GPT-6 Astra
⧉GPT-6 Astra is the next-generation flagship OpenAI released in September 2026: a 1,050,000-token context window, up to 128,000 output tokens, reasoning_effort up to xhigh, image input, tool calling and prompt caching. Fully OpenAI-SDK compatible - set model to gpt-6-astra - and served on both /chat/completions and /responses. Pay-as-you-go, failed calls never billed; cached input bills at one tenth, and we do not add the long-context surcharge the official API applies.
What is the GPT-6 Astra API and what does it cost?
GPT-6 Astra is the flagship of OpenAI's GPT-6 family, released on September 3, 2026, with a 1,050,000-token context window and up to 128,000 output tokens. On NezhaGate it costs $3.00 per 1M input tokens and $15.00 per 1M output tokens (OpenAI list price: $10 / $50 for prompts up to 272K tokens). Base URL: https://nezhagate.com/v1 (OpenAI-compatible).
Measured performance
Real streaming requests sent through the public NezhaGate API. Test window 2026-10-20 18:17 – 2026-10-20 18:37 UTC; sample size 100.
Failed requests in this sample: HTTP 429 × 1. Failed calls are not billed.
How this was measured
POST /v1/chat/completions with stream: true, one at a time with a 1-second gap, rotating through 10 prompts that each ask for about 200 words on a software topic, max_tokens 2500, reasoning_effort not sent (model default).
Time to first token runs from sending the request to the first streamed token of any kind (answer or reasoning text).
Output speed is answer tokens divided by the time from the first to the last answer token; reasoning tokens are subtracted when the API reports them, and a request is left out when the token count cannot be separated from hidden thinking.
Success means HTTP 200 and a completed stream with a non-empty answer. Percentiles are nearest-rank over successful requests. The test key belongs to an internal account, so these calls are not part of the public success-rate figures elsewhere on the site.
Measurements are repeated every week and this page updates itself; the date above is the latest run.
Cost example
NezhaGate figures use the live price on this page; vendor figures use the vendor's published list price (sources below).
1,000 coding-assistant calls
Each call sends 20,000 input tokens (code plus instructions) and receives 2,000 output tokens, without prompt caching. Reasoning tokens count as output tokens.
| Where | Cost | Difference |
|---|---|---|
| NezhaGate | $90.00 | — |
| OpenAI API, standard tier (prompts up to 272K tokens) | $300.00 | NezhaGate 70% cheaper |
OpenAI charges 2x input and 1.5x output for the whole request once a prompt passes 272K tokens ($20 / $75 per 1M). NezhaGate bills one flat rate at every prompt length, and cached input at $0.30 per 1M tokens.
Vendor list prices checked 2026-09-25: OpenAI: GPT-6 Astra model page, OpenAI API pricing.
Parameters & capabilities
"Vendor spec" is what the model's maker publishes. "Via NezhaGate" is what our own requests through the gateway showed (last checked 2026-10-20).
| Item | Vendor spec | Via NezhaGate |
|---|---|---|
| Model id | gpt-6-astra OpenAI: GPT-6 Astra model page |
gpt-6-astra |
| Endpoints | — | /v1/chat/completions and /v1/responses |
| Price per 1M tokens | $10 input · $1 cached input · $50 output (prompts up to 272K); $20 · $2 · $75 above 272K OpenAI: GPT-6 Astra model page | $3.00 input · $0.30 cached input · $15.00 output, at any prompt length |
| Context window | 1,050,000 tokens OpenAI: GPT-6 Astra model page | Not measured by us |
| Max output | 128,000 tokens OpenAI: GPT-6 Astra model page | ✗ max_tokens is not enforced on this route (see limits below) |
| Streaming | — | ✓ Yes (SSE), with usage in the final chunk |
| Tool calling | — | ✓ Yes, OpenAI tools / tool_calls format on /v1/chat/completions |
| Image input (vision) | — | ✓ Yes (image_url, base64 data: URL tested) |
| JSON mode | — | ✓ Yes, response_format: json_object returned bare JSON |
| Structured output | — | ✓ Yes, json_schema with strict: true returned matching JSON |
| Reasoning effort | low, medium, high, xhigh, max OpenAI: GPT-6 Astra model page |
✓ low to max accepted; omitted means medium; minimal and none are sent as low |
| temperature / top_p | — | Accepted without error but have no effect on this route |
Tool setup
Base URL https://nezhagate.com/v1, your NezhaGate API key, and the model id gpt-6-astra. Request and response samples (cURL, Python, Node.js) are in the GPT-6 Astra API docs.
