Prompts up to 272,000 tokens
GPT-6 Astra
GPT-6 Astra: a 1,050,000-token window with the pricing cliff mapped
openai/gpt-6-astra runs on the NextModel gateway at $10/M input, $50/M output, and $1/M cache hits for prompts up to 272,000 tokens — then $20/$75/$2 above that line, with the window topping out at 1,050,000 tokens. Catalog specs, the pricing cliff, early-tester quotes with the skeptics included, and the official sources, on one page.
Pricing
The 272K cliff, to scale
Per-million catalog rates for openai/gpt-6-astra, plotted against prompt size. The dashed lines mark the flat $10/$50 listing Claude Fable 5.1 carries across its whole 1,050,000-token window — a comparison of published catalog rates, not a benchmark.
Source: NextModel catalog listings for openai/gpt-6-astra and anthropic/claude-fable-5.1. Rates are USD per million tokens.
Catalog
The catalog listing, 1:1
Every number below is transcribed from the NextModel catalog entry for openai/gpt-6-astra.
Prompts up to 272,000 tokens
Single-request window
2,048 by default; raise it per request
Tier table
| Prompt size | Input / 1M | Output / 1M | Cache hit / 1M |
|---|---|---|---|
| Up to 272,000 tokens | $10 | $50 | $1 |
| Above 272,000 tokens | $20 | $75 | $2 |
The window tops out at 1,050,000 tokens.
APIs
Chat Completions · Responses
Capabilities
ThinkingStreamingComputer UseFunction Calling
Early testers
What early testers say
Quotes are third-party opinions from people running their own tests — NextModel has not measured this model. Every quote is verbatim and linked, and the skeptics are part of the record.
Early wins
“I got early access to GPT-6 Astra: when I say this model broke through tasks I couldn't crack with 5.6 Sol or Fable, I mean it specifically”
“Astra caught approximately 4% more labeled bugs through actionable findings than GPT-5.6 Sol, and 22% more than Opus 5”
Measured takes
“With the Provider Adapter harness, Astra (high) scores 99.9% for $19K.”
ARC's own caveat, from the same post: the ARC-AGI-3 Semi-Private Standard harness scores 62.7% for $26K, and ARC explicitly notes it is "not claiming that it is AGI".
“a significant decrease in hallucination rate from 92% to 51% at max effort”
“GPT-6 Astra and Claude Fable 5.1 both cost $10/$50 per million tokens, but win on different scoreboards.”
The skeptics
“It's very expensive. After 15 message I burned through my 5 hour limits.”
“GPT-6 Astra (on Light) stops a lot and often in unintuitive ways that I haven't seen in a while”
“First chart: from score 61 (GPT-5.6 Sol) to drumroll 61 (GPT-6 Astra)”
Primary sources
Official sources
Read the vendor's own material before deciding. These are OpenAI's pages, not ours.
OpenAI announcement
The launch post, including OpenAI's own benchmark claims.
OpenAI API docs
Parameter reference for the model on OpenAI's developer docs.
OpenAI system card
OpenAI's published safety and capability documentation.
Reading guide
Start here
Seven posts and two reference pages cover the decision end to end.
GPT-6 Astra on NextModel: 1M context, computer use, one API
The spec sheet read closely: what is listed, and what is still unverified.
GPT-6 Astra vs GPT-5.6 Sol: when the 2x upgrade pays for itself
Where the published long-context scores justify the price step, and where they do not.
GPT-6 Astra vs Claude Fable 5.1: reading the launch-week debate
What launch-week coverage actually published, and how to A/B the debate on your own traffic.
GPT-6 Astra pricing explained: the 272K cliff, cache, and worked examples
The boundary math between the two rate bands, and when cache hits change it.
Migrating to GPT-6 Astra: model ids, responses API, and a rollback plan
A one-string id change on an OpenAI-compatible gateway, plus the fallback leg worth planning for.
1M context in production: what GPT-6 Astra's retrieval numbers actually buy you
OpenAI's published 96.3% retrieval past 512K, and what that does not cover for RAG decisions.
GPT-6 Astra on a budget: cache, routing, and paying frontier rates only when it matters
Cache hits, sub-272K prompt discipline, and routing routine work to cheaper catalog legs.
Catalog page: gpt-6-astra
The live listing with price tiers, health, and code snippets.
Early testers roundup
A digest of launch-week feedback as it accumulates.
FAQ
GPT-6 Astra questions
Which model id do I send?
Send openai/gpt-6-astra to https://api.nextmodel.app/v1 with the OpenAI SDK or any OpenAI-compatible client. The catalog lists both the Chat Completions and Responses APIs, with Thinking, Streaming, Computer Use, and Function Calling capabilities.
What does openai/gpt-6-astra cost, and how big is the context?
$10 per million input tokens, $50 per million output tokens, and $1 per million cache-hit tokens for prompts up to 272,000 tokens; above that boundary, $20, $75, and $2. The context window is 1,050,000 tokens and max output is 8,192 tokens (2,048 by default).
How does the 272,000-token tier work in practice?
The catalog prices two prompt-size bands. Prompts up to 272,000 tokens bill at $10 input / $50 output / $1 cache hit per million. Once a prompt passes 272,000 tokens, the per-token rates change: prompt tokens above 272,000 bill at $20 input / $75 output / $2 cache hit per million. Worked example: a call with a 400,000-token prompt has 128,000 tokens above the boundary, and those bill at $20/$75 rather than $10/$50. The window ends at 1,050,000 tokens.
Which claims on this page are unverified?
Three. The NextModel catalog lists no latency band for openai/gpt-6-astra yet. There is no first-party CacheSafety benchmark row for it. And the OSWorld 72.6% computer-use score plus the MRCR v2 retrieval results (100% at 256K-512K, 96.3% at 512K-1M) are OpenAI's published evaluations, not NextModel measurements. The tester quotes are opinions, not benchmarks.
When should I route to a cheaper model id instead?
When prompts stay well under 272,000 tokens, the task is high-volume and routine, or a cheaper catalog entry already clears your quality bar, route there and reserve openai/gpt-6-astra for runs that need the 1M window, computer use, or the top reasoning tier. The pricing calculator prices the same workload against both ids.
Does the review coverage include criticism?
Yes, on purpose. Alongside the positives, this page quotes a tester who burned through usage limits after 15 messages, one who found the computer-use agent stopping in unintuitive ways, and one who flagged a benchmark score that did not move from GPT-5.6 Sol. Launch-week signal is mixed, and the quotes keep it that way.
Catalog facts transcribed 1:1 from the NextModel catalog. Last reviewed 2026-09-06.