GPT-6 Astra review: what people built with it and where it falls short

3D scenes and games make up most of the builds people can actually show. In our hand-checked random sample, 15 of 44 confirmed builds were 3D scenes or models and 9 were playable games; the rest were 8 animation, video or image projects, 5 agent runs or computer-use jobs, 3 research or analysis projects, 3 apps or tools and 1 website.

Sample/method: Sample: 309 posts drawn at random from the 5,042 posts in the gpt-6-astra-fa-bu collection that do not mention Sol (2026-09-02 to 2026-10-15, UTC+8). AI classification, then every 'built' label read by hand (44 of 61 confirmed). Counts measure discussion.

Astra is usually driving another tool rather than producing the final asset on its own. Across all 5,042 posts, Blender is mentioned in 203, Three.js, WebGL or Babylon.js in 116, and Unreal in 51.

Token Gremlin (@TokenGremlin, Oct 11) built a little world inside a glass cabinet with GPT-6 Astra, Three.js and Blender; it runs entirely in the browser, and you can grab the clouds, make it rain and change the seasons. (Item: A browser glass-cabinet world built with GPT-6 Astra)

david (@0yield, Oct 4) posted Counter-Strike surf running entirely in Chrome, with GPT-6 Astra recreating the Counter-Strike: Source movement physics; the author says they still can't get over how close it feels. (Item: GPT-6 Astra rebuilds Counter-Strike: Source surf physics in Chrome)

David Kalmanson (@davidkal88) says one /goal had Astra reverse-engineer Baldur's Gate 3 and rebuild the Nautiloid in Babylon.js, walkable in the browser after 48 hours. That is the author's claim; we have not verified it.

eta (@seftsaint, Sep 20) moved a scene concept from Godot to Unreal using GPT-6 Astra, Unreal and Blender, then changed the trees and lighting. (Item: Rebuilding a concept scene in Unreal and Blender with GPT-6 Astra)

Long autonomous jobs are the second group. Minecraft comes up in 60 posts and ciphers in 47.

Carter Church (@CarterWChurch, Sep 30) used GPT-6 Astra to break an unsolved cipher to one of Napoleon's generals, unread for 217 years. His point: Astra completed the whole multi-modal workflow in about 6 hours from a single image and a goal. (Item: GPT-6 Astra decrypts a 217-year-old cipher to one of Napoleon's generals)

For comparison, OpenAI's own claim is that Astra 'brings stronger visual judgment to the websites, games, applications, and renderings it builds' and can create and host sites and games with Sites in ChatGPT.

The builders who finished something connected Astra to a real tool, brought their own assets and spent most of their time giving feedback.

Hubery (@hubery_bu, Oct 9) built an Unreal Engine 5 boss fight in ten days without ever having opened Unreal before, with Astra connected to the editor through MCP. (Item: Beginner builds a UE5 boss fight in ten days with GPT-6 Astra)

Hubery's steps: first, connect Astra to UE5 through the official Unreal MCP plugin that ships with UE 5.8, so it can build Blueprints, change parameters, lay out scenes and run the game. Second, don't ask it to make characters or animation from scratch; Hubery calls these AI's weakest area and used Epic's free Paragon assets from Fab instead. Third, let Astra write the combat system, skills, stamina and boss AI. Fourth, spend days playing and giving specific feedback: 'the move can't recover, I have to wait ages to act' works far better than 'it feels bad'. The verdict: Astra is very good at systems and scaffolding, but it cannot judge whether the game is fun.

Expert prompts matter in modelling too. Tom Krcha (@tomkrcha, Sep 9) found that naive prompts fail at clean subD topology and soft shapes, while refined, systematic prompts written by an expert succeed.

When authors gave Opus 5.5 and Astra the same prompt, Opus won 18 of the 29 posts with a verdict. Of all 48 same-prompt posts (Sep 15 to Oct 15, 2026), 18 favoured Opus, 3 favoured Astra, 8 were split or about the same and 19 gave no verdict.

Sample/method: All 48 posts in the gpt-6-astra-fa-bu collection (no Sol mention) that state 'same prompt' and compare with Opus 5.5 (2026-09-15 to 2026-10-15, UTC+8); each read in full by the editor for the author's own verdict. Mostly 3D, animation and visual tasks.

