GPT-6 Astra: From Rumors to Benchmark Dominance

Episode 1 · GPT-6 Astra's 3D modeling demos flood in, community calls it SOTA (2026-09-04, 37 posts)

Within a day of GPT-6 Astra's release, the community produced a wave of hands-on 3D modeling demos, with multiple developers calling it SOTA-level. Workflows in architectural visualization, game prototyping, and 3D content creation appear dramatically compressed.

Confirmed

• Dimillian posted a full demo thread: specifying only room count and style (modern minimalist, tree-surrounded), Astra produced design drawings, built the model in Blender, wrote its own export pipeline to Unreal Engine 5, iterated against engine fidelity checks, and delivered a standalone app with 60 FPS walkable real-time exploration of the whole house. He added technical details: Astra works via Blender headless/CLI, writing 3D models in Python rather than manipulating the GUI; for rendering, the trick is lighting prompts—asking for "cinematic golden-hour shots" yields stunning 4K lighting in Cycles. OpenAI's dev team also published a deep-dive build article from Blender scene to walkable UE5 world
• Theo used GPT-6 Astra (Codex) to one-shot a 3D game running directly in the browser; the workflow is minimal—install Blender, open Codex. He called Astra world-class at Blender and 3D spatial reasoning
• Early-access user @tomkrcha demonstrated full 3D reconstruction from a single house photo in Blender, with toys, appliances, and furniture as real, tweakable geometry at 60 FPS; OpenAI staffer npew turned real-estate listing photos into complete Blender 3D models
• Stefan3DAI showed Astra working autonomously like an artist in Blender: gathering reference images, building scenes, rendering test frames, comparing and correcting, then importing a walkable level into UE5
• iamfakhrealam compiled 12 top cases from day one: single-prompt whole houses, an Apple Park replica, a 3D-reconstructed Oval Office, and a Long Dark-style piece; users like dejavucoder called Astra's 3D modeling SOTA
• Creator cerspense built a full DJ truck and lighting console in minutes within Sentinel, arguing playability explodes when software and model/harness are co-designed; joshmillgate said Astra one-shot a tool that creates near-indistinguishable 3D mockups; another user generated smoothly interactive Three.js web 3D content, suggesting Blender may no longer be needed and compressing hours of work to 30 minutes; Nasereliver demoed a 3D Grok bot from one prompt

Why it matters

• 3D modeling has long been a professional domain AI couldn't reach; Astra's autonomous reference-gathering–modeling–rendering–correction loop mirrors a real artist's workflow, and its headless/CLI + Python approach reveals a new paradigm for agents operating professional software
• Connected toolchains across photo reconstruction, Blender/UE5/Three.js/Sentinel could compress prototyping iteration costs by an order of magnitude for architects, game developers, and creators

• GPT-6 Astra demos modeling a house in Blender into a walkable UE5 scene — ChrisGPT · 2026-09-04
• GPT-6 Astra demo: Blender house model turned into walkable Unreal Engine 5 scene — BLCNYY · 2026-09-04
• Dimillian walkthrough: building the Astra 3D demo house from Blender to UE5 — Dimillian · 2026-09-04
• OpenAI's Astra 3D demo: from a prompt about rooms to a full house plan — Dimillian · 2026-09-04
• AI agent builds 3D models by writing Python for headless Blender, not GUI — Dimillian · 2026-09-04
• Cinematic 4K renders from an AI agent hinge on golden-hour lighting prompts — Dimillian · 2026-09-04
• Astra nails cinematic golden-hour lighting in Blender Cycles 4K renders — Dimillian · 2026-09-04
• Astra wrote its own UE5 export pipeline and delivered a 60 FPS standalone app — Dimillian · 2026-09-04
• Dimillian has Astra generate storyboards comparing Blender and Unreal Engine renders — Dimillian · 2026-09-04
• Astra builds cinematic 3D scenes end-to-end: skybox, lighting, shots — all from a few prompts — Dimillian · 2026-09-04
• Theo: GPT-6 Astra one-shots a 3D browser game, world-class at Blender and spatial reasoning — charliermarsh · 2026-09-04
• GPT-6 Astra hailed SOTA at 3D modeling with wild design output — dejavucoder · 2026-09-04
• GPT-6 Astra turns one house photo into a tweakable 3D model running at 60fps — gabrielchua · 2026-09-04
• GPT-6 Astra turns realtor staging photos into a full Blender 3D model — npew · 2026-09-04
• GPT-6 Astra turns Blender scenes into walkable UE5 levels, stunning devs — Dimillian · 2026-09-04
• GPT-6 Astra demo: models a house in Blender, turns it into a walkable UE5 scene — RileyRalmuto · 2026-09-04
• GPT-6 Astra one-shots a 3D mockup tool nearly indistinguishable from reality — OpenAIDevs · 2026-09-04
• OpenAI's Astra can now layout and route PCBs, sparking hardware engineering debate — MikePFrank · 2026-09-04
• GPT-6 demo shows fully autonomous 3D modeling in Blender, shipping scenes to UE5 — petewoodbridge · 2026-09-04
• GPT-6 Astra reportedly lays out PCBs in KiCad, wowing netizens — cneuralnetwork · 2026-09-04

