| Weekly Edition |
Price: Zero tokens |
|
| The Weekly Token |
| All the week’s news that fits the context window. |
|
| Vol. I · No. 2 · October 17–24, 2026 |
|
|
A note from Apurv
Hi there,
Welcome to issue two of The Weekly Token. As always: the week’s biggest AI announcements, curated from X, with a link to every original post so you can go straight to the source.
This week was about open weights and agents that keep working after you leave: DeepSeek, Moonshot and MiniMax all shipped big open models, Anthropic launched Claude Duet and auto mode, and Agent Skills joined MCP at the Agentic AI Foundation. My picks still lean towards what matters if you’re building agentic apps.
As always, hit reply and tell me what was useful, what was noise, or what I missed.
|
|
|
Lead story · Open Source
DeepSeek released DeepSeek-V4.1-Flash-1017 last Saturday, calling it a Pareto improvement over the previous Flash: stronger scores, faster task completion and shorter outputs. Weights went to Hugging Face and ModelScope, and DeepSeek’s own table puts it at 42.3 on Terminal-Bench 4.0, up from 31.2 (company-reported). Off-peak API prices are $0.15 / $0.60 per million tokens, per kimmonismus, and DeepSeek opened its search and fetch APIs to everyone at $0.75 per 1,000 calls.
Arena ranked it No. 1 among open models in Agent Arena across more than 7,500 real sessions, at a $0.20 median cost per task, and No. 12 overall on Code Arena’s WebDev board. Then on Tuesday DeepSeek open-sourced DeepInfer, the in-house inference framework that serves V4.1-Flash at scale, with high-performance kernels, speculative decoding and structured outputs. “Open models. Open kernels. Open inference,” the lab wrote. The same week, the WSJ reported that DHS ordered federal civilian agencies to remove DeepSeek technology.
|
|
|
Claude Duet, a research preview for US Pro and Max subscribers, works as a proactive agent: it surfaces what needs attention, keeps working after you step away and checks in before acting. Starting next week, Claude's web, desktop and mobile apps also run in auto mode by default on Pro, Max and Team, with an automated check blocking risky actions.
|
|
|
Anthropic is donating the Agent Skills standard to the Agentic AI Foundation, home of MCP; it says more than 70 agent platforms have integrated it. Claude also now supports MCP Tasks and Elicitations, so plugins can run long jobs like deep research and report back when done.
|
|
|
Walmart's Agent Gateway lets approved AI agents shop its site over MCP (TechCrunch via Techmeme), Vercel Shop for Shopify is GA with MCP for product search and carts, and Figma Specs can hand work to coding agents through Figma's MCP server.
|
|
|
The agent proposes edits to its own prompts, tools, memory, skills and subagents, tests them and keeps what works, with frozen weights. The Apache 2.0 project reports gains of up to 11.5 points on WorkBuddy-Bench.
|
|
|
Apurv’s take (opinion)
MCP Elicitations are the bit I’m watching. An agent that can stop mid-task and ask the user a structured question is a UI problem, not just a protocol feature. With Claude Duet checking in before it acts and auto mode blocking risky actions, the hard design question is shifting from “how do I prompt it” to “when and how should the agent interrupt me.” Design those check-in moments deliberately.
|
|
|
|
A multimodal mixture of experts with 2.8T total and 104B active parameters and a 1M-token context, live in Kimi, Kimi Code and the API. Artificial Analysis ranked it fourth on its Intelligence Index at 55 and says it will be the leading open-weights model once the weights ship.
|
|
|
The stealth model that served more than 200 trillion tokens was an early M3.1 Flash, priced at $0.15 / $0.60 per million tokens. Artificial Analysis scored it 42, and on Friday the 428B-parameter, 23B-active weights went up on Hugging Face and ModelScope.
|
|
|
Co-founder Stefano Ermon says it outperforms GPT-6 Luna and Haiku 5.5 on quality; Techmeme reports Inception pitches it as 4x to 8x faster than those models.
|
|
|
Mistral launched Voxtral 3 plus a voice agent SDK built on LiveKit, ElevenLabs debuted Eleven Realtime speech-to-speech models that start talking in half a second, and Underdog released Bark 3.0, a 163 MB open speech-recognition model.
|
|
|
Canva says Craft is built for design and ranks second among AI labs on Design Arena's image board. fal and Black Forest Labs also shipped FLUX 3 Max.
|
|
|
|
Cursor's agents can now drive desktop apps such as the iOS Simulator to test the features they build.
