| Weekly Edition |
Price: Zero tokens |
|
| The Weekly Token |
| All the week’s news that fits the context window. |
|
| Vol. I · No. 3 · October 24–31, 2026 |
|
|
Hi,
Sorry this week’s Weekly Token is late. I was busy at React India in Goa. Here’s the video of my talk if you want to watch it: Watch my talk. I also ran a workshop there; it wasn’t recorded, but I’m planning to do a webinar on it soon.
Welcome to issue three 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 Google put Gemini 4 Carbon at the top of the leaderboards, OpenAI shipped a cheaper computer-use model, Anthropic filed to go public, and MCP picked up two big platform backers in Apple and Android. 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 · Models
Sundar Pichai introduced Gemini 4 Carbon on Wednesday as Google's next frontier model, just two weeks after Gemini 4 Argon. “I did say we'd be releasing our models periodically,” he wrote. Carbon ranks No. 1 on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59, which Google says is 9 points above Argon and just ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57 in its own comparison table.
Google is leading with security. Pichai said Carbon scores 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation and 71.0% on CWE-bench for patching, and Google DeepMind called it state of the art across cybersecurity and long-horizon work. The rollout is gradual: some trusted cyber defenders get access this week, then more partners through Google's Fairwind Program, ahead of wider availability “soon.” That evening, on Alphabet's Q3 results, Pichai said Argon and Carbon are “leading the frontier.”
|
|
|
MCP co-creator David Soria Parra said Apple now maintains the official Swift SDK for the Model Context Protocol. On Thursday, at Android Dev Summit, Google put the AppFunctions API into beta so apps can expose functions to their own agents or to Gemini over Android MCP. Google also previewed a Connected Apps console for building apps in Search and the Gemini app, with an MCP and tools setup panel and a review step. Earlier, the NVIDIA app began shipping a local MCP server, on by default and listening only on 127.0.0.1.
|
|
|
Anthropic told Reuters that Claude, on an internal statistics project, requested an access token from Georgia's dashboard and pulled aggregate voter counts the state normally sells, without human authorization. On Thursday OpenAI said it now monitors all frontier RL runs, evals and internal Codex use, and that in September a model exploited a DNS misconfiguration to get online and was flagged within 15 minutes. Anthropic described Claude self-assessment, which flags misuse in its own conversations, and on Friday showed thousands of Claude agents red-teaming a training-data monitor.
|
|
|
Agents are defined as code, with prompts in Markdown and tools in TypeScript, deployable on Cursor's infrastructure or your own, with built-in evals on every change. Later in the week Cursor added Annotate to Design Mode, and added DeepSeek V4.1 Flash on its CursorBench cost frontier at $0.60 to $0.98 per task.
|
|
|
An Anthropic Applied AI guide says a single agent with skills often beats multi-agent setups, with one customer running about twice as fast. It advises turning subagent instructions into skills, exposing actions as tools or MCP, and running evals before and after, and warns tool choice degrades past 30 to 50 tools.
|
|
|
U.S. ChatGPT Pro users can text their dot tasks over iMessage or RCS and get proactive updates back, in preview. Earlier, Elon Musk said Grok Bot can call a person or business, wait on hold, make a request and send back a summary.
|
|
|
|
Released Monday in ChatGPT, Codex and the API, Terra reads the screen and acts across apps. Posts quoting OpenAI's launch materials put API pricing at $1 / $5 per million input / output tokens, 50% below GPT-6.1 Sol, and OpenAI says it matches GPT-6 Astra on its Agents' Last Exam at medium effort for a tenth of the price per token (company figures). GitHub made it GA in VS Code with Copilot the same evening, and by Friday Arena had it at No. 8 in Agent Arena at a $0.62 median cost per task.
|
|
|
Worlds can be steered live, scenes get native audio with characters who converse in real time, and Gemini's reasoning keeps multi-step events coherent. It is in research preview with trusted testers, and Pichai said it is coming to AI Ultra subscribers and, for the first time for a Genie model, the Gemini API.
|
|
|
Fern 0.1 targets AI R&D chores such as auditing training data and reading agent traces. The lab reports 73.4% on SWE-bench Verified; access is by request.
|
|
|
|
On Thursday OpenAI doubled usage limits on every Pro tier and Business Premium and dropped a planned Pro 200 cut, a day after offering Pro and Business subscribers up to $500 a month in API credits. On Friday Anthropic renamed Max 5x to Max and Max 20x to Ultra at the same $100 and $200 prices and removed the 5-hour limit for verified Ultra accounts; weekly limits are unchanged. The Information reported OpenAI is winning Anthropic customers with 30%+ discounts, with Replit moving most of its AI spend, and Moonshot gave K3.1 Pro subscribers back a week of quota after rate-limit complaints.
|
|
|
Unveiled at GitHub Universe, Fusion asks for an effort level instead of a model and can coordinate several Anthropic and OpenAI models on one task. Artificial Analysis measured Fusion Max at 60 on its Coding Agent Index, 2 points behind Claude Code with Opus 5.5 (max), at $5.59 per task versus $11.42.
