If you search for "gpt 6 astra", you will find three different products with the same name. I summarise what the model is, what it costs and where the vendor says it is strong. Mainly, though, I describe my own test: in my own unfinished game Noční linka (Night Line) I used Astra as an independent critic of 3D models and as a fixer of animations in Blender, including the places where it failed.
GPT-6 Astra is an OpenAI model with the API ID gpt-6-astra: text and image input, text output, a 1.05 million token context. According to the press it was introduced on 3 September 2026. In the API it costs $10 per million input tokens and $50 per million output tokens (verified 20 October 2026), the same as Claude Fable 5.1 and 2.5 times as much as Opus 5.5 ($4 / $20). It is not Google Project Astra or Topaz Astra. In my game it served as an independent critic of models and fixed the mop, punch and pole-holding animations, but the coat did not pass after three rounds. I did not run a head-to-head test against Opus 5.5.
I am making my own game, Noční linka (alpha, unreleased), and I used GPT-6 Astra in it in two roles:
- An independent audit of the 3D models written by Opus 5.5 (10 to 11 October 2026): on the bus it praised the readable silhouette, frames, handrails and seats, and on the village houses it criticised bare sides, flat doors and roofs that are too regular.
- Animation fixes in Blender (11 October 2026, Blender 4.1.1 without a window): the mop and the pole grip on the bus passed in two rounds, with a palm anchor deviation of at most 4.54 mm for the mop and 2.19 mm for the pole (measured on bones, not on the full mesh). The coat for a female character ended as HOLD after three rounds and was not enabled in the game.
- What I do not have: a direct comparison with Opus 5.5 on the same task.
What is GPT-6 Astra and what is it not?
GPT-6 Astra is an OpenAI language model that the vendor presents as its most capable for the most demanding work. It accepts text and images as input and returns text. According to the press, OpenAI introduced it on 3 September 2026 and is rolling it out in ChatGPT (Plus, Pro, Business, Enterprise), in the API and through Azure and Bedrock. The official announcement was not available, so the date comes from the press.
The name Astra is also used by other products:
| Name | What it is | Who is behind it |
|---|---|---|
| GPT-6 Astra | A language model with image input, in ChatGPT and the API | OpenAI |
| Project Astra | A research prototype of a universal assistant, its features are gradually reaching Gemini Live | Google DeepMind |
| Topaz Astra | A tool for upscaling and enhancing video | Topaz Labs |
Czech search suggestions for "astra ai" also mix in third-party apps, so check which Astra any review is about. This article is about the OpenAI model.
How much does GPT-6 Astra cost?
In the API you pay per token. Standard rates per the OpenAI model page and the Anthropic overview, verified 20 October 2026:
| Model | Input per 1 M tokens | Output per 1 M tokens | Context |
|---|---|---|---|
GPT-6 Astra |
$10 |
$50 |
1.05 M |
Claude Fable 5.1 |
$10 |
$50 |
1 M |
Claude Opus 5.5 |
$4 |
$20 |
1 M |
OpenAI also lists cached input at $1, higher rates for inputs over 272 thousand tokens and half price for batch processing. A model calculation, not a measurement: a request with 10,000 input and 2,000 billed output tokens comes to $0.20 on Astra and $0.08 on Opus 5.5. Reasoning tokens count as output.
In ChatGPT you pay a subscription, not per token; check the current plans and any free access with OpenAI. For choosing a model by task and price see ChatGPT, Claude, Gemini and Kimi, and for Opus 5.5 see the separate article.
Where do the vendor and reviews say Astra is strong?
The OpenAI model page lists tools including computer use and MCP, and reasoning effort from low to max. The press summarises early testers' impressions: strong in agentic and spatial work and in computer use, weaker in writing. These are not independent measurements. Two OpenAI case studies are more relevant for 3D and games:
- Architectural visualisation. Astra built an editable scene through Blender's Python API, inspected its own previews and fixed defects, for example the shading of a sink. The author notes that geometry checks are not a certification of every detail and that the floor plan would need professional review before construction.
- Game development. The workflow starts with image references and player experience requirements, and only then comes code and continuous checking with tests. The author adds that playtesting in the browser is still needed.
What carries over is the order of work and the checking of results. I did not test writing with Astra, so I can neither confirm nor refute that it is weaker.
How did I test Astra in the game Noční linka?
Noční linka (Night Line) is my own unfinished game, not a client project, in alpha and unreleased. You drive a night bus on route 913 through the fictional Bořín district in 1990s Czechia, it runs on the Godot 4.7.1 engine and third-party assets are CC0 only. The 3D models of the world (panel blocks, village, train, cars, bus, mill and more) are Python scripts for Blender written by Claude Opus 5.5. Blender runs without MCP, only as command-line scripts. Astra was given two other roles.
Audit of the models by Opus
On 10 and 11 October 2026 Astra assessed renders of these models and fixed nothing:
- Bus: it praised the readable silhouette, frames, handrails and seats.
- Village houses: it criticised bare sides and backs, flat doors and roofs that are too regular. Up close they look like a kit (see the picture above).
- Panel blocks: entrances, loggias and aerials work, but the dirt map repeats.
- Train and mill: the same blotches on the train roof, repeated outlines of damaged plaster on the mill.
