The Brief
New briefs daily around 7 AM Eastern
Artificial Analysis Decision Index v2 finds a 60x cost spread
Billing, not list price, drives the gap: GLiDE v1.2 bills a document per question and costs nearly 8x Drex Max v1.1 per million tasks for 33% higher list prices. The score now rewards calibration, and private data rose from 25% to 45% of the index.
- 01Read
vLLM study finds request cancellation fails to end retry overload
With client retries on, only 12% to 38% of calls arriving after an overload were delivered. Aborted executions took 79% to 85% of computed prompt tokens, and vLLM's own counter missed a third to half of them, so dashboards understate the wasted work.
- 02Read
Audit finds biomedical fine-tunes halve median abstention to 4.2%
Across 25 fine-tuned models paired with 11 base checkpoints, median appropriate abstention on MedAbstain fell from 8.3% to 4.2% and dropped in 21 of 25 pairs. Clinical fine-tunes start out rarely saying they do not know and get worse at it.
- 03Try
Open WebUI 0.13 runs Claude Code and Codex as chat models
Coding agents now sit in the same model picker as chat models, with tool approvals in chat, and 0.13.3 lets scheduled automations run them in a saved terminal and folder. Teams already on Open WebUI get agent runs without a separate scheduler or client.
- 04Watch
AlixPartners: Claude Code contract analysis missed about $800K
On a merger review of more than 40,000 contracts, first answers looked complete but one total was off by about $800,000 and a renewal date was a month late. The causes were silent: unprocessed long contracts, unlinked amendments and access errors read as zeros.
- 05Read
Simon Willison's agent-built C SQLite extension beats Rust 2x
After Claude replaced strtod with a custom parser, the C parse_number() ran 16 to 19 ns per row against 41 to 55 ns for Rust, from a 14 KB binary versus about 300 KB. All four versions matched across 8 million fuzz checks, so the speedup did not cost correctness.
A dated brief from the vibe-coding frontier. Today’s Intel.