Secret Chat Logo

GPT 6 Astra Is Here:
OpenAI's Most Capable Model, Now on Secret Chat

Launch Private Chat

Published September 5, 2026

Today we added GPT 6 Astra to the ChatGPT model on Secret Chat AI. It is live, it is manual, and it is expensive enough that those first two facts belong in the same sentence. We verified on our production server on September 5 that our OpenAI key reaches the real gpt-6-astra model and gets an answer.

What GPT-6 Astra Is

OpenAI announced GPT-6 Astra on September 3 and calls it “our most capable model, built for the hardest end-to-end work.” Its model page points to complex reasoning, coding, computer use, research and document creation. The API model has a 1,050,000-token context window, accepts text and images, and can return as many as 128,000 text tokens. Its knowledge cutoff is April 30, 2026.

Those specifications describe capacity, not proof that every answer will be better. OpenAI's own published results give a better sense of the intended territory: 97.6% on FrontierMath Tier 4, 74.1% on DeepSWE v1.1, and 72.6% on OSWorld versus 65.7% for GPT-5.6 Sol. These are OpenAI's measurements, not ours. We haven't run benchmarks on Astra. We also leave the depth to OpenAI's default rather than setting a reasoning-effort level ourselves.

The Price Is the Product Decision

OpenAI charges us $10 per million input tokens, $1 per million cached input tokens and $50 per million output tokens for Astra. If one request crosses 272,000 input tokens, the rates for that entire request become $20, $2 and $75 respectively. That last condition matters in the sort of enormous conversation Astra is built to hold; it isn't a marginal surcharge on only the tokens above the line.

GPT-5.6 Sol currently costs $4, $0.40 and $20 — a promotional rate OpenAI says holds at least through November 21, 2026. Sol is the model used by Research and the highest point the Smart Agentic router can reach, so the useful comparison is simple: at today's rates Astra costs 2.5 times as much on input and output alike. Secret Chat credits track what the provider charges us. Expect the same answer from GPT 6 Astra to cost about 2.5 times what it would from Research, though there is no honest fixed price per message. Chat length changes the bill.

There is an odd coincidence across providers. Astra's $10 input and $50 output prices match Claude Fable 5.1 exactly. Fable re-reads a cached conversation four times more cheaply, however: $0.25 per million cached tokens against Astra's $1.00. Astra fills the same role on ChatGPT that Claude Fable fills on Claude: the top model kept behind a choice you make yourself.

Smart Agentic Will Not Pick Astra

Smart Agentic ChatGPT does not route to Astra, and it will not land there by accident. The router still escalates only as far as GPT-5.6 Sol.

That is deliberate. The router runs for every message, including “make this sentence shorter” and questions that deserve a one-line reply. Adding a model priced at 2.5 times Sol to that rotation would raise the cost of ordinary chat without making ordinary chat better. Astra is for the occasions when you already know the problem justifies the dearer model. One small job stays small regardless: the automatic chat name is written by the cheap GPT tier. A title does not need a model priced at $50 per million output tokens.

How to Choose GPT 6 Astra

  1. Open the app and pick ChatGPT in the model picker.
  2. Open the mode menu next to the message box. Below Smart Agentic, Fast Chat, Web Search, Research and Create Image, the last row is GPT 6 Astra, marked with a star. Pick it.

The last position is intentional. The other rows form one ladder: more work for the answer as you move from quick chat toward research. Astra is not another rung. It is a different, dearer model, separated so that choosing it cannot feel incidental.

On Secret Chat, this mode is web-grounded. We attach OpenAI's built-in web search tool, as we do for Web Search, Research and Claude Fable, so Astra can search and read pages while it works. Prompts still pass through the same automated safety check used for every Secret Chat model. Under every response, the app prints the model and the credits spent. You will see “GPT 6 Astra” and that answer's actual price there.

Give it a problem with evidence to inspect and a decision to make, not a vague request to be clever:

Read the attached material and check current sources on the web. I need a decision, not a summary. Identify the two strongest options, test each against the constraints below, show me what evidence would change your conclusion, then recommend one. Constraints:

When Astra Earns Its Credits

Use Astra when a weak answer would cost more than the model: a difficult code investigation, a dense set of documents with facts scattered across them, or research that has to survive several competing constraints. Its long context is useful when the source material is genuinely long. It is wasteful when the chat merely is.

