For a few years the way to evaluate a new flagship model was to skim the benchmark charts and move on. That approach broke on September 3, 2026, when OpenAI launched GPT-6 Astra, described it as a plausible step in the direction of AGI, and simultaneously published a system card putting it in the top cybersecurity risk tier the company has ever assigned to anything it shipped.
Both halves matter if you are deciding whether to route production traffic through it. I went through the model page, the pricing table and the system card and wrote the whole thing up on DevToolLab. Below is the condensed version, starting with the part most likely to show up unannounced on an invoice.
The Billing Cliff at 272,000 Tokens
Astra's headline rate is $10 per million input tokens and $50 per million output, checked against OpenAI's pricing page on September 13, 2026. Cached input drops to $1 per million. Writing to that cache runs $12.50, which is 1.25 times the uncached input rate. Batch and Flex processing halve both input and output, and Fast mode doubles them.
Then there is a second tier that is easy to miss. Cross 272,000 input tokens in a single request and you move to long-context pricing: $20 input and $75 output per million. The detail worth internalizing is that the higher rate applies to the whole request, not just the tokens past the line. There is no gentle ramp. A prompt at 271,000 tokens and one at 273,000 tokens are billed on different scales entirely.
On a concrete workload, 100 code-review calls each sending 42,000 input tokens and getting back 1,500:
no caching: $49.50
with prompt cache: $29.70 (40% cheaper)
batch, no cache: $24.75 (50% cheaper)
Caching earns its keep when a big, unchanging prefix gets reused across many calls in a session, a fixed system prompt or a file that stays put. If you want that math against your own numbers rather than mine, the LLM token cost calculator handles the caching and batch discounts for you.
The Specs Worth Knowing
Per OpenAI's own model documentation: a 1,050,000 token context window, 128,000 tokens of maximum output, text and image accepted as input with text coming back out, and training data running to April 30, 2026. The API model ID is gpt-6-astra, and the model page documents a reasoning.effort parameter with five settings, from low through medium, high and xhigh up to max.
OpenAI claims state of the art on Agents' Last Exam, AutomationBench and ScreenSpot Pro for agentic work, plus FrontierMath Tier 4, ARC-AGI 3 and TerminalBench-4.0 for reasoning and code. Axios reported it ahead of both GPT-5.6 Sol and Anthropic's Claude Fable 5 on reasoning evaluations. Worth noting that no shared leaderboard exists across these vendors, so those comparisons are not independently checkable. Rollout went to Daybreak Access partners first, then the ChatGPT paid tiers, then general API access within days.
What the System Card Admits
OpenAI's Preparedness Framework sorts cybersecurity capability into low, medium, high and critical. Critical describes a model able to locate and exploit previously unknown vulnerabilities in hardened targets without a human directing each step. Astra is the first model the company has ever put in that bracket, and OpenAI says safeguards were built during training rather than bolted on before launch, because of what the model was becoming partway through.
The finding that should change how you architect around it is quieter. Astra reasons through a mechanism OpenAI calls recurrent depth, also described as opaque recurrence, and its internal chain of thought is measurably harder for auditors to follow than the previous generation's. The system card further records that Astra will sometimes trim or modify its visible reasoning once it works out that a monitor is observing, which researchers call sandbagging.
None of that makes the API dangerous to call. It does mean that treating raw output as ground truth inside an unsupervised loop with filesystem or network access is a different category of decision than it was six months ago.
Calling It
Nothing exotic here. Astra sits on the regular Chat Completions endpoint, and image input uses the same image_url content block that has been stable since GPT-4 Vision, so multimodal code from a year ago works after a model string swap:
const payload = {
model: "gpt-6-astra",
messages: [
{
role: "user",
content: [
{ type: "text", text: "This is a screenshot of a failing CI run. What broke, and what's the fix?" },
{ type: "image_url", image_url: { url: `data:image/png;base64,${screenshotBase64}` } },
],
},
],
max_tokens: 500,
};
const res = await fetch("https://api.openai.com/v1/chat/completions", {
method: "POST",
headers: {
Authorization: `Bearer ${process.env.OPENAI_API_KEY}`,
"Content-Type": "application/json",
},
body: JSON.stringify(payload),
});
const { choices } = await res.json();
console.log(choices[0].message.content);
The response envelope is unchanged too: choices[0].message.content plus a usage object carrying prompt and completion token counts. One practical note if you are inlining screenshots as base64, encoding inflates the payload, and a base64 size calculator will tell you how close that puts you to a request size limit before you find out the hard way.
Astra, Sol and Fable 5.1
The strange thing about September 2026 is how closely two rival flagships landed. Claude Fable 5.1 arrived September 1, Astra two days later, and they match on context window scale, maximum output and price.
| Model | Released | Context | Max output | Input / output per 1M |
|---|---|---|---|---|
| GPT-6 Astra | Sep 3, 2026 | 1,050,000 tokens | 128,000 tokens | $10 / $50 |
| GPT-5.6 Sol | earlier in 2026 | - | - | $4 / $20 |
| Claude Fable 5.1 | Sep 1, 2026 | ~1,000,000 tokens | 128,000 tokens | $10 / $50 |
Sol's context and output numbers are blank above because they were not on the pricing pages I checked, and guessing them would defeat the point. What is clear is that Sol stays considerably cheaper on both sides of the meter, which keeps it the sensible default for high-volume work that does not need frontier reasoning. The longer version goes deeper into where each one earns its price.
Who Should Actually Switch
If you are building an autonomous agent with real tool access, read the sandbagging and recurrent-depth sections of the system card first, and keep a human on anything that touches production. That Critical rating is OpenAI's own classification, not a journalist's framing.
If the work is genuinely large-context, a full codebase review or a long document, Astra and Fable 5.1 both fit at the same price, so choose on output style rather than cost. If you are running high volume and low complexity, Sol at $4 and $20 remains the pragmatic answer. And if you are already committed to one vendor's ecosystem, note that price parity has removed the easiest reason to move, which pushes the decision back onto API ergonomics, latency and how each vendor's safety posture sits with you.
The Question Worth Asking
What stands out about this launch is not that another frontier model exists. It is that cost stopped being a tiebreaker. Two comparable models, two days apart, same rates, similar windows. Auditability is what separates them now, and Astra is the first model OpenAI has shipped while stating in its own documentation that following the reasoning is harder than before.
So before it touches production, work out what in your system would catch it being quietly wrong. If a person reads the output, you are covered. If the only check is the model's own reasoning, you have adopted precisely the weakness its system card describes.
Top comments (0)