# GPT-6 Astra: OpenAI's most capable model, its benchmarks and its price

> GPT-6 Astra from OpenAI: 1.05M-token context, $10 and $50 per million tokens, benchmarks from its system card and when to pick it. Data as of Nov 15, 2026.

*Tags: ai, models, openai, agents*

---


> [!info] Quick Definition
> **GPT-6 Astra** (`gpt-6-astra`) is OpenAI's most capable model as of November 15, 2026, according to its own documentation. Its system card was published on September 3, 2026. It is a closed model, available through the API, with a 1.05-million-token context window, presented by OpenAI for "the most demanding work".


The data on this page is as of November 15, 2026, and it changes often. Before you decide, confirm prices and limits on the pages listed under Sources.

## What is GPT-6 Astra and when did it come out?

OpenAI's documentation describes it as "our most capable model for the most demanding work." It sits at the top of the GPT-6 family: below it are cheaper variants (gpt-6-luna, gpt-6-sol and gpt-6.1-terra, with prices in the cost section).

The model page gives no launch date. What it does show is that the system card, the document where OpenAI explains how it evaluated the model, is listed as published on September 3, 2026 and opens with "Today, we are releasing GPT-6 Astra". That is the reference date.

## What is it for?

OpenAI presents it for "the most demanding work". For someone designing products, these are possible uses (ideas from this wiki, not an OpenAI list):

- **Multi-step work:** an agent that researches, writes a document and reviews it. To understand how those tasks are organized, see [[reasoning-models|reasoning models]].
- **Coding and prototypes:** complex code when a small model falls short.
- **Long documents and data:** its 1.05-million-token [[context-window|context window]] fits a lot of material in a single request (the listed maximum input is 922,000 tokens, because 128,000 are reserved for output).
- **Images as input:** you can pass a screenshot or a diagram and ask for an analysis. Output is text only. More in [[multimodal-models|multimodal models]].

Features listed in the documentation: streaming, structured outputs, function calling, file search, web search and prompt caching. The reasoning parameter accepts five levels: low, medium, high, xhigh and max.

## What is the quick spec sheet?

| Item | Value (as of Nov 15, 2026) |
|---|---|
| API ID | `gpt-6-astra` |
| Lab | OpenAI |
| Reference date | September 3, 2026 (system card) |
| Context | 1,050,000 tokens (max input 922,000; max output 128,000) |
| Knowledge cutoff | April 30, 2026 |
| Modalities | Input: text and image. Output: text |
| Standard price | $10 input / $50 output per million tokens |
| Open or closed | Closed (API access) |

## What benchmarks did it publish and how do you read them?

The numbers here come from OpenAI's system card, which reports them. They are not a basis for ranking against models from other labs, which use other tests and other settings. To learn how to read them, see [[ai-benchmarks|how to read AI benchmarks]].

| Benchmark | Astra score | Reported by |
|---|---|---|
| HealthBench Professional (length-adjusted) | 64.7 | OpenAI, in its system card |
| HealthBench Hard (length-adjusted) | 36.6 | OpenAI, in its system card |
| Gray Swan IPI Arena | 8.5% estimated attack success (GPT-5.6 Sol: 27.0%) | OpenAI, in its system card |
| ExploitBench | 100%, even at the lowest reasoning effort tested | OpenAI, in its system card |

How to read them, based on what the system card itself says:

- **HealthBench Professional** is, in the card's words, "a public evaluation of model capability and safety for clinician use cases". A high score means better performance on those cases, not that the model can replace a doctor.
- **Gray Swan IPI Arena** measures "whether instructions embedded in untrusted content can redirect an agent toward harmful actions". Lower is better here: 8.5% means that share of attacks managed to redirect the agent. It is an estimate based on 15 attempts per scenario.
- **HealthBench Hard** and **ExploitBench** are the other two scores in the table: OpenAI reports them without a short definition in the text, so treat them as lab data, not universal measures.

The system card also places Astra in OpenAI's Preparedness Framework: Critical in cybersecurity, High in biological and chemical, and below High in AI self-improvement. It is a risk assessment by the lab itself, not a quality measure.

## How much does it cost and how much context does it have?

According to OpenAI's pricing page, as of November 15, 2026, per million tokens:

| Mode | Input | Cached input | Output |
|---|---|---|---|
| Standard (up to 272K input) | 10 | 1 | 50 |
| Standard (over 272K input) | 20 | 2 | 75 |
| Batch or Flex (up to 272K) | 5 | 0.50 | 25 |
| Fast (up to 272K) | 20 | 2 | 100 |

Rules that change the bill:

- **Long prompts:** past 272,000 input tokens, input costs double and output costs 1.5 times, on the whole request, not just the excess.
- **Batch and Flex** cost half the standard rate. **Fast** (formerly Priority) costs double.
- **Cache:** repeating the same prompt prefix brings input down to $1 per million.

To understand how that bill is built, read [[cost-per-token|cost per token]] and [[tokens|tokens]]. The cheaper options in the family, at the standard rate: gpt-6-luna at 0.10 and 0.50; gpt-6-sol at 2 and 10; gpt-6.1-terra at 1 and 5.

## When should you pick it and when not?

**Pick it when:**

- The task is hard, long, and a mistake is expensive (reviewing a long document, an agent with many steps, complex code).
- You need to process a huge amount of material at once.
- You want a top-quality reference point before deciding whether a cheaper model is enough.

**Skip it when:**

- The task is repetitive and simple (classifying, extracting fields, summarizing short texts). A [[small-vs-large-models|small model]] costs a fraction there.
- You need open weights or want to run the model on your own infrastructure. See [[open-vs-closed-models|open vs. closed models]].
- Your budget cannot carry $50 per million output tokens at scale.

If you want the product experience for daily work, see [[chatgpt-for-designers|ChatGPT for designers]]. This page covers the model, not the product. To see another top model from another lab, there is [[gemini-4-carbon|Gemini 4 Carbon]], though its benchmarks come from another lab and are not comparable with OpenAI's.

## Example

Illustrative example: a product designer has 40 interview transcripts (about 400,000 tokens) and needs a synthesis with quotes and themes. She does a first pass with a cheap model to tag each interview, and with those tags she uses Astra once for the final synthesis. With the prompt under 272,000 tokens, input costs $10 per million: 200,000 input tokens and 5,000 output tokens come to 2.00 + 0.25 = $2.25, not counting the cache.

What to try tomorrow: take a task of yours that today takes several hours, run it at medium and at high reasoning levels, and compare quality and cost. Record the tokens each run used before deciding which to keep.

## Common mistakes

- Assuming Astra is better than any other model because OpenAI calls it the most capable: that is OpenAI's claim about its own lineup.
- Comparing its scores with another lab's as if they were the same test.
- Forgetting the surcharge for prompts over 272,000 tokens: it doubles input for the whole request.
- Using it for simple tasks that a cheaper model in the family can handle.
- Confusing the Critical cybersecurity level with a quality score: it is a risk classification.
- Assuming it is available in a specific product (for example ChatGPT): the API documentation lists it, and each product's own help pages need checking.

## Sources

- [OpenAI: GPT-6 Astra model page](https://developers.openai.com/api/docs/models/gpt-6-astra)
- [OpenAI: API pricing](https://developers.openai.com/api/docs/pricing)
- [OpenAI: GPT-6 Astra system card (Deployment Safety)](https://deploymentsafety.openai.com/gpt-6-astra)


---

Source: https://www.fernandoux.com/en/wiki/ai/gpt-6-astra/