For email drafts, small rewrites, quick facts and ordinary back-and-forth, stay on Smart Agentic. Research is the sensible first attempt when you need current sources. If its answer fails on depth rather than missing instructions, keep the chat and retry with Astra. The price difference is then a decision based on an answer you have seen, not on a flagship label.

Astra is not on an AI Council panel, for the same reason Fable isn't. A council seats each company's mid tier because its value comes from independent companies disagreeing, not from making every single answer the strongest obtainable. If you want both kinds of scrutiny, get Astra's answer first, paste it into a council, and let the panel check it.

The Cybersecurity Limit You Should Know Before Paying

Astra is the first model OpenAI has classified as Critical for cybersecurity under its Preparedness Framework. In its own evaluations OpenAI tested the model's exploit development against vulnerabilities disclosed between June and August 2026, and reports that these included two zero-days in software it did not name. OpenAI also reports a perfect 100% on ExploitBench, against 78.5% for GPT-5.6 Sol. That result is OpenAI's own, measured without the system-level safeguards it runs in production; it is not a Secret Chat test.

The consequence is practical. The public model refuses advanced offensive-security work, while exploit-development capability is gated behind OpenAI's trusted-access programs. OpenAI has also said API users outside those programs may have security-adjacent requests slowed or stopped outright by a safety check. A legitimate security researcher can therefore spend credits, wait, and still get no substantive answer. We cannot change that boundary, and we would rather you know it before selecting the mode than discover it after.

This is not a general claim that Astra refuses security questions. Defensive work such as reviewing code and proposing patches remains part of the public use case. The boundary tightens around advanced offensive tasks and exploit development; OpenAI, not our mode menu, decides where a particular request lands.

What Private Means Here

Your prompt reaches OpenAI through our gateway, under our credentials, from our servers, so the record on OpenAI's side carries our gateway rather than your name, account or IP. Secret Chat removes you from your queries; it does not remove the data from your messages. Write your client's name into a prompt and it is still in the prompt. Your chat history lives only in your browser, and a prompt exists on our side only for as long as it takes to fetch your answer.

GPT 6 Astra is available now as the starred last row under ChatGPT. Our recommendation is narrow: start below it, and choose it when the work in front of you makes the 2.5-times price reasonable. Manual is not a limitation here. It is the point.

Frequently Asked Questions

  1. What is GPT 6 Astra on Secret Chat AI?

    GPT 6 Astra is a manually selected ChatGPT mode that runs OpenAI's real gpt-6-astra model — the most capable model OpenAI offers, web-grounded on Secret Chat, and chosen by hand rather than by the Smart Agentic router. It is live now.

  2. How do I select GPT 6 Astra?

    Pick ChatGPT in the model picker, open the mode menu beside the message box, then choose the starred last row, “GPT 6 Astra.” It appears below Smart Agentic, Fast Chat, Web Search, Research and Create Image.

  3. Will Smart Agentic ChatGPT ever select Astra automatically?

    No. Smart Agentic escalates only as far as GPT-5.6 Sol. Astra costs 2.5 times as much per token, and routing ordinary messages to it would make everyday chat dearer without making those simple answers better.

  4. How much does a GPT 6 Astra answer cost?

    There is no fixed credits-per-message figure because the cost depends on the conversation and answer length. Secret Chat credits track provider cost, so the app shows “GPT 6 Astra” and the credits spent under each reply. As a guide, the same answer costs about 2.5 times what it would from Research.

  5. Can GPT 6 Astra search the web?

    Yes. Secret Chat attaches OpenAI's built-in web search tool to this mode, allowing Astra to search and read pages while it works.

  6. What are GPT-6 Astra's context and output limits?

    OpenAI specifies a 1,050,000-token context window and up to 128,000 output tokens. The model accepts text and images and produces text, with an April 30, 2026 knowledge cutoff.

  7. Why might GPT 6 Astra refuse a cybersecurity request?

    OpenAI classifies Astra as Critical for cybersecurity and restricts its advanced offensive and exploit-development capabilities. Outside OpenAI's trusted-access programs, a security-adjacent API request may be refused, slowed or stopped by OpenAI's safety systems, even when the user's purpose is legitimate.

  8. Is GPT 6 Astra available on the AI Council?

    No. The AI Council uses each company's mid-tier model because cross-company disagreement is the useful result. You can paste an Astra answer into a council and ask the panel to check it.