The Brief — Read the Asterisk

Every headline number this week came with fine print. A price that expires in six weeks. A benchmark whose loser says the ruler is broken. A tokenizer that quietly inflates the only unit you are billed in. None of this is fraud — it is what a market looks like when the product is improving faster than anyone can price it.

For a one-person operation, the fine print is the whole story, because that is where your September bill lives. Here is the week, asterisks included.

OpenAI stopped selling one model

On July 9, GPT-5.6 shipped as three: Sol at $5 per million input tokens and $30 output, Terra at $2.50/$15, and Luna at $1/$6. OpenAI’s framing is that the number is the generation while the name is a durable capability tier — Sol, Terra and Luna are meant to advance on their own cadence. All three carry a 1M-token context window and up to 128K output tokens.

So what: a 5x spread between the top and bottom of the same generation turns model choice into a routing problem rather than a shopping one. If everything you run goes through the flagship, you are paying Sol rates for Luna work — classification, extraction, tagging, first-pass summarising. The cheapest performance win available to you this month is a rule that sends the boring 80% down a tier.

The labs are now arguing about the ruler

OpenAI put Sol at 53.6 on Agents’ Last Exam, 13.1 points ahead of Claude Fable 5. Fable 5 then beat the entire GPT-5.6 family on SWE-Bench Pro, 80% to 64.6%. OpenAI’s response was to publish an estimate that roughly 30% of SWE-Bench Pro’s tasks are broken.

So what: both results are real, and they disagree because they measure different jobs. The signal worth taking is not which lab won — it is that vendors are now litigating the benchmarks themselves, which means leaderboards have stopped working as a shortcut. Ten prompts drawn from your actual workload, scored by you once a quarter, is now a genuine edge. It takes an afternoon.

Sonnet 5’s price has an expiry date

We covered Sonnet 5 taking the default slot two weeks ago. Here is the part that matters now: its $2-in / $10-out pricing is introductory, and it runs through August 31. After that it is $3/$15 — a 50% step up. Anthropic’s own documentation also notes that the new tokenizer maps the same input to roughly 1.0–1.35x more tokens than the previous scheme, depending on content type.

So what: those two compound. Identical usage could land your September bill meaningfully north of your July one, with nothing changed on your side. Put August 31 in the calendar now, and re-baseline against real invoices rather than the price card.

The agents left the IDE

Claude Cowork arrived on web and mobile on July 7, starting with Max subscribers. More interesting than the launch is the data Anthropic published alongside it: across 1.2 million sessions from 600,000 organisations, business process work — reports, checklists, spreadsheets — accounted for 33.4%. Content and copywriting took 16.4%. Software development was 8.7%.

So what: the agent story was sold as a developer story, and the people actually using it are not developers. Under 9% of it is coding. The rest is precisely the administrative drag a solo founder absorbs personally because hiring for it never pencils out. If you have been waiting for agents to become useful to you specifically, the usage data says you are the median user, not the edge case.

An agent ran its own funding round

Lyzr, a three-year-old agent startup in Jersey City, closed a $100M Series B at roughly a $500M valuation — and had its own agent, SivaClaw, run the raise. It fielded questions from more than 130 investors, drafted investment memos, and tracked which slides backers lingered on. Total interest reportedly reached $400M.

So what: two lessons, one durable and one temporary. The durable one is that the strongest demo is your product doing the hard thing in public, on you first. The temporary one is that there is enough capital chasing AI right now that a founder with traction barely has to leave their desk. Treat the second as a window, not as weather.

The 5% question

OpenAI floated giving the US government a 5% stake — about $42.6B against the $852B valuation it set in March — pitched as a sovereign vehicle modelled loosely on Alaska’s Permanent Fund, and potentially extended to other labs. Nothing is agreed, and Anthropic says it has not discussed it.

So what: mostly an investor story. But if your product sits on a lab’s API, sovereign equity becomes a new line in what “vendor risk” means. Concentration risk used to be purely commercial. It is quietly becoming political.

The through-line is not that anyone lied this week. It is that headline numbers now have a half-life: the price is introductory, the benchmark is contested, the token is not the token you used to buy. Read the asterisk — it is where the margin is.