| Daily issue / By the Numbers |
Tuesday, August 25, 2026 |
|
|
|
|
Today's briefing
What matters today
AI’s operating choices are moving closer to the foreground: OpenAI is splitting coding work across model tiers, enterprise buyers appear to be rewarding lower-cost options, and new labor and enforcement cases put deployment risks in focus. The rest of the picture spans rare-weather planning, short-form video from business files, and specialized legal AI.
| Inside today's briefing |
| 01 |
OpenAI’s Codex can now route bounded tasks from GPT-5.6 Sol to lower-cost GPT-5.6 Luna, but developers must explicitly request it. |
| 02 |
Stanford researchers report a 19% employment gap for 22- to 25-year-olds in the most versus least AI-exposed occupations. |
| 03 |
Taiwan indicted nine people over allegations that 74 Nvidia B300 servers were diverted to Chinese buyers through intermediary routes. |
| 04 |
Alibaba’s Wan3.0 turns documents, slides, spreadsheets, and web pages into video clips up to 30 seconds long. |
|
 |
|
Lead story / launch
OpenAI Lets Codex Route Smaller Tasks From Sol to Lower-Cost Luna Workers
OpenAI has added cross-model delegation to Codex Multi Agents v2. GPT-5.6 Sol can now break down a job and assign bounded tasks to GPT-5.6 Luna, which OpenAI positions as the faster, lower-cost option, before Sol collects and assembles the result. The setup is deliberately constrained. Luna workers cannot message other agents or spawn additional workers, and the routing works only across supported Codex models. It is also opt-in: unless developers explicitly request another model, Sol continues to spawn Sol workers with the parent model’s reasoning setting and context-forking behavior. Context is the central operating constraint. When a Luna worker does not receive the parent conversation, its opening instruction must contain everything needed to complete the assignment; OpenAI recommends using `fork_turns: none` for that arrangement. Guidance also recommends keeping subagents to six to eight, though that is not an enforced limit. The update gives developers a task-level choice between a default mode presented as higher performing and a lower-cost, more specialized execution path. Its practical test is whether teams can isolate routine work cleanly enough to save time and tokens without removing the context needed for correct results.
Read full story ↗
|
 |
A tool for your workflow
Snipman
Save your best replies once, then insert complete answers wherever you work.
Reusable writing shortcuts for repetitive work
|
| |
|
|
 |
|
benchmark
Stanford Finds a 19% Employment Gap for Young Workers in AI-Exposed Jobs
An August update to Stanford economists’ labor-market research finds that AI exposure is translating into a sharper entry-level employment divide, especially for 22- to 25-year-olds. Employment fell roughly 11% since 2022 in the most-impacted 40% of occupations, versus 10% growth in the rest, with weaker hiring—not more exits—driving the difference. Automation-oriented Claude use shows the clearest negative association; augmentation is more mixed. The result is concentrated among young workers, remains correl(ation
Continue reading ↗
|
 |
|
benchmark
Anthropic’s Fable 5 Has 11% of Enterprise Spend as Cheaper Opus 5 Pulls Ahead
Ramp payment data suggests enterprise buyers are favoring Anthropic’s cheaper Opus 5 over premium Fable 5, even as Anthropic’s reported annualized revenue surged from $9 billion at the end of 2025 to more than $65 billion by late July. Fable 5 captured roughly 11% of Anthropic-model spending two months after launch, while Opus 5 moved ahead after its late-July release. The signal is spending—not total usage—and could test whether frontier-model upgrades generate incremental demand ahead of a reportedly targeted $2T
Continue reading ↗
|
 |
|
security risk
Taiwan Indicts Nvidia Manager Over Alleged Diversion of 74 B300 Servers to China
Taiwanese prosecutors have charged nine people, including an Nvidia senior manager and two Taiwan-based Supermicro employees, in a case alleging forged installation records and the rerouting of restricted AI servers to Chinese buyers. Of 130 B300 systems documented as operating domestically, 74 allegedly reached China and 56 were stopped by customs. The case could raise scrutiny of export-control compliance, inventory verification, and intermediary routes, while a proposed US law would separately address remote U.S
Continue reading ↗
|
 |
|
platform shift
Nvidia Falls More Than 2.2% Before Earnings as Seven-Day Slide Extends
Nvidia’s Wednesday earnings are becoming a test of whether recent weakness reflects a broad semiconductor pullback or company-specific concerns. Shares fell more than 2.2% near $210 Monday, extending a nearly 7% seven-day retreat, while AMD, TSMC, Broadcom and the semiconductor index also declined. Investors will seek clarity on a Bloomberg-reported plan to raise prices by up to 15% next year for AI-chip servers as memory costs rise, alongside questions about Nvidia’s OpenAI and Anthropic investments and data-cent
Continue reading ↗
|
 |
|
Four figures that put today’s product, enterprise, research, and access stories in context.
|
Signal 01
$0.05
Key figure
|
|
Alibaba’s Wan3.0 can turn PDFs, slide decks, spreadsheets, web pages, and other media into video clips up to 30 seconds long. Pricing starts at $0.05 per second at 480p.
Source story ↗
|
|
|
Signal 02
$40 million
Key figure
|
|
Thomson Reuters launched Thomson, its first proprietary legal model, as the default engine for CoCounsel’s Tabular Analysis. Built from an open-weight base, it remains part of a multimodel product, and its performance claims are company-reported pending further validation.
Source story ↗
|
|
|
Signal 03
100 Million
Key figure
|
|
GPT-BERT, trained on roughly 100 million words, beat Meta’s Llama 2 70B on one grammar benchmark despite Llama 2 using about 15,000 times more training data. The result is a data-efficiency signal, not evidence of broad capability parity.
Source story ↗
|
|
|
Signal 04
2.5 Million
Key figure
|
|
Arm and UNICEF’s AI in Play aims to support 20 AI-powered ventures, equip 200,000 young people with foundational AI literacy, and reach 2.5 million children and community members in 12 countries. These are targets, not outcomes.
Source story ↗
|
|
|
| |
Daily tool drop 5 AI tools worth knowing today |
Selected for fit, not rank |
 |
Decawork
Centralize company accounts, access, oversight, and retirement for employee-built AI agents.
Best for / IT teams governing internal AI agents
|
Open ↗ |
 |
Navigara
Maps AI coding costs and capacity gains to roadmap work, with lower-cost model routing.
Best for / Engineering leaders managing AI spend
|
Open ↗ |
 |
Cortex by SKYNETLAB
An MCP semantic-memory layer with quality gates, typed claims, contradiction tracking, and citations.
Best for / Builders adding durable agent memory
|
Open ↗ |
 |
Trama
Creates native macOS automations from plain language using AppleScript, shell, OCR, and screen awareness.
Best for / Mac users automating repeat work
|
Open ↗ |
 |
Claude Academy
Official Claude courses, developer tutorials, multi-agent workflow training, and certification paths.
Best for / Teams building Claude capability
|
Open ↗ |
|
|
|
|
|
|
|
|
|
Reader tool / From our team
Snipman
Save reusable snippets and insert complete replies without leaving your workflow.
Reusable writing shortcuts for repetitive work
|
|
|
|
|
Reader check-in
Help shape tomorrow's briefing
One click tells us what to keep, improve, or tighten.
|
|
|
|
|
Prefer one email a week? Get the essential AI moves in the Sunday Weekly Digest.
|
|
|
