July 6, 2026
Anthropic just undercut its own best model
Cheaper agents, the 510 billion dollar number fooling founders, and a fusion machine that lit a bulb.
Two shifts this week pull in opposite directions. AI got cheaper to run, and harder to fund unless you are one of two companies. Here is what it means for your money and your freedom.
TECH: AI in production
Anthropic shipped a model that competes with its own best one, for less.
On June 30 Anthropic launched Claude Sonnet 5. The company calls it its most agentic Sonnet yet, built to plan, use a browser and a terminal, and run on its own. Anthropic says its performance is close to Opus 4.8, its flagship, at a lower price. Intro API pricing is 2 dollars per million input tokens and 10 dollars per million output through August 31, then it goes to 3 and 15.
For anyone running agents or batch jobs, the model you reach for by default might now be the cheap one. One catch. The headline rate is not the bill. Early users report the new tokenizer emits roughly 1 to 1.35 times more tokens per task, so the same job can cost more than the per-token price suggests.
The money play: Move your agent and batch workloads to Sonnet 5 and lock the intro rate before August 31, but count your own tokens first. An adjective saving is not a saving. A measured one is.
Sources: Anthropic | TechCrunch
BIZ: build, fund, scale
The venture headline this week is a record. The real story is who is not getting the money.
Global venture funding hit 510 billion dollars in the first half of 2026, more than all of 2025 at 440 billion. Sounds like the market is back. It is not, for most of us. OpenAI and Anthropic alone took 217 billion, which is 43 percent of the total, and over 70 percent of second-quarter money went to AI companies. Call it the two-name mirage. The record is real and it is not yours.
The week's biggest rounds prove where the money actually goes. National Grid put 1.75 billion dollars into Joulent for AI data-center power. Together AI raised 800 million at an 8.3 billion valuation for open-source inference. TwelveLabs raised 100 million for video AI. 8090 Solutions raised 135 million for software built by AI agents. Power, inference, and agent tooling. Picks and shovels, not apps.
Meanwhile Oracle told regulators on June 22 it cut about 21,000 jobs, near 13 percent of its people, and named AI adoption as a reason. More than 160,000 tech jobs are gone in 2026 so far, against about 245,000 for all of last year.
The money play: If you are not selling AI infrastructure, the flood is not for you, so raise on revenue and efficiency, not on a hot market. And when you plan 2027 headcount, read "we use AI" for what it is becoming. Code for fewer people. Point that same tooling at growth before someone points it at your job.
Sources: Crunchbase | Crunchbase | TwelveLabs | CNBC | TechCrunch
LAW: rights and fights
Europe just gave founders something rare from a regulator. Time.
On June 29 the EU Council gave final approval to a package that simplifies the AI Act. This is adopted, not proposed. The heaviest high-risk obligations move to December 2, 2027 for standalone systems and August 2, 2028 for AI built into products. One thing did not move. The rules for general-purpose AI models that took effect in August 2025 still stand, so transparency duties are live right now.
If you sell into the EU, your most expensive compliance work just slid out a year or more. Your transparency work did not.
The money play: Defer the heavy high-risk compliance spend to 2027 and 2028, but build model transparency into the product today. The deadline that already passed is the one that bites.
Sources: EU Council | European Commission
POL: policy, tax, borders
Same month, opposite direction across the Atlantic.
While Europe wrote deadlines, the United States kept its hands off. Executive Order 14409, signed in early June, sets no mandatory license and no preclearance to build AI models. The only federal ask is a voluntary review of up to 30 days before a frontier release. Call it the regulation gap. The EU governs by calendar. The US governs by suggestion.
Where you incorporate and where you ship first is now a strategy choice, not a formality.
The money play: If you ship models or model-heavy products, the lighter federal touch makes the US the faster place to launch first. Ship there, then meet the EU calendar on its slower clock.
Sources: White House
SCI: breakthroughs to build on
A startup pulled electricity straight out of a fusion machine.
On June 19 Realta Fusion ran its WHAM prototype at UW-Madison and converted plasma energy directly into electricity, enough current at about 100 volts to light several bulbs. The company says it is the first private fusion firm to do this on a real machine, and TechCrunch calls it an apparent first. This is a proof of concept, not a power plant. Commercial plants are a mid-2030s target.
Fusion is moving from a physics question to an engineering problem. The people who care first are the ones buying enormous amounts of power, which now includes anyone running large AI.
The money play: Keep direct-conversion players like Realta on a watch list, not a purchase order. The timeline is a decade out. The signal worth noting is that cheap power for compute is finally on someone's roadmap.
Sources: PR Newswire | TechCrunch
Before you go
One theme runs under all of it. Companies are saying "we use AI" when they mean "we need fewer people." At nomadtechnologist.com I build the AI workforce that runs a business without the headcount, except you point it at growth instead of a layoff memo. Kobe reviewed this issue and went back to sleep. See you next time.
Not legal, financial, or tax advice.
Get this every weekday.
Free. 5 minutes. Unsubscribe whenever.