17 more related posts →

Episode 2 · Unverified 'GPT-6 Astra' Reportedly Generates Impressive 3D Models (2026-09-04, 2 posts)

Users claim an unverified model dubbed 'GPT-6 Astra' can produce stunning 3D models of real locations, with one saying it made Blender part of their daily workflow; OpenAI has not confirmed the model's existence.

• User claims GPT-6 Astra is so good at 3D modeling he opens Blender daily — flavioAd · 2026-09-04
• User claims GPT 6 Astra can 3D model a local island in one shot — MickeySteamboat · 2026-09-05

Episode 3 · OpenAI Launches GPT-6 Astra and Declares the AGI Era (2026-09-04, 21 posts)

OpenAI released the rumored Astra on Sept 3, confirming it is GPT-6, with President Greg Brockman closing the launch with "welcome to the AGI era," sparking fierce debate over the AGI claim. Early hands-on reports broadly agree this is a genuine capability leap — especially in long-horizon agent tasks, coding, and frontend — though the "AGI" label and some benchmark figures remain contested. The model is rolling out, and it is already influencing developer tool choices and subscription decisions.

Confirmed

• OpenAI confirmed Astra is GPT-6; Brockman declared the "AGI era" and the model is rolling out (m9, m10, m16, m19).
• Benchmarks: record Epoch index of 169 (m16); reviewer htihle found Astra (high tier) scored 92.9% on WeirdML, tied for first with Fable 5.1 (max tier), setting new per-task highs on 6 of 17 tasks with notably less generated code (m19). AI Explained published a 22-minute video covering ARC-AGI 3, FrontierMath, and Agents Last Exam results, saying the model is strong enough that OpenAI itself is worried (m1).
• A wave of third-party demos: WorldofAI's roundup shows a Call-of-Duty-like game built in 30 minutes, Pokémon-style gameplay, a futuristic city in Blender, plus computer and robot control demos (m2). Machine Heart cataloged viral demos including runnable browser 3D scenes with TypeScript + Three.js and zero-shot robotic arm control, while noting their limitations (m20).
• Positive early user reports: ivanbezdomny called it the smartest practical coding model he has used, with more coherent and less annoying refactors (m16, m17); BLUECOW009 said that with steering it seems able to solve any problem he finds in a single run (m13); cneuralnetwork reported speed, skills comprehension, and bug-finding strengths (m14); vista8 felt the model "suddenly got much smarter" for the first time since GPT-4 and Claude Sonnet 3.5 (m15).
• YuchenjUW reported Astra fixed GPT's biggest weakness — frontend — beating Fable 5.1 in design sense, one-shot games, and tool/computer use, and is considering canceling one subscription (m9, m17, m18); EXM7777 praised its autonomous goal-setting and subagent orchestration in real business use, with an Opus 4.5-like conversational feel (m7); Scobleizer's roundup of X discussion found 3D modeling and cross-app long tasks most praised (m5).

Unconfirmed

• A Reddit post claims Astra can autonomously browse, build sites, and complete tasks (m4) — unverified via official channels, though it broadly matches the third-party demos in m2.
• Conflicting ARC-AGI-3 figures of 99.9% vs 62.7%; johnseach argued the capability jump is real but the AGI label is not, noting the reasoning process is harder to supervise (m20); on ThursdAI, Chollet still rejects the AGI claim (m10, m11).
• User assessments diverge: ZenenoDev's day-one verdict was "high intelligence, low intuition," saying Astra fails to understand existing architecture and implementation intent and may revert to Sol 5.6 (m3); haider1 saw near-universal leads in pretraining skills but only a small coding edge over Fable 5.1 (m8); bookwormengr's computer-use test took nearly 15 minutes for a one-page Pythagorean theorem slide, positive overall but slow (m12).

