OpenAI GPT-5.6 Sol autonomously post-trains the smaller model Luna
OpenAI's GPT-5.6 Sol autonomously post-trained the smaller Luna model with a "fairly underspecified prompt"
AI BriefingOpenAI’s new flagship model, GPT-5.6 Sol, can independently optimize smaller models. After researchers provided “fairly vague prompts” through the Codex platform, Sol independently identified the training configuration, selected GPUs, and ran the post-training scripts for the Luna model. On an internal benchmark measuring recursive self-improvement (RSI), GPT-5.6 Sol scored 16.2 points higher than its predecessor, GPT-5.5. OpenAI said that during internal testing, average daily token output per active researcher more than doubled compared with GPT-5.5’s peak; over the past six months, the share of compute allocated to internal coding and reasoning increased 100-fold, while agent-based token usage grew about 22-fold.
Reported by 22 additional sources
- OpenAI proposes five steps for managing AI investmentOpenAI: Official Updates · July 15
- GPT-5.6 Sol tops Design Arena with an Elo score of 1353, surpassing Claude Fable 5Greg Brockman · July 13
- GPT-5.6 Sol Ultra solves a 50-year-old math conjecture in one hour using 64 subagentsRohan Paul · July 11
- GPT-5.6 enhances health intelligence, while Luna delivers stronger performance at lower costOpenAI · July 11
- GPT-5.6 Sol Ultra uses 64 sub-agents to solve a 50-year-old mathematical problem in one hourNoam Brown · July 11
- OpenAI employee explains the five reasoning levels and applicable scenarios of GPT-5.6 SolThe Decoder: AI News · July 11
- OpenAI releases the 5.6 model series: Sol, Terra, Luna, and Ultra modeJason Liu · July 10
- GPT-5.6 launches, ChatGPT and Codex mergeKazike · July 10
- GPT-5.6 released, Codex joins ChatGPTOpenAI Developers · July 10
- GPT-5.6 launches, ChatGPT and Codex officially mergeWeChat Official Account: Digital Life Kazike · July 10
- OpenAI launches the GPT-5.6 model series, including three variants and the ChatGPT Work toolTechCrunch: AI · July 10
- OpenAI releases GPT-5.6 (Sol, Terra, Luna): a three-tier model family with programmatic tool calling supported by the Responses APIMarkTechPost · July 10
- GPT-5.6 Sol and Terra officially launchCharlie Holtz · July 10
- GPT-5.6 launches with three model tiersOpenAI Developers · July 10
- GPT-5.6 Sol approaches the Fable 5 benchmark score at one-third the costThe Decoder: AI News · July 10
- OpenAI releases the public version of GPT-5.6 and the ChatGPT Work agentThe Decoder: AI News · July 10
- GPT-5.6 launches three models on OpenRouterOpenRouter · July 10
- GPT-5.6 released: Sol excels, while Terra/Luna offer low-cost efficiencyGreg Brockman · July 10
- GPT-5.6 is rolling out in the APIOpenAI Developers · July 10
- GPT-5.6 series Sol/Terra/Luna releasedOpenAI · July 10
- OpenAI releases its strongest model and best blog postSam Altman · July 10
- Sam Altman's 5.6 livestream: Three major new ChatGPT productsSam Altman · July 10
Source: The Decoder: AI News · the-decoder.com