The Brief
New briefs daily around 7 AM Eastern
Fireworks cuts a 30-turn agent task from one minute to eight seconds
Agent loops wait on the model at every turn, so a preview tier at 1,500 tokens/s turns minute-long runs into seconds. Work that can wait gets the opposite trade: a background mode at half price, plus prompt caching that lasts days.
- 01Read
Vals AI: low reasoning effort keeps most of the score at half the cost
At the lowest setting, four of seven frontier models kept at least 80% of their Vals Index score for half the cost or less, and Opus 5.5 kept 80% at 16% of its cost. The default effort level sets a large part of the bill.
- 02Read
SemiAnalysis: SpaceX targets about 13GW of AI datacenter power in 2027
SpaceX has 3.6GW under construction and over 9GW of identified capacity, which SemiAnalysis sets against Google's historical buildout. A single new buyer at that scale changes how much compute exists, and compute supply feeds into token prices.
- 03Read
Artificial Analysis moves six job indexes to v2.0; Opus 5.5 leads all
The finance, legal, engineering and other occupation indexes now weight non-hallucination and agentic tool use more heavily. A model that answers wrongly with confidence now loses more points in these domain rankings, and Opus 5.5 tops all six.
- 04Discuss
Honnibal: Astra's Qwen math gain relied on Python, failed on Gemma
Yesterday's coverage reported Astra editing weights to fix Qwen. This rebuttal sorts each experiment by raw weights, Python or network runs, and finds the math gain came through tooling. Simon Willison also got a 27B model on a laptop to read weights.
- 05Read
Mercor and Google release engineering exam; best agent scores 48.3%
The 85 tasks come from mechanical, electrical, civil and structural work done by licensed engineers. The best setup fails about half of them while spending roughly 27M tokens and 63 minutes per task, so one run is both costly and unreliable.
- 06Read
IBM: one open model completes 30% to 81% of AppWorld tasks by harness
Across 90 AppWorld tasks only the harness changed, yet completion moved by 51 points. That spread comes from the wrapper alone, and IBM built CUGA, one of the three harnesses compared, so the margin needs outside replication.
A dated brief from the vibe-coding frontier. Today’s Intel.