Why it matters

• This is OpenAI's first flagship framed around an "AGI era," directly provoking debate over AGI's definition. Regardless of the label, the measured leap in long-horizon agent tasks, coding, and frontend is already shaping tool selection and subscription decisions (m9, m18 mention canceling one). Harder-to-supervise reasoning (m20) and claims that even OpenAI is worried about the model's strength (m1) add new safety concerns.

• AI Explained breaks down GPT-6 Astra: so capable OpenAI itself is worried — AI Explained · 2026-09-04
• Greg Brockman closes GPT-6 briefing with "welcome to the AGI era" — thursdai_pod · 2026-09-05
• GPT-6 Astra deep dive: Greg Brockman declares "welcome to the AGI era" — thursdai_pod · 2026-09-05
• Heavy User's Early Review of Astra (GPT-6): High Intelligence, Low Intuition — ZenenoDev · 2026-09-05
• Early user: Astra feels like the first real intelligence jump since GPT-4 — vista8 · 2026-09-05
• GPT 6 Astra day-one impressions: fast, good with skills, solid bug-finding — cneuralnetwork · 2026-09-05
• GPT-6 Astra (high) hits 92.9% on WeirdML, matching Fable 5.1 (max) — zainhas · 2026-09-05
• GPT-6 Astra hands-on: fixes GPT's frontend weakness, users mull canceling a subscription — Yuchenj_UW · 2026-09-06
• Real-world GPT-Astra review: exceptional long-horizon autonomy, writing finally clicks — EXM7777 · 2026-09-06
• Early user: GPT-6 Astra seems to solve any problem in a single run with steering — BLUECOW009 · 2026-09-06
• GPT-6 Astra early verdict: 3D modeling and cross-app computer work are its standout strengths — Scobleizer · 2026-09-06
• GPT-6 Astra Hits 169 Epoch Record but Its Reasoning Is Harder to Monitor — ivan_bezdomny · 2026-09-06
• Dev spends a day with GPT-6 Astra: smarter at refactoring, no reason to switch models for coding — ivan_bezdomny · 2026-09-06
• OpenAI's GPT-6 Astra touted as its most advanced model: autonomous browsing, building, tasks — CreamHoliday4754 · 2026-09-06
• User tests OpenAI Astra computer use: one PPT slide took nearly 15 minutes — bookwormengr · 2026-09-06
• GPT-6 Astra sparks AGI debate: 99.9% vs 62.7% ARC-AGI-3 scores explained — johnseach · 2026-09-06
• GPT-6 Astra leaps ahead in world modeling and reasoning, but coding gap vs Fable 5.1 stays slim — haider1 · 2026-09-06
• GPT-6 Astra Demos: COD-Style Game in 30 Minutes, Computer Use and Robot Control — WorldofAI · 2026-09-06
• Researcher hails unverified GPT-6 'Astra' as fixing GPT's frontend weakness — MaziyarPanahi · 2026-09-06
• GPT-6 Astra demos: 3D modeling, software control, and 95% zero-shot robot arm success — 机器之心 · 2026-09-06

1 more related posts →

Episode 4 · GPT-6 Astra Sets New Records Across Benchmarks With More Intelligence Per Token (2026-09-05, 19 posts)

After rolling out to some users around Sept 5, GPT-6 Astra set new records across multiple third-party benchmarks, with the core narrative being "stronger and more token-efficient": Artificial Analysis's updated index shows it outscoring GPT-5.6 Sol while defining a new Pareto frontier of intelligence versus output tokens, cutting output tokens by roughly 10% at maximum reasoning; it scored a record 98.1 (xhigh) and 97.7 (high) on the Extended NYT Connections benchmark at about 40% lower cost per puzzle; and it ranked first on Terminal Bench 4.0 with the Codex harness at half the cost of the runner-up, roughly 47% below prior solutions.

