Fresh stories
Google releases Gemini 4 Hydrogen for coding agents
Google made Gemini 4 Hydrogen available in Gemini, Antigravity, and AI Studio. Demos show Carbon coordinating Hydrogen workers, and OpenClaw has added support.
LlamaIndex releases ParseBench-2 for parsing more than 4,300 pages
LlamaIndex's ParseBench-2 tests document parsing across more than 4,300 pages. It weights over 330 error types against human severity judgments and provides dimension-specific scores.
Will DePue reports Blackwell fingerprints in six Chinese model weights
Will DePue reports Blackwell-compatible initialization matches in six Chinese models, validated against 24 models with known hardware. The findings do not establish chip location or prove export violations.
Google releases Gemini 4 Hydrogen for coding agents
Google made Gemini 4 Hydrogen available in Gemini, Antigravity, and AI Studio. Demos show Carbon coordinating Hydrogen workers, and OpenClaw has added support.
Gergely Orosz reports Microsoft email failures lasting at least 12 hours
Gergely Orosz reports widespread Outlook.com and Microsoft 365 email failures lasting at least 12 hours, with servers failing again during recovery. CPU exhaustion was reported, but the cause remains unconfirmed.
Top storiesthis week
vLLM reports 10.2× tokens per second per chip on Rubin
vLLM reports 10.2× total tokens per second per chip, including cached tokens, versus GB300 on AgentX. SemiAnalysis reports 6.5× better inference performance per dollar for the Rubin submission.
Theo finds 23 wrong claims in Gemini 4 Carbon code analysis
Theo's compiler architecture review records 23 wrong claims out of 79 for Gemini 4 Carbon and 24 out of 36 for Gemini 4 Argon. His published evaluations also compare models on eight real T3 pull requests.
Tools raise average VIALS-AA scores from 25% to 47%
Artificial Analysis reports that tools raise average VIALS-AA scores from about 25% to 47%. They increase task costs for every model, with gains and spending varying substantially by model.
Anthropic releases Fable 5.2 with API prices of $8/$40 per million tokens
Fable 5.2 is available on all paid plans and generates output up to 30% faster than Fable 5.1. API prices are $8 per million input tokens and $40 per million output tokens.
Reports: Fireworks reportedly suffered a customer-data breach
Reports citing The Information say a Fireworks breach exposed customer billing records, some datasets and incomplete model weights. The intrusion reportedly came through a third-party provider hosting research systems.