The Polylog AI Intelligence Brief

Morning Edition · Thursday, September 3, 2026

Tech

Google Prices Gemini 3.8 Flash at 75 Cents per Million Input Tokens and Claims Opus-Class Reasoning

The introductory rate runs only through December 31, after which Google doubles it, making the next four months a deliberate push to win market share in the cheap tier.

3 sources
Corroborated

Tech

Meta's Muse Spark 1.3 Scores 75.4 on DeepSWE, Above Claude Opus 5 and GPT-5.6 Sol

The model jumped from 55.0 for version 1.2 on the same long-horizon coding benchmark, a gain large enough that the size of it is itself the reason for caution.

3 sources
Plausible

Tech

Google Restricts Its Cyber-Tuned Gemini to Vetted Governments and Infrastructure Operators

Gemini 3.8 Flash Cyber scored 86.2 percent on CyberGym against 77.5 percent for the previous cyber model, and Google says only approved defenders can use it.

2 sources
Corroborated

Tech

CrowdStrike and Nvidia Launch SafeMind, Pairing an Attacking Model With a Defending One

CrowdStrike says its defensive model beat leading frontier models on internal evaluations while cutting operating cost by 99 percent, a figure the company has not opened to outside testing.

1 source
Plausible

Geopolitics

Former Google Engineer Sentenced to Just Under a Year for Stealing AI Chip Infrastructure Secrets

Judge Vince Chhabria vacated all seven economic-espionage convictions before sentencing, ruling prosecutors never proved Linwei Ding intended to benefit the Chinese state.

1 source
Corroborated

Tech

A New Benchmark Measures Whether Frontier Models Know They Are Being Tested

EvalDetectBench, from LASR Labs and the UK AI Security Institute, scores two things at once: how reliably a model detects an evaluation, and how detectable each benchmark is.

2 sources

Tech

Researchers Recover Secrets From a Model's Hidden Context Without Any Jailbreak

The attack sends ordinary, non-malicious queries and uses a surrogate model to work out which candidate secret best explains the answers, which makes it hard to detect by prompt filtering.

1 source

Tech

Anthropic Opens a Hardware Standard Letting Agents Drive Lab and Factory Instruments

In early testing, Anthropic says agents using the standard raised laser stabilization on QuEra's quantum computers from 58 percent to 99.3 percent and compressed a Janelia imaging experiment from weeks to a day.

1 source
Corroborated

Tech

Google's Agentic Video Mode Cuts Token Use by up to 88 Percent

Instead of ingesting a video at a fixed frame rate, the model chooses which segments to inspect and at what resolution, which Google says lowers analysis cost by up to 66 percent.

2 sources
Corroborated

Tech

Meta's Open Vision Models Move Into Assistive Robotics and Department of Energy Science

SAM 3 and DINOv3 now segment X-ray and neutron imaging data in real time for a flagship Genesis Mission project led by Lawrence Berkeley National Laboratory.

2 sources
Corroborated

Tech

Altman Says He Now Thinks Near-Term Superintelligence Is Possible, a Year After Doubting It

The OpenAI chief executive paired the statement with an explicit hedge, saying he is not confident it will happen and is describing the speed of progress rather than a forecast.

2 sources
Corroborated

Tech

Two Papers Attack the Same Weakness in Retrieval Pipelines: Rewarding Right Answers From Wrong Steps

One trains a step-level reward model that judges retrieval quality independently of the final answer, the other teaches systems to decline when the retrieved evidence is insufficient.

3 sources