Confirmed

• Artificial Analysis: GPT-6 Astra surpasses GPT-5.6 Sol and sets a new Pareto frontier on intelligence index versus output tokens (m2, m19)
• NYT Connections: 98.1 at xhigh, 97.7 at high, beating GPT-5.6 Sol at 40% lower cost per puzzle (m4, m18)
• Terminal Bench 4.0: first place at half the runner-up's cost, 47% cheaper than before (m7, m17)
• eyebench-v3: Astra-max scored 95% with 1/3.8 the output tokens and half the cost of Sol-max; adonissingh argues it leads on "intelligence per token" even at low reasoning effort (m12, m15, m16)
• MineBench.ai says Astra's first-generation output quality far exceeds the current top four models, the biggest quality jump since launch (m3)
• Hands-on: kimmonismus found Astra-Medium roughly matches GPT-5.6 xhigh at 1/3 the cost; Reddit user dubesor preferred its visual magazine layouts over GPT-5.6 Sol on iOS; Angaisb was satisfied at Max reasoning; Simon Willison's pelican SVG comparison across Sol, Terra and Luna peaked at 63 cents per run (m5, m6, m8, m9, m11)
• API pricing: input $10 / output $50 per 1M tokens, a 2.5x premium over GPT-5.6 Sol ($4 / $20) (m14)

Unconfirmed

• A FrontierMath T4 leak claims Astra scores above 5.6 Sol Pro (Max); commenter teortaxesTex doubts open-source models will match this within 18 months — personal opinion, no official data (m10)
• Developer xeophon's early take: incremental improvement rather than a step change — code quality similar and still needing steering, data analysis at the same level (m13)
• Most benchmark data comes from community and third-party leaderboards; OpenAI has not systematically confirmed it (m2, m3, m12)

Why it matters

• Token efficiency is seen as Astra's underdiscussed edge: Reddit user Cagnazzo82 argues that fewer tokens per task, combined with speed and quality, strikes directly at competitors' cost weakness (m1)
• DataLearnerAI notes the 2.5x API premium did not bring proportional reasoning gains — improvements concentrate in agentic tasks — so the "premium" and "token-saving" narratives coexist and real cost depends on use case (m14)
• If the "intelligence per token" advantage holds, it could reshape how model value is assessed and pressure rivals' pricing and product strategies

• GPT-6 Astra Beats 5.6 Sol Pro (Max) on FrontierMath T4; Open Models Seen 18 Months Behind — inductionheads · 2026-09-05
• Artificial Analysis: GPT-6 Astra cuts output tokens ~10% vs GPT-5.6 Sol at max effort — peterjliu · 2026-09-05
• GPT-6 Astra sets new NYT Connections benchmark record at 98.1 while costing ~40% less — zero0_one1 · 2026-09-05
• Early hands-on compares GPT-6 Astra vs GPT-5.6 Sol at max reasoning effort — Angaisb_ · 2026-09-05
• GPT-6 Astra's Untold Story May Be Token Efficiency, Argues Reddit User — Cagnazzo82 · 2026-09-05
• GPT-6 Astra tops Terminal Bench 4.0 at half the cost of #2 — charliermarsh · 2026-09-05
• MineBench Says GPT-6 Astra's First Generation Beats Current Top 4 Models — Ballist1cGamer · 2026-09-05
• GPT-6 Astra Max beats GPT-5.6 Sol at visual essays and magazine spreads, user finds — dubesor · 2026-09-05
• Hands-on: GPT 6 Astra Beats GPT 5.6 Sol at Visual Magazine Spreads — dubesor · 2026-09-05
• Simon Willison benchmarks GPT-6 Astra vs GPT-5.6 on SVG: max effort costs 63 cents per call — gaganghotra_ · 2026-09-05
• Early GPT-6 Astra tests: same intelligence as GPT-5.6 xhigh at roughly 1/3 the cost — charliermarsh · 2026-09-05
• GPT-6 Astra sets Extended NYT Connections record of 98.1 while costing ~40% less — sherwinwu · 2026-09-05
• GPT-6 Astra runs Terminal Bench 4.0 at 47% lower cost — sherwinwu · 2026-09-05
• Astra-max hits 95% on eyebench-v3 at half the cost of Sol-max, tokens ~3.8x fewer — adonis_singh · 2026-09-05
• Blogger: Astra-max's eyebench-v3 dominance is 'just not fair' — adonis_singh · 2026-09-05
• Astra-max claims vastly better intelligence-per-token even at low reasoning — adonis_singh · 2026-09-05
• GPT-6-Astra beats GPT-5.6-Sol on updated Artificial Analysis Intelligence Index — DaserTheLaser · 2026-09-05
• GPT-6 Astra Costs 2.5× More Than GPT-5.6 Sol, but the Gains Are Agentic, Not Reasoning — DataLearnerAI · 2026-09-06
• Early Astra hands-on: code quality and data analysis feel incremental, says developer — xeophon · 2026-09-06

Episode 5 · GPT-6 Astra finds up to 176x code speedups in five minutes (2026-09-05, 2 posts)

A developer reports GPT-6 Astra found 4.5x to 176x performance speedups in his codebase within five minutes, while another developer jokes the model found nothing left to optimize in his code.

