Polylog
The Polylog AI Intelligence Brief

Morning Edition · Sunday, July 19, 2026

AI Inference Prices Split Roughly 100-to-1 as OpenAI Pushes a New Value Metric

Markets

AI Inference Prices Split Roughly 100-to-1 as OpenAI Pushes a New Value Metric

Chamath Palihapitiya pegs a million tokens at about 56 dollars from Anthropic and 50 cents from Chinese models, while OpenAI's finance chief tells buyers to stop counting tokens.

3 sources
Plausible
White House Moves to Decide Who Gets Frontier Models From OpenAI and Anthropic

Geopolitics

White House Moves to Decide Who Gets Frontier Models From OpenAI and Anthropic

Reporting says a proposed "Gold Eagle" clearinghouse could gate early access, weeks after the government pulled Anthropic's Fable 5 and Mythos 5 offline over security concerns.

2 sources
Plausible
Meta Opens Its First Paid Model API With Muse Spark 1.1

Tech

Meta Opens Its First Paid Model API With Muse Spark 1.1

The million-token agentic model prices at 1.25 and 4.25 dollars per million input and output tokens, putting Meta directly into the coding-agent market against Anthropic and OpenAI.

2 sources
Li Auto's 35B Mach-Mind-4-Flash Reports Matching 120B-Class Models on Math and Code

Tech

Li Auto's 35B Mach-Mind-4-Flash Reports Matching 120B-Class Models on Math and Code

The mixture-of-experts model activates just 3 billion parameters and claims 92.7 on AIME 2026 and 80.9 on LiveCodeBench, gains it credits entirely to post-training.

2 sources
Plausible
Meta's Brain2Qwerty v2 Decodes Typed Sentences From Brain Signals at 61% Word Accuracy

Tech

Meta's Brain2Qwerty v2 Decodes Typed Sentences From Brain Signals at 61% Word Accuracy

The non-invasive pipeline reads magnetoencephalography recordings, a large gain over prior surgery-free methods and a step toward assistive interfaces without implants.

2 sources
Anthropic and Rivals Propose a Shared Scale for Rating AI Jailbreak Severity

Tech

Anthropic and Rivals Propose a Shared Scale for Rating AI Jailbreak Severity

The draft Cyber Jailbreak Severity scale runs CJS-0 to CJS-4 across four axes, published as Anthropic restored Fable 5 and Mythos 5 after a 19-day export-control block.

2 sources
Plausible
Meta Ships Muse Image and Previews Muse Video as Agentic, Editable Generators

Tech

Meta Ships Muse Image and Previews Muse Video as Agentic, Editable Generators

Muse Image runs as an agent that invokes tools and self-refines, adds region-specific editing, and embeds a crop-resistant watermark, while Muse Video ranks third on Arena's text-to-video board.

2 sources
ETH Zurich Demonstrates a Quantum Computer With Mechanical Working Memory

Tech

ETH Zurich Demonstrates a Quantum Computer With Mechanical Working Memory

A superconducting qubit acts as processor while vibrating acoustic resonators store quantum states, a design the team says packs more memory into a chip a few millimeters across.

2 sources
Anthropic Bundles Fable 5 Into Max and Team Premium at Half Quota

Markets

Anthropic Bundles Fable 5 Into Max and Team Premium at Half Quota

From July 20 the model enters top subscription tiers with access capped at 50 percent of the full quota, while lower tiers keep credit-based access and a one-time 100-dollar credit.

2 sources
Anthropic's Economic Index Puts a Number on AI's Labor Effect, and It Is Modest

Macro

Anthropic's Economic Index Puts a Number on AI's Labor Effect, and It Is Modest

Its headline estimate of 1.8 points of annual US productivity growth falls to roughly 1.0 point after discounting for task success, with computer and math work dominating usage.

2 sources
Anthropic Solicits the Public's Hardest Questions About AI

Tech

Anthropic Solicits the Public's Hardest Questions About AI

The lab says it will publish its reasoning as it addresses submitted questions, a legitimacy exercise as scrutiny of lab narratives intensifies.

1 source
A New Local Speech-to-Text Library Aims to Widen On-Device Transcription

Tech

A New Local Speech-to-Text Library Aims to Widen On-Device Transcription

transcribe.cpp offers drop-in support for whisper.cpp with Rust bindings and numerically validated models, targeting cross-platform distribution of offline speech recognition.

1 source