GPT-6 Astra: From Leak to Launch in Two Days

Episode 1 · Astra Rumored to Have Two Test Builds Focused on Long-Horizon Autonomy (2026-09-03, 3 posts)

Insider Lentils80 says two GPT Astra checkpoints, ultima-alpha and vega-alpha, are being tested with a focus on long-running autonomous tasks, while others call it a watershed model that breaks OpenAI's internal research benchmarks.

• Astra hailed as a watershed model, reportedly passing OpenAI's internal AI-research intern benchmark — haider1 · 2026-09-03
• Insider leak: OpenAI's Astra tested as 'ultima-alpha' and 'vega-alpha' checkpoints — Ok_Display_3159 · 2026-09-03
• GPT-6 Astra scores impressively on PostTrainBench Lite benchmark — maksym_andr · 2026-09-04

Episode 2 · GPT-6 Astra Launches with Benchmark Leaks: ARC-AGI-3 Hits 98.6% (2026-09-04, 41 posts)

OpenAI released GPT-6, codenamed Astra, on 09-04, with benchmark scores leaking alongside the launch that shattered records across multiple metrics. It is reportedly OpenAI's largest training run to date and the first model pretrained on Stargate infrastructure using more than 100,000 DBUs.

Confirmed

• Benchmark figures consistently relayed across multiple posts: ARC-AGI-3 up from 7.8% to 98.6%, FrontierMath Tier 4 v2 at 97.6%, DeepSWE v1.1 at 74.1%, BenchCAD 95.9%, GPQA Diamond 96%, a perfect score on ExploitBench, and across all displayed tests a comprehensive win over GPT-5.6 Sol (m4, m5, m11, m14, etc.)
• Official blog figures: AAII score of 61.2, slightly above GPT-5.6 Sol; Coding Agent Index of 67, slightly behind Fable 5 (m6, m17)
• Training details: researcher Aidan Clark says this is OpenAI's "largest training run to date," the first model pretrained on Stargate infrastructure with over 100,000 DBUs, and the first time a previous-generation model primarily supervised the training of the next generation (m10, m15, m20)

Unconfirmed

• Most scores circulated as screenshots; some posts (m8, m13) explicitly flag them as unofficial, and a few remind that pre-launch benchmark leaks are often exaggerated or fabricated; the FrontierMath Tier 4 score appears as both 97.6% and 97.4% in different posts—exact value awaits official confirmation

Why it matters

• The jump from 7.8% to 98.6% on ARC-AGI-3 is extraordinarily rare; randlongevity said they never predicted the score could be this high, seeing it as proof that "we're living in the singularity" (m19)
• After hands-on testing, bindureddy offered a finer judgment: Astra indeed beats Fable 5.1 on reasoning, math, data analysis, and research, but Fable 5.1 remains the king of coding (m12)—consistent with the official coding index slightly trailing Fable 5
• teortaxesTex lamented the breakneck iteration pace: GPT-5.6 Sol was surpassed less than two months after launch—the "hedonic treadmill" has peaked, obsolete upon release (m7)

• OpenAI's Astra trained on largest-ever run using 100K+ GPUs at Texas Stargate site — cedric_chee · 2026-09-04
• GPT-6 Astra benchmarks: how the new model scores — CounterReady4774 · 2026-09-04
• GPT-6 Astra Benchmarks Released — james_6732 · 2026-09-04
• Leaked GPT-6-Astra benchmarks claim 98.6% on ARC-AGI-3, unverified — scaling01 · 2026-09-04
• Unconfirmed GPT-6 Astra benchmark screenshots circulate online — haider1 · 2026-09-04
• GPT-6 Astra hits 97.4% on FrontierMath Tier 4, trained on 100k+ GPUs — ChrisGPT · 2026-09-04
• Unconfirmed GPT-6 Astra benchmarks leak amid 'AGI era' launch claims — mark_k · 2026-09-04
• GPT-6 Astra's SWE results show only ~6% gain over its predecessor, critic notes — robleclerc · 2026-09-04
• GPT-6 Astra beats Fable 5.1 on Terminal science and Automation benchmarks — ChrisGPT · 2026-09-04
• Leaked OpenAI Astra scores claim 98.6% on ARC-AGI-3 and near-perfect math benchmarks — teortaxesTex · 2026-09-04
• GPT-6 Astra full benchmarks: ARC-AGI-3 jumps from 7.8% to 98.6%, GPQA 96% — kimmonismus · 2026-09-04
• ARC-AGI-3 hits 98.6% with GPT-6, settling a five-month-old saturation bet — burny_tech · 2026-09-04
• GPT-6-Astra benchmark leaks: 98.6% on ARC-AGI-3, perfect score on ExploitBench — teropa · 2026-09-04
• GPT-6 Astra reportedly scores 98.6% on ARC-AGI-3, months ahead of Chollet's timeline — burny_tech · 2026-09-04
• Experts stunned as OpenAI's Astra posts a shocking ARC benchmark result — salahuddin · 2026-09-04
• Leaked GPT-6 Astra scores show 98.6% on ARC-AGI-3, beating GPT-5.6 Sol — iamaliveix · 2026-09-04
• Astra Is Out and So Are the Benchmarks — Hyleal · 2026-09-04
• GPT-6 Astra benchmarks leak: 98.6% on ARC-AGI-3, biggest training run yet — Dr_Singularity · 2026-09-04
• More GPT-6 Astra training details: first model supervised by prior models — Dr_Singularity · 2026-09-04
• OpenAI's Astra beats Fable 5.1 on reasoning and math, but Fable 5.1 remains the coding king — bindureddy · 2026-09-04

21 more related posts →

Episode 3 · OpenAI Calls Astra Its Most Aligned, Best Coding Model (2026-09-04, 2 posts)

OpenAI officially states GPT-6 Astra is its most aligned model with substantially improved intent understanding, and the best yet at computer use, software engineering, and vision.

• OpenAI says GPT-6 Astra is its most aligned model with better intent understanding — OpenAI · 2026-09-04
• OpenAI calls GPT-6 Astra its best model yet for computer use and SWE — dkundel · 2026-09-04