Subscribe
Share
Share
Embed
Anthropic's workforce data points to review as bottleneck
A daily summary of what is interesting and happening in the AI industry, with a focus on what this means for people building harness experiences that are used.
Good morning, it's Sunday, October eighteenth.
In today's briefing we see DeepSeek updating its open V4.1 Flash model and offering search tools free inside rival harnesses, Replit making OpenAI its default model, and Anthropic's own data on a shrinking junior workforce.
First up - Today in the big model news;
OpenAI
Replit has switched its default model to OpenAI. CEO Amjad Masad cited cost, partnership and performance, and said, quote, "I think ChatGPT is better now." The benchmarks he pointed to have not been published. He later clarified that Claude is still selectable and Replit remains model agnostic. An unnamed insider says the Anthropic partnership is still in place, and Anthropic has made no statement. TechCrunch notes that most users keep the default, so the default decides where spend lands. Replit was recently a top-five Anthropic customer. The default slot is the contested asset, and an unpublished benchmark claim is no basis for switching your own: run evaluations on your own workloads.
In local model developments;
There continues to be a pattern of open-weight models closing on closed frontier systems. DeepSeek has released its ten-seventeen update to V4.1 Flash, which the company itself calls minor. The weights are MIT-licensed, at seven hundred and sixty-three billion parameters, with a one-million-token context in the recommended settings.
DeepSeek reports large agentic gains. Terminal-Bench 4.0 rises from about thirty-one to about forty-two against its September checkpoint. DeepSWE rises from about seventy-four to seventy-seven. And Humanity's Last Exam without tools rises from about thirty-seven to about forty-three.
Here is the caveat, and it matters: every one of those numbers comes from DeepSeek's own model card, measured against its own earlier checkpoint. The seventy-seven on DeepSWE sits above figures logged for Kimi K3 and Reflection Beam, but those may come from different harnesses, so it is a higher number, not a demonstrated better model.
The only independent reading is an Artificial Analysis score of thirty-nine for the earlier checkpoint. OpenRouter lists thirty-nine and a half for the new one, though it is unclear whether that was re-measured. Next to double-digit vendor-reported jumps, that is a small move. An independent run on the new weights would settle it, and until then, re-routing workloads on the model card alone is premature.
In the harness, tools and orchestration world;
Also from DeepSeek, new Search and Fetch APIs. One returns search results with site, language and date operators, and the other converts a page to Markdown. Both cost seventy-five cents per thousand successful requests, and only responses with usable content are billed. Calls made through DeepSeek Harness, Claude Code or Codex are free until billing starts at midnight Beijing time on November first. That puts DeepSeek's tools inside rival harnesses, and adopting them means changing an endpoint rather than a workflow. Standalone search vendors now compete with a metered add-on from a model supplier, and DeepSeek gains a view of agent retrieval traffic it would not otherwise see. Test it before the free period ends, measuring cost and result quality against your current search provider.
In other news…
Anthropic's own workforce report, via TechCrunch, says the junior share of its technical staff fell from twenty-one percent to fourteen percent since early 2025. The junior share of technical hires fell from twenty-seven percent to nine percent. In an April survey, twenty percent of employees said their work had shifted to managing Claudes, and meeting time per employee rose twenty percent year over year. Anthropic names human attention and judgment as the bottleneck, and says more code does not necessarily mean better code, though it reports no quality decline. All figures are self-reported. This extends the running storyline of AI generation outpacing human verification: the lab with the heaviest agent use is shrinking its entry-level pipeline while growing coordination load. Budget reviewer hours as the scarce input, and ask what training path replaces the junior roles being cut.
Meta has removed language from its commercial terms that said it could use any content from a business account's domains, pages or URLs, including for AI training. Nieman Lab's legal experts called it a "remarkable provision" that could allow scraping of publisher sites. Meta first said crawling was unchanged, then pulled the language, saying the terms only applied to its ads and business products. Gizmodo says Meta tweaked the terms only slightly, so the scope of the removal is unclear, and the rollback is by Meta's own account. Training rights can arrive inside ad-product paperwork, so data-use terms deserve a review on every platform update.
And Cody Wilson, founder of Defense Distributed, has unveiled the WASP, which The Verge calls the world's first AI-made gun. Wilson says iterating with an AI could speed new 3D-printed designs. A firearms historian told The Verge the design is flawed, with an unfinished trigger guard leaving the trigger exposed and an inverted grip. The generating model is not named. The capability claim looks weak, but the episode still pressures model providers' weapons-design policies, such as Anthropic's usage update that codified weapons software bans. Expect journalists and regulators to test whether those clauses are enforced in practice.
That's the briefing. Have a great day, and don't forget to subscribe.