AI Twitter/X Highlights Digest · 2026-10-16 (Fri)

Today’s Highlights

Meta releases the open-weight Muse Spark Open: @AIatMeta calls it “the most capable US open-weight model to date” — 816B total / 42B active, 1M context, natively multimodal, built for coding and long-horizon agent work; @ollama shipped a cloud version the same day with Claude Code support; @cursor_ai added Muse Wonder at 46.1% on CursorBench.
Gemini 4 Argon lands in Antigravity: @antigravity opened Argon to paid Google AI Pro users with a separate weekly quota that doesn’t eat into 3.8 Flash; reception is mixed — @signulll says it codes “worse than Opus, even worse than Sonnet in many cases.”
Anthropic tells investors it’s profitable: @FT reports Anthropic told investors the company is profitable as it rallies support for an IPO, though growth has been slower than some backers expected; @theo’s “if Anthropic drops Fable 5.5 today it might be all over” got 6k+ likes.

1. Muse Spark Open: Meta’s 816B open-weight coding model

Key points: @AIatMeta released Muse Spark Open, calling it “the most capable US open-weight model to date”: 816B total / 42B active parameters, a 1M-token context window, native multimodality across text, image, video and audio, and a focus on long-horizon tasks, complex software engineering and multi-step reasoning. @alexandr_wang’s thread repeated the specs. @ollama shipped muse-spark-open:cloud the same day, and ollama launch claude muse-spark-open:cloud plugs it straight into Claude Code (Max / Teams first, Pro soon). Also that day, @cursor_ai added the closed Muse Wonder: 46.1% on CursorBench at default effort, the highest of any Muse model, priced at $4 / $20 per million input / output tokens.

Why it matters: A top open-weight model aimed squarely at agentic coding, plus same-day Claude Code support through Ollama, makes it easier still to bring open-weight models into coding workflows. Note that “most capable” is Meta’s own claim.

AIatMeta: Muse Spark Open launch: https://x.com/AIatMeta/status/2110779754274292212
alexandr_wang: specs and positioning: https://x.com/alexandr_wang/status/2110778736434184611
ollama: on cloud, works with Claude Code: https://x.com/ollama/status/2110833609829695718
cursor_ai: Muse Wonder in Cursor, 46.1% CursorBench: https://x.com/cursor_ai/status/2110867556693430293

2. Gemini 4 Argon in Antigravity: a separate weekly quota for Pro

Key points: @antigravity is rolling out Gemini 4 Argon to paid Google AI Pro subscribers, saying demand has been “incredible” and that it scaled capacity deliberately to get here; Pro gets an extra weekly-only Argon quota on top of the existing 3.8 Flash bucket (confirmed by @_mohansolo). @Google also posted a thread on how teams across Google used Argon on their hardest engineering problems. The counterpoint came from @signulll: in his coding tests Argon struggled even on tiny codebases, hallucinated a lot, and was “doing worse than something like Sonnet in many cases.”

Why it matters: Last week Argon’s story was the launch and benchmarks; now it is actually in a coding IDE with a quota. The gap between hands-on reports and official messaging is worth watching.

antigravity: Argon for Pro: https://x.com/antigravity/status/2110866531932316011
_mohansolo: separate weekly quota: https://x.com/_mohansolo/status/2110866446020710657
Google: internal engineering thread: https://x.com/Google/status/2110746360047603834
signulll: falls short of Opus for coding: https://x.com/signulll/status/2110906507776667709

3. Grok Bot: a purchase guarantee, growth with no free tier, plugin submissions

Key points: Quoting a user who booked a flight with Grok Bot, @elonmusk said that if Grok Bot makes a mistake with a purchase or leaks your card (“never happened yet”), SpaceX will make you whole; in another post he said he has “never seen anything grow as fast,” and that with no free tier, people are getting real value for what they pay. @poteto pointed out that Grok Bot (and Grok, Cursor) plugins can be submitted; @grok described the free Spam-Call Scrubber skill, which finds your number on people-search sites and data brokers, submits opt-outs, handles confirmations, and rechecks on a schedule you set.

