Four things worth your attention today, and three of them are being reported wrong somewhere in the feed. Nothing got smarter this morning. What changed is what you can verify, so here is the operator’s read with the primary sources attached.
Anthropic’s September 1 Price Increase Is Cancelled, and Several AI News Sites Have It Backwards
Start here, because a handful of AI aggregators are telling you something false today. Several are reporting that Claude Sonnet 5 “reprices today” with a cut of around 17%. That is not what happened, and the primary source is a page anyone could have opened. Anthropic’s pricing documentation states that the $2 and $10 per million input and output tokens, announced at launch as introductory pricing through August 31, 2026, is now the standard price, and that the previously scheduled increase to $3 and $15 on September 1 “will not occur.”
There is no repricing event today. There is a 50% price increase that sat on every 2026 budget built around Sonnet, and it has been quietly deleted. If you build on Sonnet, tomorrow was going to hurt. It won’t.
Now the part none of the coverage mentions, printed on the same page all of it is citing. Sonnet 5 is not 33% cheaper than Sonnet 4.6, despite sticker prices of $2 against $3. Anthropic notes that Claude 4.7 and later models use a newer tokenizer producing roughly 30% more tokens for the same text, while Sonnet 4.6 and earlier use the previous one. Run the arithmetic on identical content:
- Sticker price: $2 per million input tokens against $3, so 33% cheaper per token.
- The same text through the newer tokenizer: roughly 30% more billable tokens, so about $2.60 of Sonnet 5 input for every $3.00 of Sonnet 4.6 work.
- Actual saving on the same document: closer to 13%.
Anthropic is explicit that the exact increase depends on content and workload shape, so treat 30% as a planning figure and measure your own traffic. My read: this is a real cut, Sonnet 5 is still the cheapest sane way to run production work in that family, and it is worth roughly a third of what the headline implies. Every word on this site is drafted through that family, so I remeasured my own numbers this morning rather than trusting the sticker. If you moved budget on 33%, go do the same. Matters.
The DeepSeek Round Everyone Is Reporting Was Paused a Month Ago
A cluster of outlets spent this week reporting that DeepSeek is closing in on roughly $7.4 billion at a $74 billion pre-money valuation, with an IPO filing possible by year end and a Shanghai Star Market debut in 2027. The South China Morning Post puts the round at about 50 billion yuan against a pre-money of about 500 billion yuan, expected to close before the end of August. Today is the end of August. No close has been confirmed.
Here is what almost none of that coverage mentions. This is the second round, and DeepSeek paused it in late July. Fortune reported that the company verbally told would-be backers it would not be signing investment agreements, after a leaked transcript attributed to founder Liang Wenfeng discussed DeepSeek’s reliance on Nvidia chips and China’s continuing lag behind the United States. Bloomberg, which broke that account, said it had not verified the posts were authentic.
The sizes do not line up either. When the round was paused it was described as targeting at least 10 billion yuan at a pre-money of at least 480 billion yuan. The version being reported now is roughly five times larger at approximately the same valuation. Either the raise was restructured or somebody’s number is wrong, and nobody covering it has said which.
I am not calling the round fake. DeepSeek’s June financing was real, closing at $7 billion with Tencent, CATL and China’s national AI investment fund behind it, and it lifted Liang’s net worth to $36 billion. The point is that “nears $7.4 billion” is being written as a clean upward line when the actual sequence is closed, paused, resized, and still unconfirmed on the day it was supposed to finish. If you are calibrating how fast Chinese labs are capitalizing against American ones, that difference is the entire story. Marketing.
Google’s A2A Protocol Moved In With MCP
Axios reported this month that Google’s Agent2Agent protocol, which defines how agents built by different vendors discover each other and hand off work, has moved under the Agentic AI Foundation. That is the Linux Foundation body formed in December 2025 around Anthropic’s Model Context Protocol, Block’s goose and OpenAI’s AGENTS.md, with AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI all sitting in the top membership tier.
This is the least dramatic item here and probably the most durable. The two protocols an agent developer actually has to care about, one for tools and context and one for agent-to-agent handoff, now answer to the same neutral foundation with every major lab and cloud at the table. I flagged Anthropic’s push to make its own specs the default in a recent roundup, and this is the counterweight to it. A year ago, building on MCP meant betting that Anthropic stayed friendly. It does not mean that now.
If you have been deferring agent interoperability work until the standards fight settled, it settled, and more decisively than the coverage suggests. It is telling that this got a fraction of the attention of a licensing spat. Matters.
Six Chatbots Beat Four Search Engines at Catching State Propaganda
NPR, working with NewsGuard, built 30 questions from 15 false narratives pushed by Russia, China and Iran between December 2025 and July 2026, then put them to six chatbots (ChatGPT, Gemini, Copilot, Meta AI, Grok and Claude) and four search engines. The chatbots correctly debunked the narratives about three-quarters of the time. AI-generated search summaries landed below the chatbots and above plain search, with Google’s AI Overview debunking most of the time and Bing’s failing most of the time.
This earns a slot because it puts a number on a story you have been sold for two years, that chatbots are a propaganda delivery system and search is the responsible alternative. Measured against real state narratives, it went the other way.
I am not going to oversell the good news either, and this is where I part company with the headline. Three-quarters is a failing grade for anything customer-facing. One in four is the rate at which these systems repeated or declined to challenge a narrative a nation state built on purpose, and if you have a bot answering questions on anything politically contested, that is your error rate, not NPR’s. The finding is that chatbots beat the alternative, not that they are reliable. Marketing, and the marketing here was the fear.
Still Live From Yesterday
Two stories from the weekend are still the biggest things on the board and neither moved materially overnight: OpenAI’s November 12 shutoff of Cursor’s access to its models after the SpaceX acquisition, and the Sony Music Publishing and Warner Chappell suit naming Dario Amodei personally. Yesterday’s roundup has the full read on both. Re-reporting them a day later to pad this list would waste your time.
The Pattern Today Is the Coverage, Not the Capability
Three of the four items above are cases where the repeated version is wrong or missing the part that matters. Sonnet 5 is not repricing today, and the tokenizer quietly eats most of the discount people think they are getting. DeepSeek’s round is not a clean upward line. The chatbots-as-propaganda-engine story inverted the moment somebody measured it. The one item where the reporting and the substance agree is the protocol handover, which is also the one almost nobody wrote about.
Take that as the operating lesson. The gap between the headline number and the number printed on the vendor’s own pricing page was about twenty percentage points wide today, and closing it took four minutes and one click. Open the primary source before you move budget.