The criticism mostly targets repetition that does not matter from afar but shows up close. For models written as scripts that is useful, because patterns are easy to copy. The audit did not recommend replacing the models wholesale. Its biggest finding was in hands, contacts and character animations. It worked with renders and did not verify motion or performance.
Animation fixes in Blender
On 11 October 2026 a batch of fixes followed. Astra, working as an agent, ran Blender 4.1.1 without a window with Cycles rendering on the CPU. It had at most three rounds of previews per item, and every preview had to be actually inspected (35 of 35 opened).
| Item | Rounds | Result | Measurement |
|---|---|---|---|
Mop |
2 |
accepted |
palm anchor deviation at most 4.54 mm across 73 frames |
Punches (jab, cross, hooks) |
2 |
accepted |
sideways wrist travel at most 0.252 m instead of 0.376 m, return to guard |
Holding the bus pole |
2 |
accepted |
heights 1.60 / 1.75 / 1.90 m, anchor deviation at most 2.19 mm |
Coat for a female character |
3 |
HOLD |
gap at the waist, elbows clipping |
The mop originally had a grip problem. After the fix both hands hold one solid shaft:
The audit rated holding the bus pole 4 out of 5 (needs a fix): open palms and no documented grip. After the fix, characters of three heights hold the pole. The pole in the bus model is 40 mm in diameter, while the clip assumed a 35 mm reference diameter:
An important limit: contacts were measured on the character's bones (a proxy), not with a full mesh intersection test. "Accepted" means an accepted source and previews, not finished gameplay: the animations are not yet wired into the game and the in-game check of passengers in Godot remains.
What did not work
The coat for the female character got real sleeves, a longer hem, a collar and buttons, but after three rounds a gap at the waist remained and the elbows passed through the body. An independent critic confirmed the HOLD verdict and the coat was not enabled in the game:
The follow-up batch of animations, V4 (nausea, swaying, provocation, calming hands), ended up all HOLD after the first round. A hand was at the temple instead of in front of the mouth, an index finger stayed curled and the fingers were too clenched. A record of the next phase from the same evening says that after further rounds the inspected previews passed (nausea and provocation in the third round, calming hands in the second, swaying after a check of the samples) together with loading the assets in headless Godot. Wiring into the game and contact on the characters, however, remain HOLD, and the coat did not change.
What does this mean in practice?
- Author and independent critic. The critic judges real renders, not the author's claims. That it is a different model, I take as a bonus, I have not measured it. It ties into the agent harness.
- Measurable acceptance criteria. "Holds the pole" is not enough. A deviation in millimetres, character heights and a frame count give the critic something to compare against.
- A cap on rounds. Three rounds per item and a HOLD verdict are a legitimate outcome. The coat showed it: better a HOLD than an accepted defect.
- Say what was not measured. Bone proxies, no in-game test, stills instead of motion.
In the game these principles fit into eight production gates, from reference and scale check to independent critique.
Is GPT-6 Astra better than Claude Opus 5.5?
I cannot say that from my test. A head-to-head run of both models on the same task did not happen: Opus wrote the 3D models, Astra audited them and fixed animations, which are different jobs. The results in Anthropic's table are vendor-reported, not independent, and differ in effort settings:
| Test | Opus 5.5 | GPT-6 Astra |
|---|---|---|
| Terminal-Bench 4.0 | 66.4% (xhigh) | 57.9% (high) |
| FrontierCode v1.1 | 54.4% | 53.3% |
| AutomationBench | 40.0% | 41.4% |
| Terminal-Bench-Science 0.1 | 58.7% | 64.6% |
Astra leads in two rows, Opus in two. The same table gives Astra lower values elsewhere, for example Humanity's Last Exam (57.2% versus 67.7%, both with tools). For your own decision, compare the fixes needed and the cost on the same set of your own tasks, not just a chart.
If you want to bring a similar discipline (an author, an independent check, measurable criteria) into company processes, AI automation can help: it starts with an audit and continues with one live process measured before and after.
FAQ
How much does GPT-6 Astra cost?
In the API $10 per million input tokens and $50 per million output tokens (verified 20 October 2026), the same as Claude Fable 5.1 and 2.5 times as much as Claude Opus 5.5 ($4 / $20). Batch processing costs half. In ChatGPT it is a subscription, not per token; check the plans with OpenAI.
What is GPT-6 Astra?
An OpenAI language model with the API ID gpt-6-astra: text and image input, text output, a 1.05 million token context, tool support including computer use. According to the press it was introduced on 3 September 2026, and reviews call it strong in agentic and spatial work, weaker in writing.
Is GPT-6 Astra better than Claude Opus 5.5?
It cannot be said in general. Anthropic's table is vendor-reported and results split by test. I did not compare the two models on the same task.
Can Astra create a 3D model?
Not directly, the model's output is text. It can create a 3D scene through code, for example a Python script for Blender, as OpenAI's case studies show. In my test the coat for a female character, a new clothing model, was part of its batch of fixes. It did not pass after three rounds (HOLD), and the game's world models were written by Opus 5.5.
Is Astra the same as Google Project Astra?
No. Project Astra is a research prototype of a universal assistant from Google DeepMind, and its features arrive through Gemini Live. GPT-6 Astra is an OpenAI language model. Topaz Astra is yet another product, a tool for upscaling video.