Ephemeris.
SAT 25 · JULIssue ninety-five
A frontier model that doubled its task budget without moving price, and four smaller things you can act on before Monday.
The price held. The task budget doubled.
Opus 5 landed Thursday at $5 in and $25 out per million tokens — the same tier Opus 4.8 sat on — while more than doubling its Frontier-Bench score at a lower cost per task. If you run long agent loops, the line to read twice is in the API notes: you can now change the tool list mid-conversation without invalidating the prompt cache.
- Frontier-Bench v0.1
- 2× 4.8 More than double Opus 4.8, at a lower cost per task.
- ARC-AGI 3
- 3× next Three times the next-best model's score.
- CursorBench 3.2
- 0.5% gap Within half a point of Fable 5, at half the cost.
Seventy percent of your routes arrive edited.
Cloudflare measured what transit providers do to the BGP ORIGIN attribute in flight and found roughly 70% of observed paths come back rewritten. Resetting ORIGIN to IGP hands the rewriter 18% more IPv4 paths and 40% more IPv6 — a thumb on the scale four steps down in route selection. Iliana Xygkou and Bryton Herdes argue the attribute should stop deciding anything at all.
The image model shipped with nothing to check.
Qwen-Image-3.0 arrived on 21 July aimed squarely at work: 4,500-token prompts, text claimed legible down to 10 pixels, twelve native languages, UI mockups and document layouts instead of portraits. It came through the chat interface only — no downloadable weights, and none of the technical report that shipped with 1.0 and 2.0 under Apache 2.0. Alibaba's own sample reel is the whole evidence base a developer gets.
- 01Rich contentPrompts up to 4,500 tokens, multi-panel layouts.
- 02Authentic detailsLegible type down to 10px; photographic skin and hair.
- 03Deep knowledgeTwelve languages, UI mockups, web-connected world facts.
- 04WithheldBenchmark table, parameter count, license, downloadable weights.
One DuckDB per tenant, awake in 300ms.
PostHog's "context warehouse" pitch is a warehouse an agent can query without an ETL pipeline in front of it. Strip the positioning and there's a real shape underneath, and it is worth copying: a DuckDB instance per organisation in a Firecracker microVM over S3, waking from idle in roughly 300ms, exposed on the Postgres wire protocol so an agent reaches it through MCP with no bespoke connector.
| Engine | DuckDB, one instance per organisation |
|---|---|
| Isolation | Firecracker microVM |
| Storage | S3 — product events sync automatically |
| Cold start | ~300 ms from idle |
| Query surface | Postgres wire protocol |
| Agent access | MCP, Slack, any Postgres client |
| External sources | 200+ via Warehouse Sources |
Show it once. It writes the skill.
Cowork in Claude for Mac now builds a skill from a screen recording. Hit record, do the task the way you always do it, narrate as you go, and Claude turns screen, clicks, keystrokes and voice into a skill definition you review before saving. That replaces the part of automation everyone stalls on: writing down steps you only know by doing.
That's Saturday.
Five picks from eighteen feeds, chosen for what a senior engineer or founder could act on this week. Two candidates were dropped as retreads: OpenAI's account of the Hugging Face evaluation incident ran here on 20 July from Hugging Face's side, and Import AI 465 leads on the AISI open-weights study that ran on 21 July.