• Dev says GPT-6 Astra found 4.5x to 176x speedups in his codebase in 5 minutes — charliermarsh · 2026-09-05
• Users report GPT-6 Astra finding up to 176x code speedups — or nothing at all — ivan_bezdomny · 2026-09-06

Episode 6 · GPT-6 Astra demos flood social media, sparking both 'Cambrian explosion' hype and skepticism (2026-09-05, 9 posts)

OpenAI announced GPT-6 Astra, claiming it can do anything a human can do on a computer, only faster. Viral demos flooded social media after launch; blogger minchoi rounded up 10 striking cases within two days and another 10 within 91 hours, while DeryaTR described a 'Cambrian explosion' of creativity. Meanwhile, the official demo drew criticism and parody, and the gap between claimed capability and actual creative quality remains to be seen.

Confirmed

• minchoi posted roundups on 09-05 (10 cases two days post-launch) and 09-06 (10 more within 91 hours), covering anatomy diagrams, LEGO sets, GTA scenes, a V8 engine, and even an entire forest.
• Video creator @davis7 had GPT-6 Astra auto-process freshly recorded footage—ingesting clips, color grading, aligning audio tracks, and preparing a Final Cut project for human polish—calling it another 'holy shit moment.'
• Indie developer @adamlyttleapps generated an app with GPT-6 Astra Ultra in 26 minutes, consuming 10,460,032 tokens.
• Podcaster Duncan Trussell had the model build a full Backrooms scene in Blender with VHS effects in about 5 prompts—roughly 30 minutes of modeling plus 20 minutes of export, with sound effects also done by the model; not a pure AI-generated video.
• DeryaTR posted on 09-07 that just 3 days after launch X is seeing a rare creative surge, arguing the 'Cambrian explosion' will accelerate as millions gain access, giving more people their own AGI moment.
• Some users produced response videos mocking or pushing back against OpenAI's official demo.

Unconfirmed

• Claims of 'near-AGI capability' reflect only some users' views; OpenAI made no such statement.
• Most demos are personal shares without independent verification.

Why it matters

• These cases showcase the model's automation in real workflows—video editing, 3D modeling, app development—notably by directly operating professional tools like Blender and Final Cut rather than generative output.
• Critics note the first 17 seconds of the official demo expose a flaw common to nearly all AI creative tools—endlessly chasing an outdated creative paradigm—suggesting a gap between marketing and actual creative quality that warrants continued scrutiny.

• 'First 17 seconds expose the huge deficiency in AI creative tools' — GPT-6 Astra demo critiqued — plopesresearch · 2026-09-05
• GPT-6 Astra Goes Viral in 2 Days: 10 Wild Demos Rounded Up — minchoi · 2026-09-05
• GPT-6 Astra Builds Backrooms Entirely in Blender in ~5 Prompts — minchoi · 2026-09-05
• GPT-6 Astra Ultra Generates an App in 26 Minutes Using 10.46M Tokens — minchoi · 2026-09-05
• GPT-6 Astra Auto-Preps Entire Final Cut Project: Import, Color Grade, Sync Clips — minchoi · 2026-09-05
• GPT 6 rebuilt the Backrooms in Blender with ~5 prompts, sound effects included — QuintinPope5 · 2026-09-06
• 10 Wild Demos in Under 91 Hours After OpenAI's GPT-6 Astra Launch — minchoi · 2026-09-06
• Someone made a response video to OpenAI's GPT-6 Astra demo — 4KTV · 2026-09-07
• GPT-6 Astra sparks a 'Cambrian explosion' of creativity just 3 days after launch — DeryaTR_ · 2026-09-07

Episode 7 · OpenAI's Astra Reportedly Trained on Over 100,000 GPUs (2026-09-05, 2 posts)

OpenAI's GPT-6 Astra was reportedly pre-trained on more than 100,000 GPUs at the Stargate data center in Abilene, Texas, making it what Nina Schick calls the first billion-dollar AI model.

