GPT-6 Astra: what the independent benchmarks actually show
Every figure below is traceable: each comes from a named, dated public source or from Scalarware's own live catalogue, listed in full at the end.
TL;DR
- GPT-6 Astra launched 3 September 2026 with OpenAI willing to attach the phrase "AGI era" to it.
- Independent score: 61 on the Artificial Analysis Intelligence Index v4.1.1 at highest effort — the same as GPT-5.6 Sol, the model it succeeds.
- List price is $10 / $50 per million tokens against Sol's $4 / $20 — a 2.5x increase for an unchanged composite score.
- The claimed advance is operating software through screens rather than APIs. No general-intelligence composite isolates that, so a flat score is weak evidence about the thing Astra is for.
- It is the first OpenAI model rated Critical for cybersecurity, so access is enterprise-first through the Daybreak programme — you cannot route production traffic to it yet, on Scalarware or anywhere reselling the public API.
OpenAI released GPT-6 Astra on 3 September 2026 at $10 / $50 per million input / output tokens. On the independent Artificial Analysis Intelligence Index (v4.1.1) it scores 61 at highest effort — identical to GPT-5.6 Sol, the model it replaces. The advance it claims is computer use, which that index does not measure.
The headline and the scorecard disagree
OpenAI's framing was unusually strong: Astra is described as setting the state of the art for computer use, browsing, software engineering, cybersecurity, science and professional work, and the company was willing to describe the moment as the "AGI era". Independent measurement lands somewhere flatter.
| Model | AA Intelligence Index v4.1.1 | List price / 1M in–out |
|---|---|---|
| Claude Fable 5.1 (max) | 66 | $10 / $50 |
| Claude Opus 5 (max) | 63 | — |
| GPT-6 Astra (highest effort) | 61 | $10 / $50 |
| GPT-5.6 Sol (max) | 61 | $4 / $20 |
Index scores as published by Artificial Analysis and reported on 3 September 2026. A composite index is one measurement, not a verdict — and Astra is explicitly built for a capability this index does not isolate.
The pricing reversal is the real story
Astra lists at 2.5x its predecessor — $10 / $50 against Sol's $4 / $20 — while scoring identically on general intelligence. That inverts the pattern the market has been trained on, where each generation delivered more capability at equal or lower cost. The premium buys computer use, not headroom on the work most teams already run.
For comparison, on the Scalarware catalogue today openai/gpt-5.6-sol is $2.00 / $10.00 per million tokens at 0% markup. If your workload is chat, retrieval or ordinary tool-calling, the honest read is that Astra offers nothing over Sol on the axis being measured, at several times the bill.
What Astra is actually built for
The interesting claim is not the score, it is the interface. Astra is described as operating software through screens rather than through APIs — driving applications the way a person does rather than calling endpoints. That is a different product from a faster chat model, and it is not what a general-intelligence composite measures.
This matters for how to read the 61. It is strong evidence that Astra does not advance general reasoning over Sol. It is weak evidence about GUI-driving agents, because no public composite currently scores that well. Anyone claiming the benchmark "disproves" the launch is over-reading it; anyone citing the launch as proof of a capability jump on ordinary work is over-reading it in the other direction.
Availability is the near-term constraint, not price
Astra is the first OpenAI model rated Critical under the company's cybersecurity preparedness framework. Access begins with enterprise customers through the Daybreak programme, with paid ChatGPT plans and the API following. Practically, that means you cannot route production traffic to it yet — not through the public API, and therefore not through any platform that resells it.
This is why openai/gpt-6-astra does not appear on the Scalarware catalogue. When the public API opens, it becomes a model id like any other.
What to run instead, today
Everything below is live on Scalarware at 0% markup on language-model inference, behind one OpenAI-compatible endpoint and one balance.
| If you want | Model id | Input / Output per 1M |
|---|---|---|
| Top measured intelligence | anthropic/claude-fable-5.1 | $10.00 / $50.00 |
| Strong general work, half the price | anthropic/claude-opus-5 | $5.00 / $25.00 |
| Astra's score, a fraction of the cost | openai/gpt-5.6-sol | $2.00 / $10.00 |
| Cheapest frontier-class option | meta/muse-spark-1.3 | $1.25 / $4.25 |
Calling any of them
curl https://scalarware.com/api/v1/chat/completions \
-H "Authorization: Bearer $SCALARWARE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "openai/gpt-5.6-sol",
"messages": [{"role": "user", "content": "Explain this stack trace."}]
}'Detecting when Astra becomes available
Rather than watching announcements, poll the catalogue — it lists exactly what is routable right now, so an agent can decide for itself when to switch:
import os, requests
catalog = requests.get("https://scalarware.com/api/v1/models").json()
ids = {model["id"] for model in catalog["data"]}
preferred = ["openai/gpt-6-astra", "anthropic/claude-fable-5.1", "openai/gpt-5.6-sol"]
model = next(candidate for candidate in preferred if candidate in ids)
print("routing to", model)The same endpoint returns live pricing and context windows per model, so a cost ceiling can be enforced in code instead of assumed.
How to read a launch like this
- Separate the claim from the measurement. "State of the art for computer use" and "61 on a general-intelligence composite" are both true and are about different things.
- Check availability before architecture. A model gated to an enterprise programme is not a dependency you can build on this quarter.
- Price the alternative you already have. Sol scores the same on this index at a fraction of the cost; that comparison is available to anyone, and it is the one that shows up on the invoice.
- Wait for independent evaluation on the axis that matters. Computer-use benchmarks are immature. When they arrive, they will say more about Astra than any composite does.
Astra is days old and independent evaluation is thin. Treat this as the published evidence as of 4 September 2026, not a settled ranking.
Frequently asked questions
- Can I use GPT-6 Astra through an API today?
- Not generally. Astra is the first OpenAI model rated Critical under the company's cybersecurity preparedness framework, so access starts with enterprise customers through the Daybreak programme, with paid ChatGPT plans and the API to follow.
- How much does GPT-6 Astra cost?
- It is listed at $10 per million input tokens and $50 per million output tokens — 2.5x the $4 / $20 of GPT-5.6 Sol, the model it replaces.
- Is GPT-6 Astra better than GPT-5.6 Sol?
- Not on general intelligence. Both score 61 on the Artificial Analysis Intelligence Index v4.1.1. Astra's claimed advance is computer use — operating software through screens rather than APIs — which that index does not isolate.
- What should I use instead while Astra is gated?
- On measured intelligence, anthropic/claude-fable-5.1 scores 66. For the same score as Astra at a fraction of the cost, openai/gpt-5.6-sol is $2.00 / $10.00 per million tokens on Scalarware. For the lowest cost, meta/muse-spark-1.3 is $1.25 / $4.25.