DeepSeek V4.1-Flash at 20 cents, Anthropic's workforce and ChatGPT ads

By AI Secret (Magna & Ben) · October 19, 2026.

The AI Secret newsletter includes, among others, three stories. All three revolve around the cost of AI agents, their effect on employment and the monetization of assistants. Links to the original articles do not appear in the body of the email, so the text is summarized as received.

DeepSeek sells wishes for 20 cents. On October 17, China's DeepSeek released V4.1-Flash-1017 as an open-weights model. Its launch page is titled "Wish-Driven Programming Made Real". As a demo, it shows a recreation of Portal 2 that the model built in the Godot game engine.

According to DeepSeek's own coding benchmark, the model scores 77% at under $0.20 per task. Its predecessor cost $0.50. The newsletter claims that Claude Opus 5 and GPT-6 Astra cost between $2 and more than $10 per task and score lower. It stresses that there are no independent tests yet.

DeepSeek also opened its search engine as an API at $0.75 per 1,000 calls. It will be free inside Claude Code and Codex until October 31. Its API accepts Anthropic and OpenAI formats, so migrating Claude Code or Codex requires changing a single URL. The newsletter's reading is that from October 31, when DeepSeek starts charging for search, Claude Code will become its "cash register". It argues that Anthropic and OpenAI will have to explain what the remaining $9.80 of each "wish" buys.

Anthropic and the pulled-up ladder. On October 16, Anthropic studied its own workforce and described itself as a possible "leading indicator" for the economy. The share of juniors among technical hires fell from 27% to 9% since early 2025, while headcount tripled in a year.

Its teams observed that Claude could increasingly take over writing code and catching bugs. In August, agents run by employees logged six times their human hours. Existing staff were given an early-warning system that flags at-risk positions and moves them first into new roles.

The newsletter's analysis argues that Anthropic applies a "humane" automation that protects those already on the payroll. According to the text, agents spend around 80% of session time waiting on a person. That is why the scarce resource is senior judgment, which is forged in junior roles. Anthropic needs other companies to keep training the seniors it hires, while selling them the agent that lets them stop doing so. The newsletter notes that staff who encode their expertise for Claude "don't do it for free", and observes that the graduates who were never hired never got that offer.

OpenAI copies Meta's advertising machine. 404 Media, a tech investigative outlet, bought an ad on ChatGPT reading "ChatGPT isn't private" and asked OpenAI to show it to people having conversations about privacy. OpenAI did.

Advertisers can also upload their customers' email addresses and phone numbers. OpenAI finds the matching ChatGPT users and shows them the ads, without telling the advertiser who matched. Turning off ad personalization does not remove advertising, because ads are still chosen based on the ongoing conversation.

According to the newsletter, OpenAI rebuilt Meta's system piece by piece, down to the product names: custom audiences, Pixel and Conversions API. The difference lies in the data. Meta has to infer what you want from "likes" and scrolling, while ChatGPT users type it out in full sentences. Meta at least hashes uploaded contact lists, and OpenAI does not say whether it does. Image generation got ads this month and OpenAI says more formats are coming.

The newsletter recalls that in 2024 Sam Altman called the mix of ads and AI "uniquely unsettling" and a "last resort". It adds that this last resort already brings in $1 billion a year.