• GPT-6 Astra reportedly pretrained on 100,000+ GPUs at Stargate in Abilene, Texas — NinaDSchick · 2026-09-05
• Astra reportedly trained on 100,000 GPUs, a staggering compute scale — sudoraohacker · 2026-09-05

Episode 8 · GPT-6 Astra reportedly scores 95% on robot control at ~43% of prior cost (2026-09-05, 6 posts)

A set of widely circulated numbers on X (Twitter) claims GPT-6 Astra scores 95% on a robot control task, versus 40% for the previous Fable 5.1, while using 6.2x fewer output tokens and costing 2.3x less (about 43% of the prior generation). The figures originated from a thread by @chooijeq and then propagated through a multi-level repost chain with identical numbers.

Confirmed

• @chooijeq's thread first gave these numbers: 95% on robot control (vs 40% for Fable 5.1), 6.2x fewer output tokens, 2.3x lower cost.
• @willcb, @AaronBergman18, @maxpaperclips, @chrisjpaxton, and @TinfoilTricorn (reposted by beffjezos) relayed the same figures, forming a clear propagation chain.

Unconfirmed

• The specific benchmark, evaluation setup, and source of the "robot control task" have no primary documentation; the 95% and cost figures exist only as secondhand thread claims without official papers or docs.

Why it matters

• If accurate, the numbers imply a general-purpose chat model substantially beating its predecessor on robot control while roughly halving inference cost, with significant implications for model selection in embodied AI.
• Community reactions split: @AaronBergman18 marveled that "robot control has been stuffed into a chat Transformer"; beffjezos declared "embodied AGI is near"; DJiafei quipped that GPT-6 should just call the MolmoAct2 checkpoint via API; Dorialexander used @maxpaperclips's repost to revisit a prior European debate about LLMs — underscoring the clear divide between general-model and specialized-model camps.

• GPT-6 Astra Scores 95% on Robot Control, 6.2x Fewer Tokens Than Fable 5.1's 40% — scaling01 · 2026-09-05
• GPT-6 Astra hits 95% on robot control, 2.3x cheaper than prior model — AaronBergman18 · 2026-09-06
• GPT-6 hits 95% on robot control; joker asks if it just API-calls MolmoAct2 — chris_j_paxton · 2026-09-06
• GPT-6 Astra Scores 95% on Robot Control, 2.3x Cheaper Than Fable 5.1 — willcb · 2026-09-06
• GPT-6 Astra scores 95% on robot control task, 2.3x cheaper than Fable 5.1's 40% — max_paperclips · 2026-09-06
• GPT-6 Astra scores 95% on robot control, 40% for Fable 5.1 at far lower cost — TinfoilTricorn · 2026-09-06

Episode 9 · GPT-6 Astra Reportedly Beats Gemini in Vision Benchmarks (2026-09-05, 3 posts)

An unconfirmed model dubbed GPT-6 Astra is making waves, reportedly leading the EyeBench vision benchmark and outperforming Gemini in detailed visual recognition.

• GPT-6 Astra shows massive leap on EyeBench visual reasoning benchmark — Waiting4AniHaremFDVR · 2026-09-05
• GPT-6 Astra called best vision model yet, dethroning Gemini on detection — jonstephens85 · 2026-09-06
• GPT-6 Astra hailed as best vision model yet, unverified claims say data labeling is dead — iamrobotbear · 2026-09-07

Episode 10 · OpenAI Launches 24-Hour GPT-6 Astra Demo Challenge (2026-09-05, 4 posts)

OpenAI's developer account challenged builders to ship GPT-6 Astra demos in 24 hours. Submissions ranged from an interactive tutorial on looped Transformers to a full Blender-to-Three.js seasonal cabin website, while one user claimed to be using the model to reverse-engineer and pirate games.

• OpenAI Devs launches 24-hour GPT-6 Astra shipping challenge — OpenAIDevs · 2026-09-05
• GPT-6 builds a 9-stage interactive explainer on looped transformers — cto_junior · 2026-09-06
• Building a four-seasons 3D cottage site with GPT-6 Astra, Blender and Three.js end to end — op7418 · 2026-09-06
• Developer tells OpenAI he's using GPT-6 Astra to reverse engineer and steal games — JasonBotterill · 2026-09-06

Episode 11 · GPT-6-Astra tops MathArena leaderboard (2026-09-06, 2 posts)

GPT-6-Astra has taken first place on MathArena with expected performance around 90%, while tests show it rarely makes math errors and is notably token-efficient, making it arguably the most reliable and cost-effective math model.

