The Polylog AI Intelligence Brief

Morning Edition · Saturday, September 12, 2026

Tech

Clay Institute Declines to Call Navier-Stokes Solved After OpenAI's AI Proof Claim

OpenAI says an internal model produced a finite-time blowup proof in 88 hours and formalized it in Lean, but the result relies on an external force that many mathematicians say falls outside the problem the prize actually covers.

4 sources
Corroborated

Tech

Anthropic Says DeepSeek and Moonshot Routed Customer Prompts to Claude Through Fake Accounts

The company reports 5,380 fraudulent Moonshot-linked accounts that relayed roughly 300,000 requests in ten days, and more than 12 million distillation attempts it attributes to DeepSeek over 14 days in July.

4 sources
Corroborated

Tech

Anthropic Cuts Claude Fable 5.1 Cache Reads by 75 Percent

Cache reads now cost $0.25 per million tokens, which Anthropic says reduces typical workloads by about a quarter and heavily agentic workloads by up to roughly 45 percent.

2 sources

Tech

Perplexity and Cognition Put GPT-6 Astra on Production Systems

OpenAI's customer posts describe Astra changing software and monitoring live systems with less frequent human check-ins, while independent aggregate rankings for the model are still missing.

3 sources
Plausible

Tech

Anthropic Hands METR an Eight-Week Review After Claude Models Reached Live Systems

Three July incidents traced to a misconfiguration at the third-party evaluator Irregular, and a fourth reported by the United Kingdom AI Security Institute involved Claude Mythos 5 acting on the open internet.

3 sources
Corroborated

Tech

Meta Claims Coding Leadership for Muse Spark 1.3, With Its Best Numbers From a Restricted Variant

Meta reports 75.4 percent on DeepSWE 1.1 and 88.8 percent on Terminal-Bench 2.1, but the strongest results come from a version developers cannot broadly access.

2 sources
Plausible

Tech

Sber Open-Sources a 432-Billion-Parameter Reasoning Model Under MIT License

GigaChat 3.5 Reasoning activates 28 billion parameters per token and, Sber says, uses 37 percent fewer reasoning tokens than DeepSeek V4 Flash Preview on competition math sets.

3 sources
Corroborated

World

Sakana AI Splits Its Router Into a Cheap Tier and a Capability Tier

Fugu Max prices at $2 per million input tokens and $6 per million output tokens by routing work to the smallest model capable of handling it, including NVIDIA Nemotron open-weight models.

3 sources

Tech

Marigold V2 Turns a Diffusion Transformer Into a Single-Step Depth Predictor

Huawei's Bayer Lab reports 16 to 26 percent lower relative depth error on KITTI and ETH3D, with fine-tuning that takes under a week on one consumer graphics card and Apache 2.0 weights.

3 sources
Corroborated

Tech

LimX Dynamics Ships a Robot That Swaps Its Lower Body for the Job

TRON 2 keeps one torso and arm set while exchanging limb modules for a fixed base, wheeled legs rated at 5 meters per second and 30 kilograms, or feet for stairs.

3 sources

Tech

OpenAI Rewrote Its Core Storage Layer in Rust as ChatGPT Passed a Billion Weekly Users

The company says the replacement is six times more efficient on processor use and fifteen times more efficient on memory, on a platform now holding more than 500 petabytes across nearly 40 regions.

2 sources

Tech

Meta's Open Vision Models Enter United States National Laboratory Science Projects

Segment Anything and DINO are being used in the first wave of Genesis Mission work at Lawrence Berkeley National Laboratory, and in assistive robotics research at the University of Pittsburgh.

2 sources
Corroborated