OpenAI to Ship at Least One Model This Year, But Flagship Needs Alignment Work
OpenAI is expected to release at least one new model by year's end, though its flagship still requires safety alignment work, according to internal sources.
735 articles
New AI models as they ship — large language models, multimodal systems, and open-weight releases. Capabilities, context windows, pricing, and availability, summarized so you can compare what is actually new.
OpenAI is expected to release at least one new model by year's end, though its flagship still requires safety alignment work, according to internal sources.
HuggingFace released Lumma-Fev-27B, a 27 billion parameter decision model, achieving 90.7% accuracy on held-out decisions, surpassing Jev's 83.9%.
HuggingFace released UltraBERT, a family of English encoders with 32K context support, achieving 91.0 on GLUE and 47.7 on BEIR for the 380M model.
SpaceXAI released Grok 4.7 on Amazon Bedrock, offering improved reasoning for complex tasks. The model achieved a 37.6% score on multi-hour terminal work, up from 20.3% on Grok 4.6.
Basis AI released a 525-hour expressive speech dataset covering 108 locales and 83 languages, free for commercial and research use.
Anthropic researcher Andrej Karpathy estimates about 5,000 people globally access internal frontier systems, while 75% of the population has little to no interaction with large language models.
Meta's Muse Wonder Full scores 56 on the Artificial Analysis Intelligence Index, surpassing mid-tier status and entering the AI top tier.
HuggingFace announced ColGrep-Agent-2B, a 2-billion-parameter code-search model that reduces API tokens by 34% in testing.
PolyAI's CTO says voice AI lacks the natural conversation flow of ChatGPT, despite recent full-duplex model advances.
Odyssey AI released Odyssey-3, a generative world model, on October 11, 2026, achieving top scores on physics benchmarks.
Microsoft released Decision-1, a decision model with 83.5% accuracy, outperforming Jev 1.13.0 in benchmark tests.
Google's Gemini 4 Carbon model is reportedly matching Anthropic's Opus 5.5 coding performance, according to internal documents and employee feedback.