• GPT-6-Astra takes first on MathArena with 90% expected performance, using just 12k tokens per question — scaling01 · 2026-09-06
• GPT-6 Astra reportedly almost never wrong on math, called most trustworthy model — gabrielchua · 2026-09-06

Episode 12 · GPT-6 Astra Tops Code Arena at 1797, Fourth Leader Change in a Week (2026-09-06, 7 posts)

LMArena's Code Arena (WebDev leaderboard) published its latest results: OpenAI's new GPT-6 Astra (Max) took first place with 1797 points, 35 ahead of runner-up Claude Fable 5.1 (Max) at 1762, reshaping the Pareto price frontier. This confirmed earlier community speculation and marked the fourth leader change in a week, underscoring the intensity of coding model competition.

Confirmed

• LMArena officially announced GPT-6 Astra (Max) topping Code Arena: WebDev at 1797 points, 35 ahead of Claude Fable 5.1 (Max) at 1762; the third spot belongs to a Claude model.
• Evaluation method: models plan and build real web apps with tools, judged by pairwise human voting.
• Qwen 3.8 Flash Next entered the top ten.
• Reshares noted this was the fourth leader change within a week.

Timeline (09-06 to 09-07)

• On 09-06, user py-net posted that Astra and Fable 5.1 formed a new top tier clearly ahead of other models, and inferred Astra should rank first by comparing score changes after upgrading from Fable 5 and Sol.
• The same day, LMArena's official account and multiple reshares confirmed GPT-6 Astra (Max) at 1797 points in first place.
• On 09-07, annbordetsky and others reshared, highlighting the four leader changes in one week.

Why it matters

• The leaderboard is based on real web app building with human pairwise comparison, giving it strong reference value for practical coding ability; py-net argued Arena's coding benchmark this time accurately reflected real-world programming skill.
• Astra's 35-point lead and reshaped Pareto frontier suggest OpenAI's new model sets a new benchmark on both performance and price-performance, potentially influencing future model pricing and competition.
• Frequent leader changes within a week show top coding models are iterating rapidly and the competitive landscape remains in flux.

• GPT-6 Astra (Max) tops Code Arena WebDev at 1797 pts, priced at $40/Mtoken — Polymarket · 2026-09-06
• Arena Coding Benchmark Declared Fixed as User Argues Astra Deserves Top Spot — py-net · 2026-09-06
• Astra and Fable 5.1 take #1 on Code Arena, in a class of their own — py-net · 2026-09-06
• GPT-6 Astra Tops Code Arena WebDev at 1,797, Qwen 3.8 Flash Next Cracks Top 10 — pbaylies · 2026-09-06
• GPT-6 Astra (Max) tops Code Arena at 1797 pts, opens +35pt lead over Claude Fable 5.1 — jxnlco · 2026-09-06
• GPT-6 Astra takes #1 on Code Arena with 1797 pts, 35-pt lead over Claude — arena · 2026-09-06
• GPT-6 Astra tops Code Arena at 1797 pts — fourth #1 coding model in a week — annbordetsky · 2026-09-07

Episode 13 · Leaked Benchmarks Claim GPT-6 Astra Aces Enterprise Tasks (2026-09-06, 2 posts)

Leaked Signal65 PINNACLE benchmark results claim the rumored OpenAI GPT-6 Astra completed 279 of 280 enterprise agent tasks with zero hallucinations. Neither the model nor the data has been officially confirmed.

• Unverified: GPT-6 Astra Reportedly Completes 279 of 280 Enterprise Agent Jobs — ryanshrout · 2026-09-06
• GPT-6 Astra leaks in benchmark: 279/280 enterprise tasks done, zero hallucination — ryanshrout · 2026-09-06

Episode 14 · GPT-6 Astra Rebuilds Maya Ruins 3D Model in Two Hours (2026-09-06, 3 posts)

A researcher tested GPT-6 Astra, which autonomously searched the web via the /goal feature and generated a 3D model of the Maya site Xunantunich in about two hours, also nailing image replication in one shot.