- Open Cursor Settings → Models.
- Under API Keys, paste your NezhaGate key into OpenAI API Key, turn on Override OpenAI Base URL and enter
https://nezhagate.com/v1. - Add a custom model named
gpt-6-astra, enable it, and pick it in the chat model menu.
Differences from the official API & known limits
max_tokensandmax_completion_tokensare accepted but not enforced on this route: in our weekly test a request withmax_tokens: 60ran to its natural length. Limit answer length in the prompt instead.temperature,top_p,presence_penaltyandfrequency_penaltyare accepted without an error but have no effect.- If you omit
reasoning_efforton/v1/chat/completions, NezhaGate sendsmedium(the model's own default), so a reasoning summary, when the model writes one, streams back indelta.reasoning_content; most short answers have none.minimalandnoneare sent aslow. - Reasoning tokens are part of
usage.completion_tokensand are billed at the output rate;usage.completion_tokens_details.reasoning_tokensshows the count when it is non-zero. - No long-context surcharge: the 2x input / 1.5x output multiplier OpenAI applies to prompts over 272K tokens does not exist here.
- No Batch, Flex or priority tiers: every call is billed at the single rate shown on this page.
- Claude Code cannot use this model (it speaks only the Anthropic Messages API, which NezhaGate serves for Claude models). Use it from Cursor, Cline, n8n, Open WebUI or any OpenAI SDK.
- If a line fails before a response starts, NezhaGate retries on a backup line; a stream that has already started is not moved. Failed calls are never billed.
GPT-6 Astra API: frequently asked questions
How do I call GPT-6 Astra with the OpenAI SDK?
https://nezhagate.com/v1, use your NezhaGate API key, and set model to gpt-6-astra. No other code changes are needed; both /v1/chat/completions and /v1/responses work.How much does GPT-6 Astra cost on NezhaGate compared with OpenAI?
How fast is GPT-6 Astra through NezhaGate?
Can I limit the answer length with max_tokens?
max_tokens is accepted but not enforced on NezhaGate's GPT-6 Astra route, so ask for a shorter answer in the prompt.Are failed requests billed?
Live Test · Playground
Try out GPT-6 Astra right here (available after login).
Input
Advanced
This model is served over a ChatGPT-subscription upstream that does not accept temperature / top_p / max_tokens (they are silently ignored). Use reasoning_effort below for thinking depth, and prompt wording for length.
Conversation
About GPT-6 Astra
GPT-6 Astra is the next-generation flagship OpenAI released in September 2026, served on NezhaGate through the OpenAI-compatible API: a 1,050,000-token context window, up to 128,000 output tokens, reasoning_effort for thinking depth (up to xhigh), image input, tool calling and prompt caching, on both /chat/completions and /responses. Set model to gpt-6-astra and keep your SDK and request shape. Pay-as-you-go, failed calls never billed, and cached input settles at one tenth of the input rate.
Use cases
1,050,000 tokens holds a whole repository, a contract set or a long report in one call, with no chunking and stitching of your own.
It speaks the OpenAI-compatible endpoint, so Codex, Cursor and Claude Code switch to GPT-6 Astra by changing the model field alone.
reasoning_effort goes up to xhigh, so you can spend the thinking budget on the hardest part and keep simple requests on low.
Image input with text output: screenshot understanding, chart questions and document parsing all use the same endpoint.
How to choose
Pick GPT-6 Astra for the strongest reasoning and the hardest tasks; GPT-6 Sol is cheaper for complex coding and agentic workflows, and GPT-6 Luna fits high-volume, low-latency work. All three belong to the GPT-6 family and share a 1,050,000-token context, so switching is a one-field change.
Related Models
Explore other models you can integrate.
GPT-6.1 Sol
The GPT-6 Sol upgrade - near-Astra coding and agents
Gemini 3.1 Pro
Next-generation flagship reasoning model from Google
Claude Opus 5
Next-generation flagship from Anthropic
Gemini 3.8 Flash
Latest Flash · adaptive thinking
Claude Fable 5
Anthropic Fable series · narrative and long-form writing
