The Polylog AI Intelligence Brief

Morning Edition · Monday, September 7, 2026

Tech

OpenAI Says Coding Agents Now Supply 3.1 Agent-Workdays of Effort for Every Human Workday

The company says it met its internal "automated research intern" target on schedule, with the median researcher consuming more than $600 a day of inference at list prices for application programming interface (API) access by mid-August.

2 sources
Plausible

Tech

Nvidia's Jensen Huang Declares AGI Has Arrived With GPT-6 Astra as OpenAI's Chief Scientist Urges Voluntary Slowdowns

Huang says Astra was trained on more than 100,000 Grace Blackwell NVLink72 systems, with 400,000 more accelerators coming online. OpenAI chief scientist Jakub Pachocki writes that no AI lab has solved alignment well enough to justify scaling at maximum speed.

2 sources
Plausible

Tech

Anthropic's Claude Fable 5.1 More Than Doubles Its Score on an Agentic Science Benchmark and Cuts Cache Reads 75%

Fable 5.1 scores 52.6% on Terminal-Bench-Science 0.1, up from 24.7% for Fable 5, while the price of cached input drops from $1.00 to $0.25 per million tokens with headline pricing unchanged.

1 source
Plausible

Tech

Anthropic Opens a Research Preview of a Standard for AI Agents to Operate Laboratory and Factory Hardware

Early testers report an agent raising laser stabilization on QuEra's quantum computers from 58% to 99.3% and reducing a Janelia imaging experiment from several weeks to a single day.

1 source
Plausible

Tech

Berkeley Lab Runs Meta's SAM 3 and DINOv3 on Government GPUs, Cutting Beamline Annotation From a Month to 15 Minutes

The SYNAPS-I project under the Department of Energy's Genesis Mission runs both open-weight vision models on 300 Nvidia A100 accelerators entirely inside federal infrastructure.

1 source
Plausible

Tech

Google DeepMind's WeatherNext 3 Produces Hourly Global Forecasts at 5-Kilometer Resolution

The model drops physics-based simulation for a scaled-up Functional Generative Network and ranks first in Brightband's independent live evaluation of operational AI forecasts.

2 sources
Corroborated

Tech

Anthropic Turns On Text Watermarking in Claude Fable 5.1 and Mythos 5.1 to Meet EU AI Act Commitments

The method changes the randomness used when the model selects tokens rather than editing the output afterward, follows Google DeepMind's SynthID-Text approach, and survives copying but not a full rewrite.

1 source

Tech

Germany Pushes a Data Center Buildout It Says Will Quadruple AI Capacity by 2030

Chancellor Friedrich Merz says construction is proceeding at a scale Germany has not previously reached, against a national strategy targeting up to 2,000 megawatts of AI and high-performance computing capacity.

1 source
Plausible

Tech

New Evaluation Infrastructure Targets the Cost of Actually Running Agentic Benchmarks

Harbor Adapters offers a unified harness across agentic benchmarks and ships with Harbor-Index, a curated meta-dataset, addressing the environment and integration overhead that keeps most agent evaluations from being reproduced.

2 sources

Tech

Pittsburgh Researchers Run Meta's Vision Models On-Device for a $41.5 Million Robotic Wheelchair Programme

The Human Engineering Research Laboratories are using DINOv3 and Segment Anything locally for real-time perception, under an award of up to $41.5 million from the Advanced Research Projects Agency for Health.

1 source
Corroborated

Tech

Meta Has Held Muse Spark API Pricing at $1.25 and $4.25 per Million Tokens Across Three Model Generations

The rate has not moved since the developer preview opened in July, while the underlying model advanced from version 1.1 to 1.3, and a $0.10 tier is available to developers who let Meta train on their traffic.

1 source
Corroborated

Tech

Two New Preprints Push World Models Toward Physical Quantities Rather Than Pixel Reconstruction

One proposes spectral targets to structure the latent space of JEPA-style predictors, the other releases a video dataset of falling coffee grounds paired with per-frame scale readings as ground truth for weight estimation.

1 source