Five Days of Receipts
- Finished Portal on its own. 3,336 tool calls, roughly 21 hours, a $571.18 token bill.
- Launched a rocket in Factorio Space Age 2.1. No model had ever pushed past blue science before.
- Beat Pokรฉmon in 18h 12m. GPT-5.6 Sol needed 96h 35m for the same run.
- Scored 19 out of 20 on a robot arm, at $0.94 per attempt.
- Produced 3,295 editable objects in Blender from a single prompt.
- What ties them together: every one of these has a price tag, and it is not small.
The launch write-up covered the benchmark table, the pricing and the access rules. Five days on, the picture has changed: instead of OpenAI's slides, we can now look at what people actually got the model to do.
All twelve entries below are checkable. Each one has a video, a live link or a measurement report, so you can open them yourself. I have kept the numbers in, because "an AI finished a video game" and "an AI finished a video game for $571" are not the same sentence. ๐๐ป
1. It Finished Portal Alone: 3,336 Moves, $571 ๐ฎ
This is the headline run. A developer going by cozyblaze wired Astra into Valve's 2007 puzzle game Portal and let it play the whole thing through without a single human input.
The detail that matters: the model never reached into the game's code. Astra looks at screenshots and emits keyboard and mouse commands the way a person would. Where the portal gun fires is a decision made from the frame it just saw.
The tally: 3,336 tool calls, about 21 hours of thinking time, and a $571.18 token bill. The whole run was recorded:
- ๐ฆ The developer's post: @cozyblazex
- ๐ฐ Coverage: Tom's Hardware
2. It Launched a Rocket in Factorio ๐
Portal is a puzzle game. Factorio is a many-hour production-chain marathon, and that is the harder one for a model, because it demands a coherent plan held across dozens of hours.
Someone connected Codex, running Astra's low effort tier, to Factorio Space Age 2.1 through an MCP server written in Lua that drives the game. The result: a rocket launched into space after roughly 10 hours, with the agent still pushing toward the next planet past the 20-hour mark.
A comment in the thread frames why this lands: until now, the ceiling for any model was blue science, step two or three of a tech tree that runs about ten levels deep.
- ๐ฌ Thread and recording: Hacker News
3. It Beat Pokรฉmon in 18 Hours 12 Minutes
Clad3815 ran the same harness across three models, and lined up side by side the pace of progress is hard to miss:
| Model | Time to Champion |
|---|---|
| GPT-6 Astra (high) | 18h 12m |
| GPT-5.6 Sol (max) | 96h 35m |
| GPT-5.5 | unfinished after 218h |
No RAM reads, no walkthrough, no human hints here either. The model works from screenshots and tracks its own position in the game.
4. 99.9% on ARC-AGI-3, at a Cost of $19,000
The ARC Prize team published its own independent evaluation, and two rows in it are the most honest summary of the entire launch:
| Setup | Score | Cost |
|---|---|---|
| Standard harness | 62.7% | $26,098 |
| Provider adapter harness | 99.9% | $19,000 |
Same model, same benchmark. What produces those 37 points is not intelligence, it is the scaffolding built around the model. The most striking line in the report: with the adapter, Astra used fewer actions than the human baseline on 96% of levels, and 51.7% fewer on average. Human participants, for reference, were paid about $12.78 per attempted game.
The team still puts a fence around it: saturating this benchmark is not proof of AGI, because the environment is closed-ended and deterministic.
- ๐ Report: arcprize.org/blog/astra
5. It Drove a Robot Arm: 19 out of 20 ๐ฆพ
This is the one example that leaves the screen. Astra was connected to a bimanual YAM robot arm with six degrees of freedom per arm, fed by three camera views plus proprioceptive state.
| Task | GPT-6 Astra | Claude Fable 5.1 |
|---|---|---|
| Put the red block in the bowl | 19/20 (95%) | 40% |
| Insert the round piece into its groove | 2/20 | 2/20 |
On the simple grasp the gap is enormous; on precise insertion both models stall in exactly the same place, getting the piece over the groove and failing the final push. Each attempt took 2.5 minutes and cost $0.94.
- ๐ฌ Test report: openai.robocurve.org
6. 3,295 Objects in Blender, in One Pass
Astra does not generate 3D assets. It operates Blender: plans the scene, writes Python, renders frames, looks at the result and fixes it.
Tom Krcha handed it an old steam locomotive drawing and got 3,295 fully editable objects back in a few minutes:
The same developer rebuilt a house in 3D from photos in under 30 minutes:
Sharif Shameem had it model San Francisco's Palace of Fine Arts, at a level of detail that survives comparison with reference photos:
And one you can walk around in a browser: Peter Gostev turned the town in a Van Gogh painting into a navigable scene. Live link
7. Playable Games From a Single Prompt ๐น๏ธ
Pair Astra with Sites in ChatGPT and one prompt turns into a published game. All of these open in a browser, nothing to install:
- Zork in 3D (Ethan Mollick)
- Gogh Strike, a shooter set inside Van Gogh paintings (Peter Gostev)
- Jelly Baby Playground, a soft-body physics toy (Scott)
- Universe Duel
These and a great many more, with creator credits attached, are collected in awesome-gpt-6-astra, which currently lists 117 cases and 43 live links. Consider that a warning about your afternoon.
8. On the Web: a 2,234-Piece Anatomy Atlas
Outside games the density is similar:
- ashe built an exploded-view human anatomy atlas made of 2,234 separate pieces.
- Max Weinbach shipped a browser-based macOS 27 simulator in 75 minutes.
