| Weekly Edition |
Price: Zero tokens |
|
| The Weekly Token |
| All the week’s news that fits the context window. |
|
| Vol. I · No. 1 · October 10–17, 2026 |
|
|
A note from Apurv
Hi there,
Welcome to the very first issue of The Weekly Token, the first of many. Every Saturday I’ll send you the week’s biggest AI announcements, curated from X, with a link to every original post so you can go straight to the source.
I build and teach agentic UI (MCP Apps, generative UI, MCP servers, WebMCP, ChatGPT apps), so my picks lean towards what matters if you’re building agentic apps. No hype, just the signal, kept short enough to read with your chai.
This is issue one, so I’d love your feedback. Just hit reply and tell me what was useful, what was noise, or what I missed. I read every reply.
|
|
|
Lead story · Models
Google began rolling out Gemini 4 Argon globally on Tuesday: in the paid tier of the Gemini API, to Ultra subscribers in Google Antigravity, and to Ultra and Pro subscribers in the Gemini app and Search. Artificial Analysis scored it 53 on its Intelligence Index, level with GPT-6 Astra at about 60% of the cost per task, and launch pricing is $2 / $10 per million input / output tokens until November 30 (then $4 / $20).
Cursor reported 52.8% on CursorBench and GitHub made it generally available in Copilot the same day; on Thursday it reached Google AI Pro subscribers in Antigravity, and on Friday Google DeepMind said four Argon agents beat Minecraft from scratch in 27 hours.
|
|
|
Muse Wonder (built for agents, 1M-token context) stays closed for now, while Muse Spark Open, an 816B mixture-of-experts model with 42B active parameters, was ranked the No. 1 U.S. open-weights model by Artificial Analysis.
|
|
|
Fast mode is off the waitlist: up to 2.5x faster at $8 / $40 per million tokens, live in Microsoft Foundry with Google Cloud coming soon.
|
|
|
|
Mercury's MCP lets agents in ChatGPT or Claude queue payments that users approve in the Mercury app, Vercel's AI SDK added MCP events so agents can subscribe to server updates, and OpenAI showed annotations paired with WebMCP in ChatGPT's browser.
|
|
|
Dynamic workflows (public beta) let a lead agent's plan run across up to 1,000 agents, and Claude Platform added native OpenTelemetry traces for the Messages API, Managed Agents and SDKs.
|
|
|
In ChatGPT, multiple agents can now work on different parts of a task at once, with each agent's progress visible and completed tasks open for review.
|
|
|
Bret Taylor's Personal Agent Protocol lets consumers choose what their agents can access; Cloudflare, one of 35 design partners, says it builds on OAuth, MCP and OpenAPI.
|
|
|
Apurv’s take (opinion)
The Mercury item is the pattern I’d bet on: the agent proposes, a real UI confirms. A payment queued over MCP and approved in the app is human-in-the-loop done right, and with MCP events and WebMCP annotations in the mix, MCP is starting to look like something interfaces can react to, not just a request/response pipe. If you’re building agentic front ends, design the confirmation surface first.
|
|
|
|
Merges grew from about 2,500 a day in May to more than 30,000 in September, so Anthropic built its own version control system; it also says Claude writes about 90% of Claude Code's code.
|
|
|
Five tiers from Free to Ultra starting at $8 a month, with one usage pool across Cursor, Grok Build and other Grok products; third-party models in Cursor move to API pricing.
|
|
|
|
OpenAI says it used its own models to work out what the device was and who planted it, referred the case to the FBI, and has no indication that model weights, source code or customer data were taken.
|
|
|
Beam's weights, tokenizer and reference implementation are out under Apache 2.0 in BF16, FP8 and NVFP4, a week after reports that Nvidia is in early talks to acquire Reflection or deepen its stake.
|
|
|
| ■ |
Basecamp ships an MCP server for ChatGPT, Claude, Cursor and Grok |
| ■ |
IHME releases a Global Burden of Disease MCP for Claude — Claude users can explore virtually all of the GBD 2023 study estimates.
|
| ■ |
Cursor cloud agents can now build shareable interactive canvases — Ask for a dashboard or report while the agent works, then share it as a link.
|
| ■ |
Figma introduces Specs, collaborative markdown for context — Announced at Config India alongside Code Layers and Guidelines.
|
| ■ |
Claude Code Projects opens to everyone on the waitlist — Each task runs as its own parallel thread.
|
| ■ |
Nous Research ships Hermes Agent v0.22.0 — Gmail, Linear and Notion connectors, one-click local models and a 600-plugin catalog.
|
| ■ |
Codex raises GPT-6.1 Sol's default context to 500K — Up from 272K, with no long-context fees on subscriptions, part of a promised 28-day run of daily improvements or usage resets.
|
| ■ |
Liquid AI releases LFM2.5-24B-A2B for agents on laptops — 2B active parameters, 131K context, base and post-trained weights on Hugging Face.
|
| ■ |
Deno's whole team joins Cloudflare — Secondary posts say Deno Deploy shuts down in six months while JSR carries on.
|
| ■ |
Coinbase says a fine-tuned Qwen3.5-9B beat frontier models at fraud prevention |
| ■ |
Karpathy describes a "sharp funnel" in who has seen what LLMs can do — His napkin math puts about 5,000 people with internal access to frontier systems.
|
| ■ |
Nadella says companies should treat powerful models as potential insider threats — His essay asks companies to build an emergency brake for agentic models.
|
|
|
|
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
|
|
Sourced from X only · every headline links to its post The Weekly Token · Vol. I · No. 1 · Week of Oct 10–17, 2026 |
|