AI Models

JetBrains has released Mellum2.1, a small open model for coding agents, and Hugging Face has released Carbon-A, a 1.2B-parameter model that predicts protein-coding regions in DNA, plus a database of 566 million predicted loci. What each does, and which numbers are the developers' own.

A new benchmark, TasteVal, says the best AI model now reaches expert-level results on eight AI research tasks with 2.3 times less compute, though its authors list reasons that may overstate progress. Preference Model has open-sourced Karotte, a framework for training environments that are harder to game, and DatologyAI has released Zephon, a data loader that keeps experiments repeatable when the GPU count changes. What each claims, and who measured it.

Scale AI and Elorian's Humanity's Sixth Sense benchmark finds people score 93.1% on everyday visual questions while the best model, GPT-6 Astra, scores 53.6%, mostly because models misread the picture. A separate preprint, not yet peer reviewed, says base models can reason almost as well as RL-trained ones when their answer starts with the right word.

On 6 October OpenAI launched a Decisions API that returns yes/no probabilities, picks and rubric scores from GPT-6 Luna, and cut its API usage tiers from five to three. Perplexity released pplx-decider-v1.1-27b, an Apache 2.0 open-weight decision model. What decision models are, what each costs, and which scores are the vendors' own.

Liquid AI open-sourced two "decision models" that give a structured answer in a single pass, in about 8 ms on an RTX 4090 by Liquid's figures. Perplexity open-sourced two search models that read text, images and PDF pages without OCR, where a small model can query an index built by a big one. What each does, in plain words.

Anthropic's Claude Haiku 5.5 costs around 75% less to run than Haiku 4.5, by Anthropic's estimate, and jumps from 15.7% to 72.4% on Anthropic's OSWorld 2.1 computer-use test. Sonnet 5.5 cache reads are also halved. What is confirmed, what is company benchmark, and who it is for.

OpenAI is rolling GPT-6 into ChatGPT's Chat tab for every plan: GPT-6 Sol for paid tiers from 7 October, GPT-6 Luna for Free and Go from 8 October, plus "Intelligent UI" answers with charts, forms and mini tools. What changed, which Sol this is, and what OpenAI's own safety card flags.

NVIDIA’s LoGRA paper shows low-rank gradient sketches cutting average RL training memory by up to 45.7% on tested Qwen reasoning setups—and making 27B RL runnable on one 8×H100 node where dense Adam OOMs. What the method is, what Table 1 shows, and what it does not prove.

Avi Chawla’s 7 October explainer contrasts standard RAG with a Jev gate between retrieval and generation. What hybrid search already does, what typed probabilities add, why thresholds stay in code, and what this architecture does not replace.

Scientific American: OpenAI released 372 math/CS results from an unreleased model on 6 October 2026. Many are Lean-verified; independent check pending.