5.6 Sol's 65.7% (and OpenAI reports a 47% cut in task completion time), 57.7% on Terminal-Bench 4.0 against Sol's 37.3%, 88.0% on SRE-Bench against 55.9%, and 100% on ExploitBench against 78.5%. Long-context retrieval is the other real gain: 96.3% on OpenAI MRCR v2 8-needle in the 512K–1M band, where Sol scores 73.8%. On maths and science it leads too — 97.6% on FrontierMath Tier 4 v2 and 96.0% on GPQA Diamond. Be aware of what does NOT change: on general coding the gains are marginal, with FrontierCode 1.1 at 53.3% against Fable 5.1's 53.5% and Opus 5's 53.4%, and DeepSWE v1.1 at 74.1%; on Humanity's Last Exam with tools it trails Claude at 57.2% against Fable 5.1's 65.0%. The widely-quoted 99.9% on ARC-AGI-3 needs OpenAI's stateful adapter harness — plain API callers should not expect it. Context is 1,050,000 tokens with 128,000 max output, input is text + vision, and requests above 272K input tokens bill at 2× input / 1.5× output, the same step as GPT-5.6. Call it on /v1/chat/completions or /v1/responses with one API key, reachable from China with no organization verification, and only successful requests are billed. ⚠️ What this API does and does not give you (measured on our own endpoint on 2026-09-06, not copied from a spec sheet): function calling works on both /v1/chat/completions and /v1/responses; the built-in web_search tool works; code_interpreter works. The computer_use_preview tool and file_search are NOT available — both of our upstream channels reject them. So if you saw the ChatGPT app drive a browser or a desktop on its own and you need that exact behaviour, this API cannot give it to you. Worth separating: much of that experience is the app, not the model — the design tooling, deep research and sandboxed computer environment are scaffolding OpenAI built around it, and a first-party API key does not hand you those either. What transfers to any API is the model plus function calling, and that part we serve in full. Through APIMODELS platform, you can access this model via a unified API with transparent pay-as-you-go pricing. Current pricing: Input: $2.40, Output: $12.00 per 1M tokens.
Build intelligent conversational systems to automatically answer user queries and improve service efficiency.
Automatically write articles, emails, ad copy, and other text content to boost productivity.
Assist with code writing, debugging, and code review to accelerate software development.
Understand and analyze unstructured data, extract key insights, and generate summary reports.
GPT-6 Astra is available through APIMODELS at: Input: $2.40, Output: $12.00 per 1M tokens. Billing is pay-as-you-go — you only pay for what you generate.
Sign up at APIMODELS, get your API key, and call our unified API endpoint. We provide detailed API documentation with code examples in cURL, Python, and Node.js.
APIMODELS offers the same GPT-6 Astra model through our aggregation platform. We provide a unified API interface so you do not need separate accounts for each provider - one API key to access all models.
Not Computer Use. Most of the rest, yes. We measured this on our own endpoint on 2026-09-06 rather than copying a spec sheet. **Works:** custom function calling (on both /v1/chat/completions and /v1/responses, returning real tool_call items), the built-in `web_search` tool, and `code_interpreter` — all three return genuine tool-call output. **Does not work:** `computer_use_preview` (Computer Use) and `file_search`; both of our upstream channels reject those tool types. So if what you need is the model driving a browser or a desktop on its own, this API cannot give it to you, and we would rather tell you now than have you find out after integrating. One thing worth separating out: much of what you saw in the ChatGPT app is **the app, not the model**. The design tooling, the deep-research pipeline, the long-horizon orchestration and the sandboxed computer environment are product scaffolding OpenAI built around the model. Even a first-party API key does not hand you that experience — you would still write the agent loop yourself, using function calling as the primitive. What transfers to any API is the model plus function calling, and that part we serve in full, at 76% below list. On long-horizon work we add no cap of our own beyond the model context window (1.05M; requests over 272K bill at 2× input / 1.5× output by OpenAI's own rule). Very long single non-streaming requests can hit a ~100s edge timeout — use streaming for those.
Depends on the job. **Computer use and long-horizon agents are a genuine step up**: 72.6% on OSWorld 2.0 against Sol's 65.7% (OpenAI also reports a 47% cut in task completion time), 57.7% on Terminal-Bench 4.0 against 37.3%, and 88.0% on SRE-Bench in one attempt against 55.9%. Long-context retrieval too: 96.3% on MRCR v2 8-needle in the 512K–1M band, where Sol scores 73.8%. Maths and science lead as well — FrontierMath Tier 4 v2 97.6%, GPQA Diamond 96.0%. **But general coding is flat**: FrontierCode 1.1 sits at 53.3% against Fable 5.1's 53.5% and Opus 5's 53.4%. So switch for browser/terminal automation, SRE work and very long context; for plain code writing, Sol ($1.324 / $6.618) is 45% less and barely behind.
$2.40 per 1M input tokens and $12.00 per 1M output, with cached input at $0.24. OpenAI's own list is $10 / $50 (cache read $1.00, cache write $12.50), so this is **76% below** — the same footing as our GPT-5.6 Sol at 67% below. ⚠️ One tier to know about: when a single request exceeds 272K input tokens, the **whole request** bills at 2× input and 1.5× output. That is OpenAI's own rule and we pass it through. Failed requests are never billed, there is no minimum top-up or subscription, and the endpoint is reachable from China with no organization verification.
**No.** That figure requires OpenAI's stateful adapter harness, where the model carries state across turns. A plain API call is stateless, and OpenAI says so directly: stateless API callers should not expect ~99% out of the box. We spell this out because it is the most misquoted number from this launch.
A 1,050,000-token context window with up to 128,000 output tokens, and text + image input. It speaks the OpenAI-compatible API: POST /v1/chat/completions or /v1/responses — point base_url at https://api.apimodels.app/v1 and set model to gpt-6-astra, no SDK change. Streaming, function calling and structured output are all verified working.
Three worth knowing up front. First, **general coding gains are marginal** (see the first question). Second, **it trails Claude on Humanity's Last Exam with tools**: 57.2% against Fable 5.1's 65.0% and Opus 5's 63.6% — for broad-knowledge deep reasoning the Claude line is still stronger. Third, two things OpenAI disclosed itself: a regression in chain-of-thought monitorability (shorter reasoning output, harder to audit under adversarial prompting), and heavily gated cybersecurity capabilities — at launch it refuses tasks like writing proof-of-concept exploits, and the safety checks can occasionally interrupt legitimate work.
On APIMODELS, GPT-6 Astra runs alongside 60+ models on one API key and one balance, so choosing is about fit, not lock-in. It supports 76% Below Official, Function Calling + Web Search, Code Interpreter, 1.05M Context, No Computer Use, No file_search, and you can weigh it on price and capability against other Large Language Model models, then switch by changing a single model-name string — no new account or integration. Browse every Large Language Model option with live pricing at apimodels.app/models.
GPT-6 Astra supports: 76% Below Official, Function Calling + Web Search, Code Interpreter, 1.05M Context, No Computer Use, No file_search. See the APIMODELS docs for full parameters and call examples.
Yes. APIMODELS exposes GPT-6 Astra through a single unified API and one key — no separate provider accounts, and no need to handle each provider's regional network access yourself.
We support Stripe (Visa, Mastercard, and other international cards) and Alipay. Credits are available instantly after payment.
How to get access, regional availability, and how this model compares with its alternatives.