// AI Briefing · October 13, 2026
AI Briefing
Today's theme is who gets to check the homework. Anthropic and OpenAI are reportedly haggling over what independent auditors may publish while a SemiAnalysis audit shows most Chinese labs barely publish safety evals at all. Meanwhile the small-model price war heats up with Microsoft and Cloudflare shipping pennies-per-million decision models, and Yandex's AI compute is literally under fire.
In this video
- 01Anthropic and OpenAI Reportedly Want a Say in What Outside AI Auditors Can PublishPolicy
- 02SemiAnalysis: Only 31 of 857 Chinese AI Model Releases Published a Safety EvaluationResearch
- 03Microsoft Launches Decision-1, a Tiny Model That Scores Choices for $0.042 per Million TokensProduct
- 04Cloudflare Open-Sources Clef-omni, Letting One Call Judge Text, Images, Audio and VideoOpen Source
- 05Yandex Loses a Third Data Centre in Four Days as Drone Strikes Hit Russia's AI ComputeIndustry
The stories
Anthropic and OpenAI Reportedly Want a Say in What Outside AI Auditors Can Publish
Reuters reports that both labs privately pushed for substantial control over what independent evaluators can make public, citing trade-secret worries, even though Dario Amodei said last month that reviewers should publish their main findings unedited. Ideas floated include lab staff shadowing review teams and barring evaluators from working for two labs at once. It matters because third-party testing is the backbone of nearly every AI safety proposal, and an auditor who needs sign-off isn't much of an auditor.
SemiAnalysis: Only 31 of 857 Chinese AI Model Releases Published a Safety Evaluation
An audit of nine leading Chinese AI developers found just 31 of 857 releases since 2021 came with a published safety evaluation, and only nine had results at launch. The median lag is 42 days, and 93% of reasoning models have none at all. SemiAnalysis argues Beijing regulates AI applications such as labeling, minors and companions, but triggers no frontier-risk duties based on compute or capability, which sharpens the contrast with US labs that publish safety reports for major launches.
Microsoft Launches Decision-1, a Tiny Model That Scores Choices for $0.042 per Million Tokens
Microsoft-Decision-1 is post-trained from Qwen3.5-9B and returns calibrated probabilities for fixed options such as yes/no, multiple choice or rubric grades, in a single pass. Microsoft claims it is about 35 times faster than GPT-6 Sol at median latency and that Xbox Research saw competitive quality at roughly 200 times lower cost. It signals a shift toward small, cheap judge models that route, verify and gate agent workflows instead of generating text. Benchmarks are vendor-reported.
Cloudflare Open-Sources Clef-omni, Letting One Call Judge Text, Images, Audio and Video
Cloudflare's Clef-omni is built on Qwen3-Omni-30B-A3B with speech generation stripped out, and returns schema-bound decisions from text, images, audio or video in a single call. A 21-second video with sound is scored in about 1.5 seconds hosted, at $0.15 per million input tokens, and the weights are on Hugging Face for self-hosting. Cloudflare also cut Clef-flash to $0.038, so the small decision-model price war is now real.
Yandex Loses a Third Data Centre in Four Days as Drone Strikes Hit Russia's AI Compute
Yandex said its Vladimir-region data centre went offline after a drone attack on October 11, following strikes on its Sasovo site, which houses two of three supercomputers used to train its AI model, and a Kaluga-region facility. Yandex Cloud is running with limited functionality. Data centres are becoming military targets, a risk for AI infrastructure anywhere that depends on concentrated compute.
Test your understanding
Quiz
1 / 8
What is the labs' stated reason for wanting control over what outside evaluators publish?
Previous digests
Oct 12Meta finally shows up to the frontier party with Muse Wonder, and everyone gets to find ou...5 stories- Oct 11Anthropic owns the news cycle again, but not with a model: it's a week of rule-writing and...5 stories
- Oct 10Today's theme is AI colliding with limits. Anthropic is owning up to Claude hacking a serv...5 stories
- Oct 9Today's digest is about AI running into the real world and the real world pushing back. Co...5 stories
- Oct 8Today is all about the stack getting cheaper, friendlier and more exposed at the same time...5 stories