A good chunk of the AI coverage sitting at the top of search results this weekend is describing a world that stopped existing on August 10. Today’s five stories share a shape once you line them up: the announcement ships, the thing you can actually use ships later, and a lot of teams are budgeting against the press release.
Anthropic Cancelled the Sonnet 5 Price Hike and the Coverage Never Caught Up
Verdict: matters.
On August 10, Anthropic’s Claude account said it is making Sonnet 5’s introductory pricing permanent: the model launched in June at $2 per million input tokens and $10 per million output tokens through August 31, and that price stays. The company’s own pricing documentation now carries the note in plain text, confirming that the previously scheduled increase to $3/$15 on September 1 will not occur.
I run my whole content operation on Claude, so I went looking for the migration plan I would need before month end. There isn’t one. There is no deadline.
That is not what you would conclude from the search results. Datafloq is still running a headline announcing that Anthropic confirms a 50% price rise on September 1. Enterprise DNA has a piece telling you that you have 30 days to audit your AI bill. FinOps LLM has an entire research page on the August 31 intro-pricing deadline and the 50% hike behind it. All of that was accurate in July. None of it is accurate now, and it is what ranks when an operator searches the deadline.
Here is the part almost nobody is doing the math on. Sonnet 5 at $2/$10 against Sonnet 4.6 at $3/$15 reads like a 33% cut. It is not, because the same Anthropic pricing page carries a second note: Claude 4.7 and later models run a newer tokenizer that produces roughly 30% more tokens for the same text, while Sonnet 4.6 and earlier use the previous one. Same paragraph of English, more billable tokens.
Per unit of actual text, the honest comparison looks like this:
- Sonnet 4.6: $3.00 in, $15.00 out
- Sonnet 5: roughly $2.60 in, roughly $13.00 out
Call it 13% cheaper rather than 33%. Still cheaper, still worth the move. But if you built a Q4 forecast on the sticker difference, you will come up short, and for a reason that has nothing to do with the price change everyone was watching. Delete the September 1 migration ticket, then re-run the cost model against measured token counts from your own traffic instead of the per-million numbers.
The contrast is sharp: DeepSeek moved the opposite way this week, raising prices in the same stretch that Anthropic locked its own down.
Z.ai Shipped an Open-Weights Model Without the Weights
Verdict: breaks your stack.
Z.ai released GLM-5.3 on August 14. Per SiliconANGLE’s writeup, it is a 753-billion-parameter mixture-of-experts model with a 1 million token context window, built on the same base as GLM-5.2 from mid-July with the gains coming from post-training alone. It posted the highest score of any open-source model on Terminal Bench 3.0, performed 50% better than its predecessor on Z.ai’s internal coding benchmark, and beat Claude Mythos 5 on the CyberGym security benchmark while falling behind Anthropic’s flagship on two others.
The weights are not out. Z.ai plans to publish them to Hugging Face under an open-source license within two weeks, after safety evaluation and hardening. The reason sits in the model’s own results: it surfaced more than 2,400 vulnerabilities across 269 software projects, roughly half of them rated medium severity or higher.
I do not think the delay is wrong. Two weeks of hardening on a model that good at finding exploits is a defensible call, and I would rather vendors make it than not. What is wrong is calling this an open-weights release on day zero.
If your architecture assumes self-hosting, “open weights, in about two weeks” is not a dependency you can schedule against. It is a maybe with a date attached. You can hit the API today. You cannot run it on your own hardware, cannot audit the weights, and cannot pin a checkpoint, which are the three reasons most teams want open weights in the first place. Any private-deployment evaluation you slotted into this sprint just moved to next month.
OpenAI’s Answer to Apple Is Due Tomorrow
Verdict: matters.
Apple escalated its trade secrets case on August 4, asking for a preliminary injunction that would bar OpenAI from using the allegedly misappropriated information while the case proceeds. OpenAI came back a day later with a 31-page motion to dismiss, and Axios reported the filing argues Apple never adequately described what it claims as secret or showed misappropriation by anyone. OpenAI also published its own response, bluntly titled Apple is getting this wrong.
Apple should not be permitted to use a baseless and pretextual lawsuit to make up for its shortcomings in the market for talent and retaining its employees, and its failures to integrate AI into its products.
That is from the motion itself, and it tells you the temperature here. OpenAI further contends Apple reviewed employees’ personal iMessages stored on corporate devices after directing staff to use personal iCloud accounts for work.
The operator-relevant date is August 17, tomorrow, which is OpenAI’s court-ordered deadline to respond to the injunction request. Arguments are set for October 1. A trade secrets verdict lands in 2027 or later and changes nothing about your year. An injunction lands in weeks and can pull capabilities out from under a product you already shipped. Tomorrow’s filing is the one worth reading.
Anthropic’s “First Profit” Is a Slide in a Fundraising Deck
Verdict: marketing.
The numbers going around are $10.9 billion in Q2 revenue and $559 million in operating income, Anthropic’s first operating profit, arriving roughly two years ahead of its own guidance. Those are real figures. They are not results.
As PYMNTS laid out, they come from financial projections reviewed by The Wall Street Journal and shared with investors during an ongoing funding round. Projections, in a raise. Anthropic is private: there is no filing, no audit, and no earnings call behind any of this. Plenty of coverage this week has rendered it as “Anthropic reported” and “Anthropic achieved,” which is a different claim than the one the reporting supports.
The caveat sits in the same reporting. Anthropic cautioned it may not sustain profitability across the full year given planned spending tied to computing infrastructure. The company being described as profitable is telling its own investors not to expect that to hold.
The growth underneath is genuine, with Q1 at $4.8 billion against Q2’s projected $10.9 billion. And put this next to the first story: the same company, in the same month, cancelled a 50% price increase while showing investors a first profit projection. That is not a business that thinks it needs the margin. That is a business buying the next two years of developer defaults, and it is the more interesting read than the profit headline.
Everyone Is Writing About an S-1 Nobody Has Read
Verdict: marketing.
OpenAI submitted a confidential draft registration statement to the SEC on June 8 and said so on its own site, which Fortune covered accurately at the time. That is the entire public record. No ticker, no exchange, no date, no pricing, no roadshow. The document is not on EDGAR.
Search the topic now and you get valuation breakdowns, how-to-invest guides, expected-date tables, and at least one tutorial on reading OpenAI’s S-1 with ChatGPT. That last one is a guide to reading a document that has not been published.
For an operator there is nothing to do here. A confidential filing changes nothing about OpenAI’s pricing, rate limits, or model availability. The day the public S-1 lands is the day this becomes real, because that version carries revenue, compute commitments, and the Microsoft arrangement in writing. Until then it is a content category, not a story.
What Actually Has Receipts
Three of today’s five items share the same defect. Anthropic’s profit is a projection, Z.ai’s open weights are a promise, and OpenAI’s S-1 is sealed. In each case the announcement arrived and the artifact did not, and the coverage treated those as the same event.
Two things this week came with receipts you can act on: a pricing page you can read, and a court deadline you can put in a calendar. Plan against those. The rest is a slide.