GPT-6 Astra Reportedly Beat Portal and Wrote a Bach Chorale — Unverified

A viral X essay from Dr. Alex Wissner-Gross — the physicist and AI researcher who co-founded Physical Superintelligence (PSI) earlier this month — claims GPT-6 Astra cleared Portal without human assistance, procedurally "grew" a three.js forest scene containing 3,808 individual trees, and composed what's described as the first AI-generated Bach-style chorale with a correctly resolved passing tone. All three claims come from one person's essay, not a primary OpenAI announcement, and none carries a playthrough video, source code, or score anyone else has reviewed. Here's what the claims actually are, why the grouping is more interesting than any one of them alone, and why a reader currently has no way to check any of it.

This is a companion read to explainx.ai's GPT-6 Astra launch benchmark coverage and its verified launch-week demo roundup — neither of which covers Portal, the three.js forest, or the Bach chorale, because none of those three specific claims appeared in the launch materials or the demo videos we checked. This post treats them separately, and treats "one commentator reported" as the operative phrase throughout.

TL;DR

The three claims, as reported

Wissner-Gross's essay frames GPT-6 Astra's capability jump through three specific, named results rather than a benchmark percentage. None of them appear in OpenAI's own published launch benchmarks — ARC-AGI-3, ExploitBench, the Intelligence Index — which is itself worth noting: these are claims from outside the official benchmark table, relayed by a single credible-but-independent voice.

Portal, cleared unaided. Portal is a well-known first-person puzzle-platformer built around a portal gun that lets the player create linked entry and exit points on flat surfaces — clearing it requires multi-step spatial reasoning, momentum tracking through connected portals, and sequencing several non-obvious moves in the right order. That much is public knowledge about a real, famous game. What's attributed to the commentator's report, not confirmed by OpenAI, is the claim that Astra cleared it without human assistance — notable in part because OpenAI itself set "beat Portal" as a specific, publicly stated capability goal for its research programs back in 2016, long before Astra existed. A decade-old, self-set target being reportedly cleared is a good story. It is still, as reported here, one person's account with no attached playthrough.

A 3,808-tree three.js forest, "grown." three.js is a real, widely used JavaScript library for rendering 3D graphics in a browser — the same category of tool behind several procedural-world demos explainx.ai has covered, including Karpathy's Opus 5 Lord of the Rings world. The specific claim — that Astra "grew" a procedural forest scene containing exactly 3,808 individual trees — is attributed to the commentator's report. The precision of the number is part of what makes the claim interesting: a specific count is more checkable than a vague "photorealistic forest," but it's also unverified pending an actual look at the running scene or its source.

A Bach-style chorale with a correct passing tone. A chorale, in the sense meant here, is a four-part harmonized hymn setting — a compositional form most closely associated with J.S. Bach, who wrote hundreds of them as some of the most rigorously rule-governed music in the Western classical tradition. A passing tone is a specific, well-defined music-theory term: a non-chord melodic note that connects two chord tones by stepping through the scale between them, rather than a note that belongs to the underlying harmony. Getting a passing tone to resolve correctly is a real, nontrivial rule to satisfy — voice-leading and chorale harmonization carry a dense set of constraints (parallel fifths and octaves to avoid, proper resolution of dissonance, correct doubling) that a system can violate in ways a trained ear or a music theorist can immediately spot. That's what makes this claim different in kind from most AI-music hype: it's a checkable correctness signal, not an aesthetic judgment. Whether Astra actually produced a chorale that satisfies it, per the commentator's report, is unverified.

Why three benchmarks beat one

The interesting move in Wissner-Gross's essay isn't any single claim — it's the grouping. Portal, the three.js forest, and the Bach chorale test three domains that don't reduce to the same underlying skill:

• Embodied spatial puzzle-solving (Portal) requires tracking physical state through a game world, reasoning about momentum and geometry, and sequencing actions toward a goal that isn't verbally specified step by step.
• Generative 3D procedural content (the forest) requires producing structured, varied, large-scale output from code — closer to the kind of world-generation and CAD work covered elsewhere in Astra's launch coverage than to game-playing.
• Rule-constrained creative composition (the chorale) requires satisfying a dense, formal rule set — voice-leading, harmony, resolution — while still producing something musically coherent, which is a different discipline than either spatial reasoning or 3D generation.

A model doing well at one of these tells you almost nothing about the other two. A model doing well at all three, if the underlying claims hold up, is a genuinely broader statement about generalization than any individual result — which is exactly why grouping them into one essay, rather than three separate posts, is a smart rhetorical (and potentially substantive) choice. The catch is that "if the underlying claims hold up" is doing all the work in that sentence, and right now nothing lets a reader check it.

