DeepSeek Search and Fetch APIs, HybridAgents, Anthropic's fund

A thinner day for model launches, but four primary pages are worth reading: one tool release, one paper and two Anthropic posts. Every figure below comes from the page it links to, and I did not run any of it.

DeepSeek adds Search and Fetch

DeepSeek's docs now describe a Search API and a Fetch API, both raw JSON over the same bearer authentication as its model API. Search takes a query of up to 128 characters and returns between 1 and 50 results, with an optional flag to include full page content. It supports operators such as site:, lang:, after: and before:, forwarded to DeepSeek's in-house search engine. Fetch takes one URL and returns the page as Markdown.

The pricing is the useful part. Both are billed per successful request, $0.75 per 1,000, not by tokens or result count. A request only bills if it returns HTTP 200 with a real result, so empty searches, failures and timeouts cost nothing. The docs also say server-side tool calls made through DeepSeek's coding integrations are free for now, with per-request billing starting on 1 November 2026.

For anyone building agents, a flat per-call price turns web access from a variable cost into a line you can forecast.

HybridAgents: think in latent space, act in text

The paper Hybrid Reasoning Agents, submitted on 13 October, tackles a split in agent design. Text reasoning is transparent and works with tools, while latent reasoning is cheaper but cannot call tools and is hard to inspect. The authors add a lightweight link that lets one model switch between the two modes during a single generation, plus a working memory that carries the hybrid context across turns.

They report tests at three model scales on 11 benchmarks covering math, science, medicine, search and agentic reasoning. Against text-based and latent baselines, they claim a 10.4% improvement, 35.5% to 54.7% fewer tokens, 1.8x to 2.6x faster inference and 2.7x to 3.7x lower cost. These are the authors' numbers and I have not reproduced them, but the claim that latent thoughts stay inspectable is the part I would test first.

Anthropic's public benefit fund

Anthropic says it will commit 5% of its shares to a new Anthropic Public Benefit Fund, subject to stockholder approval. Another 2% can be added at the board's discretion for every 50% rise in company value above a set baseline. The fund is a wholly-owned LLC rather than a 501(c)(3), which the post says allows flexibility between forms of giving, and none of the equity will go to lobbying or political giving. Anthropic adds that roughly 20% of its shares are now dedicated to the public good once founder pledges and employee matching are counted.

Employment inside Anthropic

A separate research post studies Anthropic's own workforce. The share of junior technical hires fell from 27% to 9%, and by August interactive agent hours outnumbered employee hours six to one. Output per person rose as people adopted agents, which moved the bottleneck to human attention: the authors say reading all the output of their Claude agents would take half the work week for staff outside engineering.

What I'd watch next

Whether other labs price agent tools per call the way DeepSeek has, what happens to DeepSeek's costs after 1 November, and whether HybridAgents code and data appear in a form others can run. I also want to see the fund's governance details Anthropic says are coming.