|
|
|
The Jules team called the async coding agent an experiment; remote background tasks move into Antigravity, which also launched a plugins marketplace.
|
|
|
|
Greg Brockman announced AI-assisted results in harmonic analysis, complexity theory and quantum field theory on Monday, including a claim that the Exponential Time Hypothesis is false; posts relaying it say two are checked in Lean and will be reviewed by outside experts. Hours later Anthropic's Levent Alpöge shared a paper saying an internal Claude research model found the core of its own anti-ETH 3-SAT algorithm on Oct 16 with no human input. Neither has been peer reviewed. Earlier in the week, PDE expert Scott Armstrong said OpenAI's Navier-Stokes blow-up construction “has no precedent.”
|
|
|
60% of the new index comes from private or held-out evaluations, including Cognition's FrontierCode 2.0. At launch Claude Opus 5.5 (max) led at 58 until Carbon arrived; a new Progress Index finds open models trail proprietary ones by a median of about 4 months.
|
|
|
Pichai said Google sent a command to its prototype satellite, booted the TPU, ran a prompt and got a complete response back on Earth: a haiku.
|
|
|
|
On Monday Anthropic said it had filed an S-1 with the SEC for a Nasdaq listing under “ANT”; share count and price are not yet set. Posts summarising the filing report first-half 2026 revenue of $16.3 billion and an $80.7 billion net loss, mostly a non-cash charge tied to convertible notes, and Reuters says it discloses a month-long training pause and AI security breaches. The week opened with Thariq Shihipar and Adam Morris, who created Claude Code on desktop, announcing their departures, and on Monday Axios reported the FTC wants Anthropic executives to testify in an expanding AI safety probe.
|
|
|
Thursday's release has 1.05T total and 52B active parameters and a 1M-token context, per Artificial Analysis, which ranks it the top open-weights model built outside China, at 26 on Intelligence Index v5. Mistral says it is No. 1 among open models on Cognition's FrontierCode. Pricing drops about 50% to $0.70 / $2.10 per million tokens, and the new MEOW license requires a separate commercial license for model-as-a-service businesses. NVIDIA says it was trained on 3,800 Grace Blackwell GPUs.
|
|
|
A tri-hybrid architecture spans a 1.7B text model with 256K context and 2B vision, audio and omni models, each fine-tunable on one GPU. Ollama added them the same day.
|
|
|
Arena put K3.1 Pro (Max) at No. 11 in Agent Arena, the top open model, at $1.03 median cost per task. Moonshot then apologized for rate limits and gave subscribers back 25% of a month's quota, while VentureBeat reported scrutiny of undisclosed paid influencer promotion.
|
|
|
Apache 2.0 weights with a 512K context; the lab says it leads most agentic coding benchmarks in its class (self-reported).
|
|
|
The FT reports a stock offering with the chipmakers as cornerstone investors; CNBC says their interest is up to $5 billion in shares.
|
|
|
| ■ |
Alphabet reaches a half-trillion-dollar revenue run rate; Google Cloud revenue grows 103% — Google says its first-party model APIs process nearly 32B tokens a minute.
|
| ■ |
Microsoft reports $92.3B in quarterly revenue, with Azure up 48% |
| ■ |
Mistral opens a Responses API for agents in public preview — Answers, reasoning and tool calls come back as separate items.
|
| ■ |
Vercel CI launches in beta, built for agents through Vercel MCP and the CLI — It ran 2 to 6 times faster in Vercel’s own repos.
|
| ■ |
A planted AGENTS.md in your Downloads folder can hijack coding agents — In tests Codex followed such a file more than 90% of the time.
|
| ■ |
Anthropic shares early drug-discovery results with six partners |
| ■ |
OpenAI reorganizes its Model Spec around five core values |
| ■ |
Claude for Legal adds a gallery of ready-made artifacts |
| ■ |
Gemini can now generate and edit documents, PDFs, slides and spreadsheets |
| ■ |
Anthropic offers Claude to U.S. governments at 40% off under a new GSA OneGov deal |
| ■ |
Suhail Doshi launches Particle; particle-1 matches performance with about 45% fewer tokens |
| ■ |
Google AI Studio is getting a managed Agents hub, TestingCatalog spots |
| ■ |
Cognition open-sources agent skills that patch PHP, Node.js and GitLab flaws |
| ■ |
Cognition is acquiring Corridor, a software security company |
| ■ |
Gemini 4 Argon found a 2.85x Gemma speedup across 8 vLLM pull requests |
| ■ |
Unsloth shrinks MiniMax M3.1 Flash to run locally on about 150GB |
| ■ |
Tencent open-sources Hy-SimulMT, a live Chinese-English interpreter with voice cloning |
| ■ |
OpenBMB open-sources MiniCPM-OCR, a 1B parser paired with a 2B checker |
| ■ |
NVIDIA Nemotron 3.5 Transcribe ties for No. 1 among open speech-to-text models |
| ■ |
Swiss lab Giotto launches Alto and Solo as sovereign European 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
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. 3 · Week of Oct 24–31, 2026 |
|