The double-edged part: checkable in principle, unchecked in practice

Here's the tension worth sitting with. Each of these three claims is, in principle, one of the more falsifiable kinds of AI capability claim you'll see this year:

• A Portal claim can be checked by watching the actual playthrough — portals, moves, and puzzle order are all visible on screen.
• A passing-tone claim can be checked by a music theorist looking at the actual score — resolution either happens correctly or it doesn't.
• A tree-count claim can be checked by loading the actual three.js scene and, in principle, counting.

None of that supporting material — video, score, live link — is provided here. That's a meaningfully different situation than a vague, unfalsifiable claim ("the AI is incredibly creative now"), and also a meaningfully different situation than a fully verified one. It sits in between: specific enough that someone could check it, and nobody currently has. Treat it the way explainx.ai treated Matt Shumer's multi-agent Unreal Engine anecdote in the demo roundup — a genuinely interesting, viral, unverified report from a credible source, worth someone reproducing and documenting properly, not evidence of anything beyond that yet.

Why calibrated skepticism, not dismissal

Wissner-Gross isn't a random account — he's a physicist with a documented research background who co-founded PSI with a stated mission of discovering and commercializing physics breakthroughs, which lends the essay more credibility than an anonymous hype thread would carry. That's a reason to take the claims seriously enough to want them checked — not a reason to accept them as settled.

It's also worth reading this alongside two things explainx.ai has already documented about Astra specifically. First, the verified launch-week demo roundup found that even among demos with actual video evidence, roughly a fifth of the "best" clips turned out to be OpenAI's own official marketing reshared rather than independently reproduced — a reminder that "someone posted a video" and "someone reproduced the claim" are different evidentiary categories, and this essay doesn't clear even the lower bar of having a video attached. Second, OpenAI's own published benchmark numbers for Astra were quietly revised, twice, within days of launch — a hallucination-rate figure moved from 4.2% to 2% and back, and a cybersecurity comparison score turned out to rest on a reasoning tier not actually available to customers. Neither incident proves this new essay's claims are wrong. Both are good reasons to hold any new, single-source Astra capability claim — official or independent — a little more loosely than the headline number invites.

What this actually means for builders

The honest scope of these claims is narrower than "GPT-6 Astra is more capable, period." Creative and generative benchmarks like Portal, procedural 3D scenes, and formal music composition are genuinely useful signals for a specific slice of teams: people building game asset generation pipelines, procedural content tools, or AI music composition software. If a model really can generate rule-correct musical structure or large-scale procedural 3D content reliably, that's directly relevant to what those teams ship — closer in spirit to explainx.ai's coverage of tools like Google's Lyria 3.5 music generation or image-to-3D rigged character pipelines than to a general capability claim.

What it doesn't tell you is how the model performs at coding, math reasoning, or enterprise agentic tasks. Creative-domain competence and formal-reasoning competence are only loosely correlated in practice — a model tuned or particularly strong at generating musically coherent, rule-following output isn't guaranteed to carry that same rigor into, say, debugging a production codebase or proving a math result, and vice versa. The skill of satisfying voice-leading rules in a chorale is a different kind of "correctness" than the skill of satisfying a compiler or a unit test. Builders evaluating Astra for coding or reasoning-heavy workloads should keep leaning on the benchmark table in the main launch post, not this essay — and builders in creative tooling should treat this essay as an interesting lead worth watching for a reproducible follow-up, not as a result to build a roadmap on yet.

Honest limitations

• All three claims trace to a single commentator's essay, not a primary OpenAI announcement, a benchmark leaderboard, or an independent reproduction.
• No supporting artifacts are available — no Portal playthrough video, no three.js source code or live scene link, no published chorale score.
• The specific numbers (3,808 trees) and qualitative claims (correct passing tone, unaided Portal clearance) have not been independently checked by explainx.ai or, as far as we can find, by anyone else publicly.
• This is not the same evidentiary category as OpenAI's own published benchmarks, which at least carry a methodology and a number, even when that methodology has itself been revised post-launch.

Related on explainx.ai

• GPT-6 Astra's actual launch: every benchmark, the pricing, and the ARC-AGI harness controversy
• The 11 best GPT-6 Astra demos from launch week, verified
• OpenAI changed GPT-6 Astra's benchmark numbers after launch — twice
• Physical Superintelligence raises $58M — who is Dr. Alex Wissner-Gross
• GPT-6 Astra vs Claude Fable 5.1: which model wins where
• Karpathy's $10 Opus 5 Lord of the Rings three.js world
• Google Lyria 3.5: AI music generation comes to Gemini
• Image-to-3D rigged game character pipeline, explained

This post is sourced entirely to one commentator's viral X essay as of September 7, 2026. Every claim about Portal, the three.js forest, and the Bach chorale is attributed as reported, not as confirmed by OpenAI or independently verified by explainx.ai. Check back for updates if a playthrough, source code, or score surfaces that would let any of these claims actually be checked.