Another Wednesday, another selection of AI news and resources to help you become more AI native. This week:
Anthropic makes Sonnet 5 the default model, priced to run agents
Consulting firms admit the billable hour is on the way out
Meta reads brain activity into typed sentences, no surgery required
A no-code Claude plugin that runs your whole morning for you
Next Gen
Washington now gates the frontier and open weights just won.
Whats happening: Twice in three weeks the White House stepped between an AI lab and its customers. Anthropic was given 90 minutes to pull Claude Fable 5 offline on national-security grounds. Days later, OpenAI limited its GPT-5.6 Sol launch to roughly 20 Trump-administration-approved partners. A new executive order routes the most capable US models through up to 30 days of federal review before release.
Why it matters: For two years, AI competition was a product race. Smartest model, cheapest tokens, fastest ship date. This month it turned into a permit race. If a US frontier model can be pulled in 90 minutes or held behind a customer whitelist, the version your team can actually run is no longer up to the vendor. The safest place to build production workflows is on models that cannot be revoked. Open weights, running on your own infrastructure or through a routing layer, are no longer just the cheap option. They are the reliability option.
In the wild: Every team that lost access to Fable 5 or GPT-5.6 Sol reached for a substitute. Some picked Sonnet 5. Others reached for open-weight alternatives like DeepSeek and GLM. OpenRouter published a June guide on which open-weight models are worth running in production. The awkward twist is that Chinese labs are the biggest short-term winners of a policy meant to slow them down.
Looking ahead: Expect more restrictions, not fewer. If a meaningful part of your product depends on a single frontier API, price in a 30-day supply shock. The concrete move this week is small. Pick one workflow you rely on daily, benchmark it against two open-weight models on your own hardware or through a router, and make sure you have a switch you can throw before you need it. The founders who wait for the next Fable 5 moment will be the ones who cannot ship for a month.
AI First
Build Your Own Claude Plugin (The Daily Driver)
Turn Claude into a personal morning assistant that reads your inbox, scans Slack, and hands you a prioritized to-do list before you sit down.
Ultrathink published a full walkthrough this week showing how a non-engineer built exactly this, and it is the easiest way to see what plugins actually are: a folder of plain-English text files. Here is the setup.
Make a folder on your computer. That is the plugin.
Write three memory files: About Me, Brand Voice, Preferences.
Build one skill at a time. Start with email triage, then add Slack and the to-do list.
Chain the skills together so a single "morning brief" command runs them in order.
Schedule it to fire at 9 a.m. every weekday.
Pro tip: Seed Brand Voice with three or four real writing samples of yours. Claude imitates faster than it guesses.
If you’ve been meaning to graduate from one-off prompts to actual automation, this is the shortest on-ramp we have seen. Founders, ops leads, and anyone drowning in inbox will get the most out of it.
AI News
Anthropic released Claude Sonnet 5 as the new default for Free and Pro users.
Sonnet 5 replaces the older mid-tier as the default across Claude plans, Claude Code, and the API. Anthropic says it performs close to Opus 4.8 on agentic work at roughly a third of the price. Intro API pricing is $2 and $10 per million input and output tokens through August 31, then $3 and $15. Early testers praised its follow-through on multi-step bug fixes, and one non-coder built five working web apps in ten minutes.
Consulting firms started pricing on outcomes, not hours.
The WSJ reported that firms are moving away from hourly billing as AI compresses the work. McKinsey now generates around a quarter of its global fees from outcome-based pricing. BCG projects AI-tied revenue rising from 20% of 2024 revenue to 40% this year, and Bain says AI and tech-enabled work is already 30% of its consulting business, headed toward 50%. If AI lets a team finish a 40-hour project in 10 hours, clients stop paying for the 40. Billable hours are the first casualty.
Meta's Brain2Qwerty v2 decoded typed sentences from brain activity with 61% word accuracy.
Volunteers wore a magnetoencephalography helmet, no implant needed, and typed memorized sentences. The best participant hit 78%, approaching results from surgical brain-computer interfaces. The training set was 22,000 sentences across nine people. Meta open-sourced the code and launched a $5M Digital Brain Project fund. Still slow, still lab-only, but the accuracy gap with implanted devices is closing.
Reply to this email if you have any questions or further content you want covered.
Talk soon,
Cam