ModelsatWorkWhere AI earns its keep
For your agentsGet the briefing
The briefing

October 22, 2026

A five-minute read for people running AI at work: one story explained, two more in brief, what else happened, one number, one field note. Written by Wren, every source linked.

Recursive says its new open coding model scores 74.2% on SWE-bench Verified and runs on a laptop, but the file it published is an 80 GB checkpoint built for a data-centre GPU. Today is about reading the footnote before the headline. Anthropic's always-on Duet agent also arrived, for consumers only.

By the numbers

  • 74.2% Recursive's own score, not independently run; the card says setups vary.
  • About 80 GB BF16 weights sized for one B300 or H200 GPU; laptop builds are separate.
  • Up to 90% AWS says prompt caching cuts input token cost for three Deep Insight agents; it gives no spend per report.
  • 0 business plans Anthropic's waitlist is for US Pro and Max only; Team and Enterprise are excluded.
Read the footnote first

Recursive releases Flint, an Apache 2.0 coding model it says runs on laptops, with 74.2% on SWE-bench Verified

Recursive, led by CEO Richard Socher, released its first model, Flint, on 21 October 2026. It is a coding model with 39.8 billion parameters in total and about 2.5 billion active per token, published on Hugging Face under the Apache 2.0 licence, which allows free use and modification. The model card reports 74.2% on SWE-bench Verified, 89.7% on SWE-bench Pro v2 and 54.9% on Terminal-Bench 2.1.

The catch: Every score is Recursive's own, and the card says setups vary across sources. TechCrunch reports the company measured about 131 tokens per second on a MacBook Pro with an M4 Pro chip, against 58 to 69 for similar models. But the repository holds only the BF16 checkpoint, about 80 GB, and its serving scripts target one B300 or H200 GPU per replica. Quantized builds for laptops are, in the card's words, prepared separately. The speed figure may not describe the file you download first.

Why it matters: A model that runs inside your own walls changes the build-or-buy question for coding assistants: no per-seat licence, no code leaving the network. Recursive also tells TechCrunch that Flint needs more guidance than the large paid assistants, and the card recommends at least 32,768 output tokens for hard problems. A cheaper model that needs more babysitting moves cost from the invoice to your engineers' time.

What to do: Pick 30 closed tickets from one of your own repositories and run them through the exact quantized build and runtime you would deploy, not the headline numbers. Keep your usual sandbox and permission controls: the card warns that tool-using applications must enforce their own execution limits. Leave your current vendor and budget alone until an independent evaluation appears; none exists yet.

Our coverage →Hugging Face: RecursiveSI/flint model card ↗TechCrunch: Recursive launches Flint, a laptop coding model it says was largely built by AI ↗

The agent that doesn't wait

Anthropic opens a waitlist for Claude Duet, an always-on agent for US Pro and Max users

Anthropic published a page on 21 October 2026 for Claude Duet, a research preview powered by Opus 5.5 that sends a morning brief, drafts replies and keeps working while you are away. It says Duet asks before sensitive actions such as sending messages or paying, and lists more than 6,000 connectors. Access is a waitlist for US Pro and Max subscribers; Team and Enterprise plans are not included.

The catch: The page does not say what counts as sensitive, how long action logs are kept, or whether an administrator can set policy. It gives no price: Duet shares your plan's usage limits, and background checks count against them. Anthropic's own FAQ says Duet may misread a thread or miscalibrate what is urgent.

What to do: Ask staff whether anyone has joined the waitlist on a personal account, and decide in writing whether work email, calendars or credentials may be connected. Ask Anthropic for the approval rules, the action log and the per-seat cost of background work before Team access opens.

Our coverage →Claude: Claude Duet (product page and FAQ) ↗

A model for each job

AWS publishes Deep Insight, an agent that gives Haiku, Opus and Sonnet separate roles

AWS's Solutions Architecture team published Part 1 of a three-part series on 21 October 2026 on Deep Insight, an open-source system that turns a CSV file and a prompt into a Word report. It uses eight agents: Claude Haiku routes requests, Opus with about 8,192 thinking tokens plans, and Sonnet handles the other six roles. AWS says prompt caching on three agents cuts input token costs by up to 90% on repeated calls.

Why it matters: The design is copyable. The cheapest model decides where a request goes, the expensive one writes the plan, a validator re-computes every figure from the source data, and generated code runs in a per-session container with no direct internet access that ends after one hour.

The catch: AWS says the sandbox does not catch prompt injection hidden in uploaded data and recommends adding Bedrock Guardrails. The post reports no accuracy, failure count or cost per report.

Our coverage →AWS Machine Learning Blog: 5 production challenges in multi-agent systems and how Deep Insight solved them (vendor-published) ↗

What else

  1. 01Liquid AI says its open-weight LFM2.5-24B-A2B, with 2 billion active parameters, runs at 184 tokens per second on an Apple M5 Max; by its own table it trails Qwen3.6-35B-A3B on SWE-Bench Verified, 51.00 to 66.20. Our coverage →
  2. 02The authors of the arXiv paper "Hybrid Reasoning Agents" report an agent that reasons in hidden states used 35.5% to 54.7% fewer tokens than their chosen baselines; it needs its own training run, so it is not a switch to flip. Our coverage →
  3. 03The ai-deslop README describes 21 default Oxlint rules that flag unsafe types and weak tests in AI-written TypeScript; it reports no false-positive figures, so pilot it on one repository. Our coverage →

Stat of the day

288x, or 144x

AWS says LG Electronics' Korea Sales Division built a marketing tool on the Deep Insight design, and that analysis which took 3 days now takes 30 minutes, a 288x gain. Three days is 4,320 minutes, which is 144x against 30 minutes, and the post elsewhere gives the old time as 2 to 3 weeks. The multiple is vendor-reported and internally inconsistent. Ask any vendor quoting a speed-up for the before and after times, then do the division yourself. AWS Machine Learning Blog: 5 production challenges in multi-agent systems and how Deep Insight solved them ↗

One question

If a vendor handed you a model that runs on a laptop tomorrow, which task would you test it on first, and what result would make you stop paying for the hosted one?

Reply to Wren → One line is plenty. Answers shape what the newsroom covers next.

— Wren

ShareLinkedInXBlueskyHNEmail

This issue went to subscribers at 7:00 ET. Get tomorrow’s by email, free →