Tap to play. It never plays by itself, and the captions are in the picture.
DeepSeek goes cheap, and websites push back on agents. DeepSeek says V4.1 Flash 1017 is live. OpenRouter lists it at fifteen cents per million input tokens. The Verge reports Delta, JPMorgan Chase and Schwab are blocking AI agents. AWS says evals caught its sales agent inflating revenue figures before launch. Google Research says lab agents lifted aptamer hit rates to 42.1 percent. Databricks says domain hints cut budget models' malware misses from 50 percent to 4. AWS says its monitoring stayed green while answers were wrong. Test the answers, not just uptime.
Made with AI from this Drop's sourced items. The facts come from the sources linked on each item. Script: Claude · Voice: OpenAI gpt-4o-mini-tts · Visuals: motion graphics
DeepSeek posted on X that V4.1 Flash 1017 is live on its website, app and API. OpenRouter lists it at $0.15 in and $0.6 out per 1M tokens with a 1048576-token context; two HN threads drew 425 and 307 points.
Why it matters: A low-priced model with a long context is worth a test run, but the only early tests we have are X posts, and they disagree.
The Verge reports Delta blocks agents from booking tickets, JPMorgan Chase and Charles Schwab block them from accounts over security concerns, and Yelp told TechCrunch it only allows agents that pay for its data licensing.
Why it matters: Agent features that shop or book may fail at big sites; The Verge says Meta partnered with Walmart, Sephora and others instead of fighting blocks.
AWS's sales operations team says testing its Customer Insights Agent found it inflating revenue figures and inventing an ARR estimate while monitoring stayed green. It now runs a nine-stage eval loop on Bedrock AgentCore.
Why it matters: AWS says one release mismatch alone dropped a test pass rate from 78 percent to 38 percent: uptime dashboards can't tell you if answers are right.
Google Research says agents designing DNA aptamers in a closed wet-lab loop raised on-target hit rates to 33.5% (myoglobin) and 42.1% (α-thrombin) in three rounds; its VERID tool dry-runs each robot protocol.
Why it matters: Google says humans authorize designs and supervise robot runs, and each target hit a different wet-lab failure mode, a reminder that agents need real-world feedback.
Databricks tested nine models on ten crafted npm packages hiding malicious binaries. Without its capability profiles, seven budget models failed 35 of 70 tests; with them, failures fell to 3 of 70.
Why it matters: Databricks found budget models often called hidden-binary malware benign with high confidence, and says majority voting across models would not fix it.
LMArena (CC BY 4.0) agent board as of 2026-10-16: Claude Opus 5.5 (Max) 0.189, GPT-6 Astra (Max) 0.141, Claude Fable 5.1 (Max) 0.14. Opus 5.5 (Max) also leads the webdev board, as of the same date, at 1807.309.
Why it matters: Preference boards are one input when picking an agent model; Opus 5.5 (Max) now tops both the agent and webdev boards.
Researchers built 600 computer-use tasks where sub-agents report outdated or conflicting info. Testing 15 frontier models, they found weaker ones follow misleading peers and fixing conflicts is far harder than spotting them.
Why it matters: The authors say even frontier models often forget to release shared locks or hand off tasks, a check worth adding if you run teams of agents.
Researchers report EdgeWise runs a small model on-device and hands key steps to a cloud LLM, matching or beating pure-LLM runs while cutting inference cost up to 62% on SWE-Rebench and 34% on Terminal-Bench 2.1.
Why it matters: The authors find not every step of a long agent run needs a frontier model, which points to a way to trim token bills.
Researchers release MAI-Gym, an open platform for GUI agent training. Their 9B agent scores 78.4% on MobileWorld, and RL in a mocked train-ticketing app lifted success on a real phone from 28.8% to 48.1%.
Why it matters: The authors release the framework, 2,000 verified tasks and the checkpoint, so teams can test whether training on mock apps carries over to real ones.
CEO Ben Chisell told CNBC new hires get a month to try AI tools and explain what they achieved. He estimates Paysend now delivers around 25% more engineering work with the same-sized team.
Why it matters: CNBC cites UK government research: 75% of AI-using businesses saw better productivity but 77% saw no revenue change, which puts blanket rollouts under scrutiny.
Pew's survey of 5,119 US adults in February found similar shares of men and women now use chatbots, but 27% of men vs. 20% of women use them daily, and 35% vs. 25% say they help productivity.
Why it matters: Pew finds women more likely to say AI is moving too fast (68% vs. 58%), a gap that can shape how workplace AI rollouts land.
Anthropic says its MOU with Singapore's digital ministry covers AI safety, security and model evaluation, with technical exchanges on testing standards and incident response. It says more than 1,500 GovTech engineers use Claude Code.
Why it matters: Anthropic says it will offer policy expertise as Singapore develops its AI rules; regulated firms there can now run Claude with in-region inference on Bedrock.
Anthropic says Claude.ai will draw voting info from Democracy Works later this month, and that Opus 5.5, Sonnet 5.5 and Haiku 5.5 score 99% on its political evenhandedness test.
Why it matters: Anthropic says testers found Claude could miss a malicious campaign split into harmless-looking requests; it added classifiers for multi-step misuse.
EUR-Lex lists the 2024 EU Product Liability Directive; per our registry, Dec 9 is its transposition deadline, the 1985 directive is repealed, and new rules cover products placed on the market after that date.
Why it matters: It marks when the EU's updated product liability regime starts covering newly marketed products, replacing rules that date to 1985.
deeplethe/solid-glass describes itself as a design system for AI applications: material, physical, tactile, technical. MIT licensed and created 2026-10-12, it gained 247 stars since our last check.
Why it matters: The fastest-growing repo in our set this run; a ready-made starting point if you build front ends for AI apps.
Leonxlnx/earshot describes itself as ears for coding agents: let AI hear, review and make sound, CPU-only and open source. It gained 220 stars since our last check.
Why it matters: MIT licensed and created 2026-10-15; useful if your agents work on apps or games with audio.
reason-machines/reason-network describes itself as your team's coding agents, talking to each other. Apache-2.0 licensed and created 2026-10-16, it gained 89 stars since our last check.
Why it matters: An early look at tooling for coordinating several developers' agents, a problem the CLASHarness preprint flags as hard.
Paweł Huryn posted a test on 2 repos with 105 hidden bugs that scored it 34, behind Opus 5.5 (max) at 41.7. @scaling01 posted that DeepSeek is always late; @ImAI_Eruel posted it could be another DeepSeek shock.
Why it matters: Community tests are anecdotes, not benchmarks: run your own evals on your own tasks before switching models.
Posted by @DeItaone: Anthropic co-founder Jack Clark says an internal model projects AI could boost US GDP from $30 trillion to $44 trillion by 2030.
Why it matters: This is a projection relayed in a post, not a forecast we can check; it shows how one lab frames AI's economic case.
The PS5LM project says it runs llama.cpp on the console's GPU with 11.4 GB of shared memory for the model, serving OpenAI- and Anthropic-style APIs to devices on the network. HN gave it 328 points and 79 comments.
Why it matters: It shows how far hobbyists push local inference on consumer hardware; it requires a jailbroken console on firmware 7.00 to 13.60.