|
|
|
| // WEEKLY 05 · JUL 08 2026 |
|
For three weeks the story here was the best AI getting locked away, by governments, export controls, and gated previews. This week it flipped. Anthropic put its most capable everyday Claude in your hands, on your phone, and gave it the ability to actually do things. There is also a prompt below that turns a pile of messy notes into a clean plan.
|
|
| // SECTION 01 · NEWS |
|
|
Claude can now run your email and calendar
On July 7 Anthropic brought Claude Cowork to web and mobile and switched on Microsoft 365 write tools. Claude can now draft and send email, manage your calendar, and create or update your files, and a Cowork task keeps running in the cloud after you close your laptop, so a 6 a.m. Monday client prep can finish before you reach the office. Running underneath is Claude Sonnet 5, released June 30, which can hold about a 700 page book in memory at once and costs roughly half of what the frontier used to.
Why it matters: for three weeks the thing between you and the best AI was a government sign-off. This week the most useful Claude is not just something you can open, it can act for you, from your pocket, at half the old price. The catch worth knowing: write access to your inbox and calendar has to be turned on by an admin first, and you should watch what you let it send on its own. Source.
|
|
› The engine, in numbers. Claude Sonnet 5 landed June 30 with a 1 million token context window, up to 128k tokens of output, at an intro price of $2 in and $10 out per million tokens through August 31, then $3 and $15. One catch: a new tokenizer counts about 30 percent more tokens for the same text, so check your real usage before you assume it is cheaper. Source.
|
|
› OpenAI is about to make GPT-5.6 fast. The Sol, Terra, and Luna family is still in a gated preview with broad access expected soon, and OpenAI says it will serve the flagship Sol on Cerebras hardware at up to 750 tokens per second this month, about ten times a typical GPU setup, aimed at agents where slow responses have been the blocker. Source.
|
|
› Google set a date for Gemini 3.5 Pro. After missing its June window, Google now points to July 17 and a full architecture rebuild, promising a 2 million token context window and stronger reasoning. As the second week of July begins it is still preview only. Source.
|
|
› Anthropic is now outrunning OpenAI. Fortune reported on July 2 that OpenAI is slowly losing ground to Anthropic and Google, as Anthropic's bet on enterprise pushes its revenue ahead. The safety focused underdog is now the one to beat. Source.
|
|
|
| // SECTION 02 · THE TIP |
|
|
Most people hand AI a one line request and take whatever comes back. Add one sentence instead: "before you answer, interview me one question at a time until you have what you need, then do it." The model stops guessing and starts pulling the specifics only you know, your audience, your constraints, the outcome you actually want. You answer a few quick questions and the result comes back tailored instead of generic. It is the difference between a stranger's guess and a good assistant who asked first.
|
|
Try it tonight: take any request you would normally rush, add "interview me first, one question at a time," and answer honestly.
|
|
| // SECTION 03 · THE PROMPT |
|
|
For the days your head is a mess of notes, half thoughts, and things you cannot lose. Paste the whole pile in and let it sort your brain into a plan.
|
|
|
|
You are my thinking partner. I am going to paste a messy brain dump: notes, half finished thoughts, tasks, and worries, in no order. Do four things:
1. SORT: group everything into a few clear themes and name each one.
2. SUMMARIZE: give me a short plain summary of what is actually going on.
3. ACT: pull out a simple next action list, most important first.
4. FLAG: tell me anything I seem to be avoiding or missing.
Keep it calm and clear, like a smart friend helping me think.
Here is the dump:
[PASTE EVERYTHING ON YOUR MIND]_
|
|
|
Try it tonight: open your notes app, paste the messiest page you have, and let it turn the pile into three themes and a next step.
|
|
| // SECTION 04 · DECODE |
|
|
The small chunks of text AI reads and writes in. A token is roughly a short word or a piece of one, so "newsletter" might be two tokens and "the" is one. It matters because almost everything is measured and priced in tokens: how much you can paste in, how much it can write back, and what you pay. When a model changes how it splits text into tokens, the same message can suddenly cost a little more or less even if nothing else changed.
|
|
Example: Claude Sonnet 5's new tokenizer counts about 30 percent more tokens for the same text, so a request can cost more than the lower per token price first suggests. Same words, more tokens.
|
|
| Two more places I post the stuff that does not fit here. |
|
|
|
| ● STATUS · ONLINE · OPUSJAKE |
© OPUSJAKE // CRT V1 |
|
You're reading the web version of OpusJake Weekly. Subscribe to get the next one in your inbox.
All issues · OpusJake, 6620 W Fort St, Detroit, MI 48209
|
|