GPT-6 Astra launch: dazzling demos clash with rocky real-world tests
OpenAI's new flagship GPT-6 Astra launched amid huge hype after NVIDIA CEO Jensen Huang tweeted on Sept 7 that it was trained on 100,000+ Grace Blackwell NVLink72 systems and declared 'AGI is here.' Early user tests, however, paint a far messier picture, with unstable real-world performance, runaway costs, and community skepticism about claimed advantages over rivals.
Confirmed
- GPT-6 Astra has 1.05M context and is priced at $10/$50 per million tokens (@docdavkitty)
- @Firm-Club-8334 burned through a $200 quota in 8 hours, then found 3 of 4 task outputs didn't actually run; official demos focused on 3D, Blender and games, while coding and agent tasks underperformed
- @onusozlm's hands-on 3D modeling test found the model merely 'okay,' well below the polished results posted by OpenAI employees and early-access users—even in the flagship demo scenario
- @ivanbezdomny complained Astra doesn't 'manage itself' on long-running tasks: unsure when to stop or how to version its work, and it exhausted his quota; he remains long-term bullish and also criticized OpenAI Codex's handling of its own issues
- @Liueroteme likened working with Astra on large tasks to a disgruntled senior engineer: without explaining why a change is needed, it doesn't much care about doing it right
- @docdavkitty noted the API's reasoning.effort defaults to low (only low/med available), contradicting 'maximum reasoning' marketing and creating a potential cost trap
- @median0rc relayed a Reddit user's test (math/ML workflows): hallucinated websites, higher-than-advertised consumption vs Sol, prompt misunderstandings unseen in years, and inconsistent capability gains
- @bendee983 cited theo's 'jagged frontier' framing: the model accomplishes previously unimaginable high-level tasks while making elementary mistakes, leaving users struggling to adapt
- @miltonian3 relayed Reddit skepticism: no data shows Astra 'far exceeding' mythos/fable—at best on par—so the 'AGI moment' looks more like marketing
Unconfirmed
- Huang's 'AGI is here' claim is a personal assertion contradicted by user tests, with no independent verification
- Reported task-failure details lack full reproducible materials; failure rates and comparisons with mythos/fable cannot be verified
Why it matters
This episode reprises the familiar gap between dazzling demos and real productivity: official showcase scenarios diverge sharply from coding and long-horizon agent work, and even the flagship 3D use case underdelivers for ordinary users. The 'jagged' capability boundary makes errors unpredictable, the default low reasoning effort risks a mismatch with advertised benchmarks, and token billing makes long tasks financially hazardous—the $200 quota evaporating in 8 hours being a case in point. Together with unresolved quantitative doubts versus mythos/fable, these details matter more than benchmark scores for developers considering a subscription.
2026-09-07 ~ 2026-09-09 · 10 related posts
- Episode 1: GPT-6 Astra Blazes Through Usage Quotas Across Paid Tiers, Users Complain(2026-09-04, 26 posts)
- Episode 2: Users across platforms report abnormally fast quota burn with OpenAI's new Astra model(2026-09-06, 20 posts)
- Episode 3: GPT-6 Astra launch: dazzling demos clash with rocky real-world tests(2026-09-07, 10 posts)
- Episode 4: Astra burns through paid tiers: users across Plus to Pro report quota woes(2026-09-07, 13 posts)
- Episode 5: Users complain GPT-6 Astra quota exhausts within an hour of reset(2026-09-11, 2 posts)
- Episode 6: Users complain OpenAI quietly cut Pro subscription usage limits(2026-09-14, 2 posts)
Primary sources
- Early user on GPT Astra: sloppy at self-managing long tasks and ran out of quota — ivan_bezdomny · 2026-09-07
- Dev slams OpenAI Codex: Astra 'doesn't know how to run itself' on long tasks — ivan_bezdomny · 2026-09-07
- Early GPT 6 Astra 3D modeling tests impress less than claimed, user reports prompt-following quirks — onusoz · 2026-09-07
- [source] "Not AGI": user burns $200 in 8 hours on GPT-6 Astra, 3 of 4 tasks broken — Firm-Club-8334 · 2026-09-07
- [source] 'Not AGI': User Spent $200 in 8 Hours on Astra, Found 3 of 4 Tasks Broken — Firm-Club-8334 · 2026-09-07
- Reddit Skeptic: Astra Looks Like Catch-Up, Not a Leap — 'AGI Moment' Smells Like Hype — miltonian3 · 2026-09-07
- [source] GPT-6 Astra ships with API defaulting to low reasoning despite 'Highest reasoning' marketing — docdavkitty · 2026-09-08
- GPT-6 Astra is brilliant and dumb at once: adapting to AI's jagged frontier — bendee983 · 2026-09-08
- Astra on Big Tasks: Slacks Off Unless You Explain Why Changes Are Needed — Liu_eroteme · 2026-09-08
- Reddit user reports Astra regression: fake website links, misread prompts, higher usage — median0rc · 2026-09-09