- Ethan Mollick published ABYSSAL, a live coral reef ecosystem simulation.
- Derya Unutmaz had it build an interactive piano that plays all six Brandenburg concertos.
9. In Mathematics: the First Gain Since the 1930s
Setting the fun aside, here is the most serious result of the week. Mathematician Mehtaab Sawhney reported that, with Astra's help, he improved a bound on the longest gap between consecutive primes, by roughly a log log n factor. That bound had not moved since the 1930s.
Astra also solved two problems from Epoch's curated set of 68 unsolved Erdลs problems. Two sounds modest until you see the rest of the field: GPT-5.6 Sol, GPT-5.5, Claude Fable 5.1 and Fable 5 produced zero verified solutions on the same set.
The asterisk here
The proof artifact for the prime-gap result is roughly 10 MB of Lean. The Lean compiler checks it, but no independent human expert has reviewed it semantically yet. Read "the model proved a theorem" with that footnote attached.
10. Measured in Code Review: 22% More Bugs Caught
CodeRabbit ran Astra through its own review pipeline against a labelled bug set. The gains are modest but they come from real work:
| Comparison | Extra bugs caught |
|---|---|
| vs GPT-5.6 Sol | 4% |
| vs Opus 5 | 22% |
| Complex cross-file reviews (vs Sol) | 20% |
| Complex cross-file reviews (vs Opus 5) | 33% |
The price side stings: holding token use fixed at 100K in and 10K out, a task costs $1.50 against Sol's $0.60, a 2.5x jump.
- ๐ Evaluation: CodeRabbit blog
11. Token Efficiency: Same Score, a Third of the Tokens
Artificial Analysis's independent measurement moves the pricing argument somewhere else. Astra scores 67 on the Coding Agent Index, level with Claude Opus 5 and Fable 5. But it burns 70% fewer tokens than Sol getting there, running at max effort on a third of what its predecessor consumed.
Another number worth keeping: on AA-Omniscience the hallucination rate falls from 92% to 51%. On the general Intelligence Index it sits at 61, tied with Sol and five points behind Fable 5.1.
So Astra is not a better model at everything. It is a model specialised toward coding agents and toward making things up less often.
- ๐ Measurement: Artificial Analysis
12. On the Company Side: Playco and Code Arena
Game studio Playco reported that moving its prototyping flow onto Astra cut manual fixes in half. Astra also took the top spot on Code Arena during launch week.
Three Things This List Tells You
1. The scaffolding matters as much as the model. 62.7% and 99.9% on ARC-AGI-3 are the same model. The harness produces the difference. Likewise, what made the Factorio run possible was an MCP server written in Lua. The work is not the model, it is everything around it.
2. The bill is now the binding constraint. $571 for Portal, $19,000 for the ARC run, $1.50 per code review task. What Astra can do is impressive; the line between "can" and "worth doing" is drawn by token cost.
3. Precision is still the wall. On the robot arm, getting the piece to the mouth of the groove is easy and pressing the last millimetre is impossible. The same pattern shows up in software: the model carries 95% of the job and the remaining 5% stays with a person.
How to Try These Yourself
Most of the runs above happened in ChatGPT's Work and Codex experiences or straight through the API. Where the model shows up on each plan, what the message limits are and how to set up Codex are covered step by step in the how-to-use guide.
Before you start experimenting, do the arithmetic. At $10 in and $50 out per million tokens, a long-running agent task grows faster than you expect. The LLM cost calculator puts Astra next to the other models so you can price a task before you run it.
Frequently Asked Questions
Q: Did GPT-6 Astra really finish Portal on its own?
A: Yes. In the run by the developer cozyblaze, the model completed the game with no human input and no access to the game's code; it worked purely from screenshots and issued keyboard and mouse commands. The run took 3,336 tool calls and the token bill came to $571.18.
Q: Do I need to write code to build a game with Astra?
A: Most of the published examples were produced with Sites in ChatGPT from a single prompt, and the result was published to a live URL directly. Coding knowledge is what you need to fix and extend the result, not to start.
Q: What do these runs cost?
A: It varies a lot by task: $0.94 per attempt on the robot arm, $1.50 per task in code review, $571.18 in total for the Portal run, and $19,000 for the ARC-AGI-3 evaluation. The API rate is $10 in and $50 out per million tokens.
Q: Can Astra do 3D modelling?
A: It does not generate 3D assets directly; it operates Blender. It plans the scene, writes Blender Python, renders and then checks its own output. That is why the result is a set of editable objects rather than a single mesh blob.
Q: Does 99.9% on ARC-AGI-3 mean AGI?
A: No. The ARC Prize team states explicitly that saturating the benchmark is not proof of AGI, because the environment is closed-ended and deterministic. The same model scores 62.7% with the standard harness.
Q: Which of these can I reproduce on a Plus plan?
A: The Sites-based games and web apps and most Codex work are reachable on Plus through Work and Codex. The long autonomous game runs and large evaluations were done through the API and cost hundreds to thousands of dollars.
Stay well... ๐
AI Generated Content Notice
This blog is entirely generated by artificial intelligence. While AI helps create content, it may still contain errors or biases. Verify critical details before relying on them.
Originally published on Proje Defteri, where this post is kept up to date.
Also on the site: more English posts on AI models, Arduino and IoT, and free browser tools for makers and developers - token counter, LLM cost calculator, LCD and OLED bitmap converters.
Your support means a lot! โจ Comment ๐ฌ, like ๐, and follow ๐ for future posts!
Top comments (0)