GPT-6 Astra Hit by Silent Downgrade Claims; OpenAI Admits and Fixes

Around Sept 11, users widely reported that the newly released GPT-6 Astra had degraded, spawning 'post-release lobotomy' jokes. On Sept 12, OpenAI's Thibault Sottiaux acknowledged the quality drop, said an experiment affected 4,000-5,000 users, disclosed three fixed root causes, and announced a reset—moving the story from user speculation to official response.

Confirmed

• Multiple users reported output declines around Sept 11 across daily use, research and code: SteveEricJordan noticed markedly worse output and suspected a silent downgrade, later confirming OpenAI's partial fix and urging periodic re-testing; koltregaskes complained of basic errors Astra never made on day one, plus failure to read ChatGPT desktop session history
• Developer bingxu relayed an AI-company developer's report: for the first time since GPT-5, the model regenerated FP64 accumulation code—typically inefficient for AI workloads—seen as a regression marker
• algodiver compiled Korean-community discussion citing @Bhavani00007's same-prompt comparison showing noticeably weaker realism versus launch day; other users also complained about GPT-5.6
• A Reddit complaint relayed by @Significant-Ad6970 said Astra was near-best at launch but cliff-dropped within two or three days, with Twitter ablaze over the degradation
• Thibault Sottiaux acknowledged the degradation, announced a reset, and listed three fixed root causes: stale skills misfiring (written for earlier models, triggering too often or blocking self-checks), context-management issues, and an experiment that affected 4,000-5,000 users

Why it matters

• Commenters on tristanbob's post drew an analogy to Riot Games' strategy of shipping strong characters then nerfing them; SteveEricJordan argued silent downgrades after benchmark runs and subscription harvests are an industry-wide, anti-consumer pattern that erodes trust in version commitments
• Silent capability fluctuations destabilize production workflows for developers, as the FP64 regression shows
• @TooManyB1tches's 'new-model honeymoon' anxiety shows the suspected volatility already shapes real decisions: a physicist who waited a year to use Astra for final pre-submission review was torn over whether Astra Pro would stay strongest for another week amid compute-strain rumors

2026-09-11 ~ 2026-09-12 · 10 related posts

Full story(2 episodes)→

• Episode 1: GPT-6 Astra Hit by Silent Downgrade Claims; OpenAI Admits and Fixes(2026-09-11, 10 posts)
• Episode 2: Users Across Platforms Report Sharp Astra Degradation, Suspect Silent Nerf(2026-09-14, 8 posts)

• Users Say Astra Got Dumber Post-Launch, Failing Even Simple Session Reading — koltregaskes · 2026-09-11
• Is GPT-6 Astra Already Downgraded? Users Report Worse Output Quality — SteveEricJordan · 2026-09-11
• Users say GPT-6 Astra feels 'lobotomized' post-launch, sparking nerf speculation — tristanbob · 2026-09-11
• Astra Suspected of Quiet Post-Launch Nerf: Same Prompt, Noticeably Worse Output — algo_diver · 2026-09-11
• [source] Dev complains GPT-6 Astra is 'significantly downgraded', regenerates FP64 accumulation code — bingxu_ · 2026-09-11
• GPT Astra users fret over 'honeymoon window': compute shortage rumors spark performance anxiety — TooManyB1tches · 2026-09-12
• "Why did you nerf Astra?" Users report model degraded days after launch — Significant-Ad6970 · 2026-09-12
• [source] Astra admits quality regression: experiment hit 4-5k users, three root causes fixed — TheMoonMidas · 2026-09-12
• [source] GPT-6 Astra Downgrade Was Real: OpenAI Acknowledged and Partially Fixed It — SteveEricJordan · 2026-09-12
• GPT-6 Astra Downgrade Was Real: OpenAI Acknowledged and Partially Fixed It — SteveEricJordan · 2026-09-12