Vib3Coded (@vib3coded, Oct 7) gave both models the same prompt for an interactive plush shark. Opus won 'for me - it actually looks like a shark'; Astra made 'a hilarious little nightmare', for $1.97 less. (Item: Claude Opus 5.5 vs GPT-6 Astra: same shark prompt, two very different results)

The complaints are about taste and coherence rather than whether the code runs. kirill sh (@shneural) ran the same 15-second motion-graphics showreel prompt and called it 'not even close'. エクセル兄さん (@excel_niisan) had both models turn the same Word documents into a 2-3 minute onboarding video; Opus told the story well, while Astra's explanation felt incoherent.

Where authors give numbers, though, Astra was usually faster, and in most of the cost figures cheaper:

Table: Same prompt on Opus 5.5 and GPT-6 Astra: time, cost and the author's verdict

Sample/method: Numbers copied from the authors' posts (user self-reports, single runs, not reproduced; dates UTC+8); cost is the author's API-priced or platform figure. Only same-prompt posts that give time or cost for both models.

Editor's reading: these are mostly visual, single-run tests by people who chose to post them. They say little about code review, where some users prefer Astra. Robin Ebers (@robinebers) keeps Opus for daily work, but whenever an Opus plan gets an adversarial review, Astra 'absolutely tears it apart': about 70% genuine holes, 30% over-engineering. Kamil Baranek (@kamilbaranek_) reports that Opus 5.5 found 7 bugs, Astra found 70+, and Opus then agreed 50+ needed fixing.

First-hand failure reports are rarer than the noise suggests: only 7 of the 309 sampled posts were one. The reports we found fall into four kinds.

Table: Beyond taste: the other ways Astra fell short

Sample/method: Examples are posts in the gpt-6-astra-fa-bu collection that do not mention Sol, each read by the editor (user self-reports, not reproduced). OpenAI's side comes from its GPT-6 Astra launch page and a Sep 12 post by OpenAI's Tibo (@thsottiaux).

Refusals: 77 posts contain refusal words, but reading them all found only 4 first-hand reports. alice (@aliceisplaying) got 'i won't do that' when asking for reverse engineering; Taelin (@VictorTaelin) only got refusals when trying to audit Bend; Alex (@AlexLi98) says Astra would not fully translate a 97-year-old Japanese WW2 machine-gun manual; LukeYoungblood.eth (@LukeYoungblood) found both Astra and Opus refused to place trades even through an MCP with write access. OpenAI says Astra refuses advanced cybersecurity tasks such as proof-of-concept exploits, and that extra safety checks 'can sometimes slow, pause, or stop legitimate work'.

Early stops and degradation: on Sep 12, OpenAI's Tibo (@thsottiaux) said an opt-in context-management experiment could cause early stops or replies to older messages for about 4-5k users and had been disabled, along with fixes for misfiring skills and badly configured engines (a vendor statement). Complaints continued: Nico投资有道 (@tychozzz) wrote on Oct 5 that simple tasks ran for 10+ minutes with poor results. BridgeMind's (@bridgemindai) own NerfBench retest on Sep 28, however, put Astra at 102.8% of its launch score: 'No nerf detected'.

Allowance: on Oct 2, BridgeMind reported that one Astra agent on Ultrafast used up a $500 plan's weekly allowance in 15 minutes, then 95% of $2,500 in credits OpenAI had given the account. For allowance and price comparisons, see our Sol vs Astra article.

Before believing an 'I built this with Astra' post, look for a link, a repo, a playable demo or concrete numbers. In our sample, 17 of 61 build-looking posts failed that test.

The same formats keep coming back. 19 posts claim an Astra agent turned $X into $Y, 5 of them from one account. At least 4 accounts posted near-identical night logs: laptop closed at 10:40 or 11:05 pm, then a chief-of-staff agent named Echo splits the work. And one quote attributed to OpenAI CEO Sam Altman, about thousands of agents running every night on Astra, appears in 10 posts.

Editor's reading: these stories have no link, repo or demo to check. They are not evidence of what Astra does, and we left them out of the build counts above.