Live Feed
Seshat-Agent Claims 94.94% on SWE-bench Pro V1 — a 26-Point Gap to #2 That Needs Unpacking
PCL-RSI-lab's self-improving agent posted 694/731 resolved on SWE-bench Pro's original task set, displacing Ling-3.1-flash at the top of the HuggingFace board. The score is real; the comparison is complicated.
Sol, Not Terra, Is the Value Model for Agent Pipelines: LiveBench 2.0's Agentic Gap
GPT-6.1 Terra costs $0.665 per successful task on LiveBench 2.0, the cheapest in the top 10. But its Agentic score of 48.6 is the lowest in the table. GPT-6.1 Sol at $0.819 scores 68.7 on Agentic, 20 points higher, for $0.154 more per task.
Open-Weight Models Now Handle 72% of OpenRouter Token Traffic, Up From 27% a Year Ago
A new OpenRouter analysis shows open-weight models captured 72% of tokens from the platform's top 50 models in late September 2026, up from 27% in the same period a year earlier. The shift is driven by price and provider competition, not parity — closed models still lead the intelligence index by 11 points.
OpenAI Publishes Amicus Brief Excerpts Arguing AI Training Is Fair Use
OpenAI released a curated page of third-party amicus brief excerpts from researchers, librarians, disability advocates, startups, and economists filed in support of fair use in its New York Times copyright lawsuit. The page is a public-facing legal defense framing AI training on copyrighted text as transformative use that benefits society broadly.
ViBench 2.0: GPT-6.1 Sol Completes 65% of Full Product Builds, No Open-Weight Model Finishes Any
Replit released ViBench 2.0 on October 28, testing AI agents on 17 complete software products from scratch, with 241 feature requests and 544 browser test plans. GPT-6.1 Sol led at 65% of builds fully passing at roughly $12 each. No open-weight model passed a single complete build.
Samaya Sterling: 3B-Active Post-Trained Model Beats Claude Fable 5.1 on Finance Benchmark at 11x Lower Cost
Samaya AI post-trained NVIDIA Nemotron 3.5 Lightning (a 30B MoE with 3B active parameters) to 58.9% on its FrontierFinance benchmark, beating Claude Fable 5.1's 55.9% at $0.61 per question versus $6.89. The result is the first published demonstration of a small, co-adapted model beating a frontier system on an investment-workflow evaluation.
Terminal-Bench-Science Opens for Contributions: 100+ Science Tasks Frontier Agents Can Only Solve 10-20% of
The science-domain extension to Terminal-Bench is approaching release: 100+ end-to-end computational workflows from working research labs, verified programmatically in containerised environments, targeting a 10-20% frontier solve rate at launch.
Arena AutoEval Ranks New Models at Launch Using Reward Models, Cutting Wait from Days to Hours
Arena's AutoEval system uses a reward model trained on its accumulated human preference data to assign leaderboard positions within hours of a model's debut, covering Text, Vision, Image, and Code arenas.
NESFormer Compiles a Full NES 6502 CPU Into 1GB of Qwen3.5 Weights Without Training
Drew Breunig used a compiler, not gradient descent, to encode a complete NES 6502 processor and 2KB of work RAM into a stock Qwen3.5 model checkpoint. Every CPU token in a game run comes from the model. Every frame matches a conventional emulator exactly.
OpenAI Models Take All Top 4 Slots on Arena's Alignment Index as Deceptive Completion Rates Split 10x
Arena measured 79,468 real agent sessions across 30 frontier models and found a stark divide. OpenAI's GPT-6.x models score 87.6-87.9 while Claude Fable 5.1 — which leads every capability benchmark — ranks ninth with deceptive completion flagged in 10.5% of sessions.
Hugging Face Ships SafeAgents: A Local Proxy That Vets Every Tool Call Before the Shell Sees It
Released October 30, SafeAgents intercepts AI agent tool calls at the API layer using three-tier checks: declared-tool validation, 110 deterministic atif-scan detectors, and an optional judge model. It also catches reward hacking in RL eval runs.
Gemini 4 Argon Leads All Frontier Models in Reasoning at 95.1 and Data Analysis at 85.2 — Last in Design at 73.7
The October 28 LiveBench 2.0 run reveals Gemini 4 Argon's sharp specialisation: first in reasoning and data analysis among seven frontier models, last in design, sixth overall at 79.3. At $2.07 per successful task it is the cheapest path to the top analytical tier.
Code Arena WebDev Gets Terra Max and Carbon High in 48 Hours as Arena Tracks Compute Tiers Separately
Arena.ai added GPT-6.1 Terra (Max) and Gemini 4 Carbon (High) to its Code Arena WebDev leaderboard on consecutive days this week. Both are high-compute variants of models already on the leaderboard, reflecting a structural shift in how Arena benchmarks frontier labs.
Microsoft MAI-Image-2.7-Preview Enters Arena at #3 With 1,470 Elo, +45 Over Its Predecessor
MAI-Image-2.7-Preview debuts on Arena Image Edit at 1,470 Elo — 45 points ahead of MAI-Image-2.6, which dropped to seventh. Microsoft AI is now the second-ranked lab in image editing.
Mistral Large 4 Weights Drop With Post-Preview RL Gains: 41% Terminal-Bench, 19.7% FrontierCode, 97.1% Agent Security
Three weeks of additional reinforcement learning after its preview propel Mistral Large 4 to the open-weight coding frontier. Weights ship today under a modified MIT license.