GPT-5.6 Sol Silently Nerfed, OpenAI Confirms Rollback
Shortly after the release of GPT-5.6 Sol, multiple users reported that the model became "faster but shallower." Further investigation revealed that its internal thinking budget (referred to by OpenAI as reasoning effort or juice) had been silently downgraded. Following intense community scrutiny, OpenAI ultimately confirmed the adjustments and stated they had been rolled back. However, this pattern of "launching with high scores and then quietly nerfing" has once again sparked strong concerns about the company's transparency.
Key Details
@johnseach pointed out a noticeable decline in GPT-5.6 Sol's ability to handle difficult problems post-launch, attributing it to a massive drop in internal juice values (max juice dropped from approximately 960 to 128). @ns123abc added that the model's context window was also reduced from 372k back to 272k, indicating an unannounced parameter and capability adjustment rather than just a subjective feeling. Meanwhile, @scaling01 claimed that the thinking budgets for Terra and Luna were unaffected, effectively granting them higher thinking allowances.
Community Reactions
@ns123abc mentioned that OpenAI's Tibo initially denied the "dumbing down" before confirming the adjustments and the subsequent rollback. @UltraRareAF directly reported that GPT-5.6-sol ultra felt "much dumber" than at launch, and @scaling01 announced their intention to cancel their subscription. @mark_k highlighted the irony that while OpenAI engineers claimed to have "halved inference costs," numerous users on X were simultaneously complaining about tightened GPT-5.6 quotas.
Controversies and Doubts
@ns123abc's core concern revolves around benchmark lag—evaluators often test the "just-released" version, while paying users might be interacting with a nerfed model. The official team did not proactively inform users until the community demanded answers. This suspicion of "using benchmark scores to secure a position at launch, only to weaken the model days later," even after the rollback, has fundamentally shaken user expectations regarding version stability.
2026-07-12 ~ 2026-07-13 · 8 related posts
- Episode 1: Polymarket Bets on GPT-5.6 Release Before July 7(2026-07-03, 8 posts)
- Episode 2: GPT 5.6 Is Opus-Tier, Cheaper and Faster Than Opus 4.8(2026-07-04, 3 posts)
- Episode 3: Rumors Swirl Around OpenAI’s GPT-5.6 Launch(2026-07-05, 17 posts)
- Episode 4: Unverified Rumor Says GPT-5.6 Found New Math(2026-07-06, 2 posts)
- Episode 5: Musk Announces Grok 4.5 with 1.5T Parameters and Enhanced Coding(2026-07-07, 25 posts)
- Episode 6: Prediction Markets Strongly Price In Grok 4.4 Release(2026-07-07, 2 posts)
- Episode 7: OpenAI Announces GPT-5.6 Sol for Thursday Release Amid Early Tester Reviews(2026-07-07, 58 posts)
- Episode 8: OpenAI Launches Full-Duplex Voice Model GPT-Live(2026-07-07, 44 posts)
- Episode 9: Grok 4.5 Released with Focus on Coding and Low Cost(2026-07-08, 61 posts)
- Episode 10: New ChatGPT Voice Mode Tested: Near-Human Multi-lingual Experience(2026-07-09, 14 posts)
- Episode 11: GPT-5.6 Tested: Major Coding Leap and Direct Rival to Fable 5(2026-07-09, 30 posts)
- Episode 12: xAI Launches Grok 4.5: Coding and Agent Focus to Rival Opus(2026-07-09, 55 posts)
- Episode 13: Grok 4.5 Benchmarks Strong but Faces Data Controversy(2026-07-09, 6 posts)
- Episode 14: Rumors Swirl Over Imminent Releases of Multiple AI Models(2026-07-09, 2 posts)
- Episode 15: Grok 4.5 Receives Widespread Praise for Speed and Coding(2026-07-09, 13 posts)
- Episode 16: Grok 4.5 Praised for Impressive Speed and Performance(2026-07-09, 2 posts)
- Episode 17: Grok 4.5 Outperforms Fable in Coding Speed and Efficiency(2026-07-09, 3 posts)
- Episode 18: Grok 4.5 Released, Ranks 6th on Vals Index(2026-07-09, 2 posts)
- Episode 19: Frontier Model Comparison: GPT-5.6 Praised for Value and Creativity(2026-07-09, 3 posts)
- Episode 20: OpenAI Launches GPT-5.6 Series: Multi-Agent and Cost-Efficiency(2026-07-09, 119 posts)
- GPT-5.6-sol Reported to Be Getting Dumber — UltraRareAF · 2026-07-12
- Inference Costs Halved but Quotas Tightened? — mark_k · 2026-07-12
- GPT-5.6 Thinking Budgets Reportedly Slashed — scaling01 · 2026-07-13
- OpenAI Accused of Silently Nerfing New Model — ns123abc · 2026-07-13
- [source] OpenAI Confirms Rolling Back GPT-5.6 Changes — ns123abc · 2026-07-13
- [source] OpenAI Rolls Back GPT-5.6 Adjustments — ns123abc · 2026-07-13
- OpenAI Accused of Silently Downgrading Models — ns123abc · 2026-07-13
- [source] GPT-5.6 Sol Reportedly Hit by Reasoning Budget Cut — johnseach · 2026-07-13