|
|
|
Shared instructions, tools and context can stay cached for 12 hours at the cached-input rate, and the Decisions API, which classifies and scores inputs without a full response, is now GA.
|
|
|
Anthropic plans to ship versions more often so developers can handle breaking changes before a new model requires them. 2023-06-01 is not deprecated.
|
|
|
|
OpenAI's models found critical PHP flaws, and it disclosed that on September 20 a model used a DNS misconfiguration to get outside its offline training web, so it paused frontier training until October 1. The red-teaming that followed turned up a remote-code-execution bug in the Brotli library used by Node.js, Bun and Envoy.
|
|
|
OSS Scanner offers free periodic vulnerability scans for critical open-source projects, and Anthropic's Frontier Red Team says Claude, with Theorem Labs, proved security properties in more than 100 open-source libraries in a few weeks.
|
|
|
The breach hit 1.4% of Console users, all with external model providers linked. Co-founder dax said attacks ran around the clock for 48 hours and that refunds are underway.
|
|
|
Thousands of GPT-Rosalind agents, with PerturbAI and Scripps Research, proposed a testable Perry syndrome mechanism involving the DCTN1 gene in 32 minutes. OpenAI also acquired Orchestra Bio.
|
|
|
|
OpenAI's frontier models and Codex are coming to Google Cloud's Model Garden, and a day later Claude on Google Cloud, managed by Anthropic, became generally available.
|
|
|
The fund starts with 5% of Anthropic's shares, subject to stockholder approval; Anthropic says about 20% of its shares are now pledged to the public good.
|
|
|
Ornith-1.5-122B (MIT) reports 0.845 on SWE-bench Verified on one DGX card, Recursive's Flint scores 54.9% on Terminal-Bench 2.1 per Cline, and Meta published the report for Muse Glimmer, a 30B open agent model for a single consumer GPU.
|
|
|
| ■ |
ChatGPT adds team tasks to automate recurring engineering work |
| ■ |
Vercel's durable-harness keeps coding agents running for months — Agents resume automatically after failures.
|
| ■ |
VS Code previews canvases, shared workspaces for you and your coding agent — Cards can become plans the agent acts on.
|
| ■ |
Cline Desktop's Goal plugin keeps the agent working until a task is verified |
| ■ |
Devin Fusion scores 68 on the AA Coding Agent Index — An Opus 5.5 lead with a GPT-6 Luna sidekick, at 31% lower cost than Claude Code, Cognition says.
|
| ■ |
Replit launches a Monitoring Agent that finds, fixes and deploys changes |
| ■ |
Gemini 4 Argon drives Blender MCP in Antigravity to build a 3D cathedral |
| ■ |
Codex desktop app now lets you edit earlier messages |
| ■ |
Grok Bot can now phone restaurants and book a table, Elon Musk says |
| ■ |
1Password launches 1Password for Agents, WSJ reports |
| ■ |
Microsoft expands Project Perception to treat AI agents as insider risks — Satya Nadella argues security controls should sit outside the model.
|
| ■ |
GitHub Security Lab's open-source agent finds 24 Android vulnerabilities |
| ■ |
Google Cloud open-sources TorchTPU for native PyTorch on TPUs — Works with vLLM, SGLang, Hugging Face and Ray.
|
| ■ |
EAGLE-4 speculative decoding hits 961.6 tokens a second on a single request |
| ■ |
Exa 3.0 Claims the Top Spot on Every Search Benchmark It Tested |
| ■ |
NVIDIA open-sources Flora, models that turn raw footage into robot training data |
| ■ |
US banks restrict third-party AI agents from customer accounts, Axios reports |
| ■ |
Replit names OpenAI its preferred Agent partner and keeps Claude as an option |
| ■ |
OpenAI nears an up to $30B round at a $1.25T pre-money valuation, CNBC reports |
| ■ |
The Browser Company will shut down Arc on December 22 to focus on Dia |
|
|
|
That’s the week. If one of these changes how you build, or you think I picked the wrong stories, reply and tell me.
See you next Saturday, Apurv
P.S. I also break this stuff down on YouTube and at apurvkhare.com.
|
|
Sourced from X only · every headline links to its post The Weekly Token · Vol. I · No. 2 · Week of Oct 17–24, 2026
You’re getting this because you subscribed to Apurv Khare’s newsletter. Unsubscribe · Bangalore Urban, Bangalore, Karnataka 562125 |