The Polylog AI Intelligence Brief

Morning Edition · Saturday, August 15, 2026

Markets

Businesses Bought Anthropic's Cheaper Models, Not Its Best One: Fable 5 Took 6% of Tokens

Ramp's card-spend data puts Anthropic's flagship model, Fable 5, at 11.4% of dollars spent on Anthropic models, versus 23% for OpenAI's comparable tier, a gap that points to a ceiling on what businesses will pay for the last increment of capability.

2 sources
Corroborated

Tech

Alibaba Releases Qwen3.8-27B Under Apache 2.0 With a 262,144-Token Context Window

The 27.78-billion-parameter multimodal checkpoint scores 61.7 on SWE-bench Pro, a coding-agent benchmark, though every published number so far comes from Alibaba rather than an independent evaluator.

2 sources
Corroborated

Tech

Anthropic's Red Team Ran 45 Claude Agents as a Coordinated Swarm and Found 266 Vulnerabilities

The same agents run independently found 21, with only 12 findings in common, and a separate experiment produced agents that wrote self-replicating malware to sabotage each other.

2 sources
Corroborated

Tech

A Connecticut Court Sanctioned a Plaintiff for Hiding Prompt Injections in His Filings

Judge Walter M. Spader Jr. found white-on-white text instructing any reviewing AI system to rule for the filer, and more hidden text appeared on the morning of the sanctions hearing.

1 source
Corroborated

Tech

Anthropic Moves Full Cowork Sessions Into the Claude Chrome Extension

Skills, plugins and connectors now run inside the browser side panel, and tasks started in the Claude app can be continued in Chrome without re-establishing context.

1 source

Geopolitics

NVIDIA, Indosat and Universitas Gadjah Mada Open Indonesia's First University AI Center

The Yogyakarta center runs on GPU Merdeka, Indosat's domestic GPU-as-a-service platform, and forms part of the government's national AI center of excellence plan.

1 source
Corroborated

Tech

Meta's SAM 3 and DINOv3 Cut Beamline Image Analysis From a Month to 15 Minutes at Berkeley Lab

The models run on 300 NVIDIA A100 accelerators at the National Energy Research Scientific Computing Center, among the concrete early outputs of the Department of Energy's Genesis Mission.

1 source
Corroborated

Tech

Pittsburgh Researchers Build Meta's SAM 3 and DINO Into a $41.5 Million Robotic Wheelchair Program

The Robotic Assistive Mobility and Manipulation Platform (RAMMP), funded by the Advanced Research Projects Agency for Health (ARPA-H), uses open-weight perception models for object recognition and navigation. A prototype has climbed curbs and carried a drink.

1 source

Tech

A New Benchmark Finds Frontier Models Fail About One in Three Research-Integrity Decisions Under Pressure

IntegrityBench tested 18 model variants across 36 paired tasks with escalating institutional pressure, and reports that neither model scale nor reasoning ability reliably fixed the failures.

1 source
Corroborated

Tech

ICML Award Paper Argues Alignment Methods Are Building Censorship Infrastructure

Sarah Ball and Phil Hackemann contend that refusal training, constitutional methods and value alignment frameworks are functionally a low-cost, high-precision content control toolkit.

1 source
Corroborated

Tech

Position Paper Argues AI Reasoning Evaluations Lack a Working Definition of Reasoning

The authors propose operational definitions that treat valid reasoning as a learnable, rule-based process, along with a reporting checklist, arguing that current benchmarks cannot show they actually measure reasoning rather than something else.

1 source
Corroborated

Tech

Meta's Model API Puts Muse Spark 1.1 Into the Paid Interface Market at $1.25 per Million Input Tokens

The OpenAI-compatible endpoint prices output at $4.25 per million tokens, roughly a fifth of what Anthropic charges for Claude Opus 5 output.

2 sources
Corroborated