• GPT-6 Astra generates 3D model of Maya site Xunantunich in about 2 hours — omarsar0 · 2026-09-06
• GPT-6 Astra's /goal feature autonomously sourced web images for 3D rebuild — omarsar0 · 2026-09-06
• Astra nails image recreation in one go where Fable kept falling short — omarsar0 · 2026-09-06

Episode 15 · Developer Says OpenAI's Astra Is First Model to Make Real Progress on His Ultra-Complex Project (2026-09-06, 2 posts)

Developer mrjonnfinger reported that while other models felt like swimming in mud on his highly complex project, OpenAI's Astra is the first to deliver substantive progress, though completion remains uncertain.

• Early user: OpenAI's Astra is the first model making real progress on an extreme-complexity task — mrjonfinger · 2026-09-06
• Dev says OpenAI's Astra is first model making progress on his 'unreasonably complex' project — mrjonfinger · 2026-09-06

Episode 16 · Rumored GPT-6 "Astra" Beats Portal Autonomously, Unverified (2026-09-06, 7 posts)

Rumors claim OpenAI's GPT-6, codenamed "Astra," has fully autonomously beaten the 3D puzzle game Portal, with user AcerFur posting a full playthrough video (edited to cut out the model's thinking pauses). Follow-up retweets also claim the model beat Rimworld, scored near-perfect on all ARC-AGI tasks, and cleared all 48 levels of the "I'm not a robot" game. The story drew wide attention, but overall authenticity remains unverified.

Confirmed

• On 09-06, @cozyblazex posted the claim that GPT-6 "Astra" beat Portal autonomously, self-labeling it unverified and recalling OpenAI's 2016 goal of "solving multiple games with a single agent"
• On 09-07, AcerFur released the full playthrough video with thinking segments cut, the first complete record of the model completing the 3D puzzle game

Unconfirmed

• Whether GPT-6 "Astra" exists and beat Portal autonomously: no implementation details were provided, and "Astra" is not an official OpenAI product name
• The Rimworld and near-perfect ARC-AGI claims appear only in retells (via DanielWhiteson, Chase) with no primary evidence
• xiaohu's claim that "GPT-6 Astra" beat all 48 levels of the "I'm not a robot" game is unverified
• Gary Marcus notes the video is edited with almost no implementation details; he says it would be stunning if real and not relying on heavy pre-scripting, but wants more information

Why it matters

• scaling01 recalls OpenAI's 2016 goal of one agent solving many games; if true, the rumor would mark that vision landing a decade later
• kalomaze proposes the community actually needs Portal's co-op mode as a multi-agent evaluation benchmark
• If verified, it would signal a major leap in 3D real-time navigation and long-horizon planning; if debunked, it is another reminder to stay skeptical of edited demos

• GPT-6 Astra autonomously beats Portal, echoing OpenAI's 2016 game-solving goal — scaling01 · 2026-09-06
• GPT-6 'Astra' reportedly beat Portal autonomously and near-aced ARC-AGI — teortaxesTex · 2026-09-06
• Full recording of GPT-6 Astra autonomously beating Portal, thinking pauses edited out — burny_tech · 2026-09-07
• GPT-6 'Astra' said to beat Portal solo, dev calls for Portal 2 co-op as multiagent eval — burny_tech · 2026-09-07
• Claimed GPT-6 'Astra' beats Portal solo; Gary Marcus flags edited footage, few details — GaryMarcus · 2026-09-07
• ChatGPT Astra-6 reportedly beats Rimworld and Portal, sparking buzz — DanielWhiteson · 2026-09-07
• Claim: 'GPT-6 Astra' beats all 48 levels of 'I'm not a robot' CAPTCHA game — unverified — xiaohu · 2026-09-07

Episode 17 · GPT-6 Astra saturates spatial reasoning benchmarks (2026-09-06, 4 posts)

GPT-6 Astra reportedly scores 91.8% on SpatialBench, beating the 80% human baseline, and fully saturates a long-unsolved community spatial reasoning benchmark, prompting claims that vision in LLMs is effectively solved.

• Independent SpatialBench fully saturated by Astra, author declares LLM vision solved — pbaylies · 2026-09-06
• Astra shows striking gains on visual-spatial intelligence benchmark SpatialBench — socoolandawesome · 2026-09-06
• Unverified: GPT-6 Astra reportedly scores 91.8% on SpatialBench vs 80% human baseline — VraserX · 2026-09-06
• Astra saturates blogger's spatial reasoning eval, prompting 'LLM vision is solved' claim — lukaszkaiser · 2026-09-06