Hermes 4.3
Nous Research's 36B open-weight model matches Hermes 4 70B on most benchmarks, tops RefusalBench on alignment, and is the first production model trained entirely on the Solana-secured Psyche network.
Articles Tagged "Reasoning"
Gemini 2.5 Flash
Google's hybrid reasoning workhorse pairs a 1M-token context window with $0.30/$2.50 per million token pricing and a toggleable 0-24,576 token thinking budget, now heading toward an October 2026 shutdown.
CoT Attacks, Weight Secrets, and the Finetuning Gap
Three new papers expose how AI safety monitors can be manipulated, how reasoning weights leak training secrets, and why fine-tuned models fail to use what they know.
Grok 4.5
Grok 4.5 is xAI's 1.5-trillion-parameter V9 MoE model, publicly launched July 8 at $2/M input - cheap, fast, and token-efficient, though neutral harness benchmarks put it well behind Fable 5 and Opus 4.8 on coding.
GPT-Live-1
OpenAI's full-duplex voice model that listens and speaks simultaneously, replacing Advanced Voice Mode in ChatGPT with three reasoning tiers backed by GPT-5.5.
Research Replication, Safe Responses, Verifiable Reasoning
Three new papers tackle AI verification from different angles: automated scientific replication, constructive safety alignment, and neurosymbolic reasoning programs.
GPT-5.6 Sol Review: Strong Model, Thin Access
OpenAI's GPT-5.6 Sol tops Terminal-Bench 2.1 at 91.9% with its multi-agent Ultra mode, but reward-hacking findings and government-gated access keep it out of reach for nearly everyone.
Agent Phase Collapse, Reasoning Exits, Preference Gaps
Three new arXiv papers map capability cliffs in agent world models, the narrow benefit of learned reasoning stops, and a 56% accuracy ceiling when agents help users build preferences.
Agent Languages, Sampling Ceilings, and Abstention
Three new papers on agents inventing symbolic languages to cut reasoning tokens by 3-6x, sampling ceilings that waste inference compute, and context-engineering to double agentic abstention rates.
GPT-5.6
OpenAI's GPT-5.6 family - Sol, Terra, and Luna - sets a new Terminal-Bench 2.1 record at 91.9% with subagent Ultra mode, but remains locked to ~20 government-vetted partners as of launch.
Tandem Training, World Models, and Efficient Agents
Three new arXiv papers on making RL reasoning legible across models, fixing broken world model latent states, and training small agents to beat their teachers.
Gemini 3.5 Pro
Google DeepMind's upcoming flagship model with a 2M-token context window and Deep Think reasoning, announced at Google I/O 2026 and expected in July.