Why it matters: Trust is the hardest part of letting an agent spend your money. A vendor guarantee plus an open plugin ecosystem show personal agents moving from “usable” to “trustworthy enough.”

elonmusk: SpaceX covers purchase mistakes: https://x.com/elonmusk/status/2110610960474079445
elonmusk: fastest growth, no free tier: https://x.com/elonmusk/status/2110866513955868924
poteto: plugin submissions: https://x.com/poteto/status/2110840796954202186
grok: Spam-Call Scrubber skill: https://x.com/grok/status/2110852882907934743

4. Anthropic: tells investors it’s profitable, new Fellowships, Fable 5.5 rumors

Key points: @FT reports that Anthropic has told investors the company is profitable as it rallies support for a blockbuster IPO, though its growth has been slower than some backers anticipated amid concern about the AI trade (revenue was reported yesterday; “profitable” is the new part). @AnthropicAI opened two new Fellowships — STEM Fellows and Conceptual Reasoning Fellows (philosophy, cognitive science, economics, policy, law) — four months at Anthropic starting January 2027, deadline October 28. In the community, @theo’s “If Anthropic actually drops Fable 5.5 today it might be all over” got 6k+ likes, and he later said he wants to make a new model tier list but fears Fable 5.5 will drop the moment he does; @evanyou joked about Anthropic’s new post on defending against distillation attacks and Bun.

Why it matters: “Profitable” carries a lot of weight in an IPO narrative; Fable 5.5 is community speculation only, with no official word.

FT: Anthropic says it’s profitable: https://x.com/FT/status/2110709513552601571
AnthropicAI: two new Fellowships: https://x.com/AnthropicAI/status/2110720940665372825
theo: Fable 5.5 speculation: https://x.com/theo/status/2110673598533050814 https://x.com/theo/status/2110827068347027508
evanyou: anti-distillation vs Bun: https://x.com/evanyou/status/2110722301536915553

5. OpenCode: heading for 2M DAU after 2.0, DeepSeek V4.1 Flash free

Key points: @thdxr says OpenCode’s numbers are “going crazy again” since the 2.0 release and are about to hit 2M daily active users; he also showed a workflow of walking outside with the OpenCode phone app, controlling the OpenCode server on his office computer, while it works in a Slack thread with his team and the OpenCode slackbot. @opencode announced DeepSeek V4.1 Flash is now free in OpenCode — a post with 22k likes.

Why it matters: The open-source coding agent keeps winning users with free models and more surfaces; driving a local server from a phone and into Slack blurs the line between terminal agents and team collaboration tools.

thdxr: approaching 2M DAU: https://x.com/thdxr/status/2110544790764794231
thdxr: phone → server → Slack: https://x.com/thdxr/status/2110793318959632615
opencode: DeepSeek V4.1 Flash free: https://x.com/opencode/status/2110635747648909472

6. Devin Oncall: Devin as your SRE

Key points: @cognition launched Devin Oncall: Devin automatically investigates alerts and bugs, helping teams diagnose and resolve production incidents, and it says it already handles thousands of alerts every day for some of the world’s largest companies. @walden_yan says usage was “explosive” even in private beta, so engineers lose less sleep on call.

Why it matters: Coding agents are stretching from writing new code to on-call debugging — one of the most painful parts of engineering time, and one of the easiest to measure.

cognition: Devin Oncall launch: https://x.com/cognition/status/2110790097369010292
walden_yan: private-beta usage: https://x.com/walden_yan/status/2110795331709632768

7. OpenAI: GPT-Live-1.1 comes to Codex voice, ChatGPT can drive Chrome tabs

Key points: @OpenAI released GPT-Live-1.1, with new voices, better accents and less “checking…”, across Voice in ChatGPT, Codex and dots, adding voices in French, German, Hindi, Mexican Spanish, Korean and Japanese; rolling out to all users today, API coming soon. The same day @ChatGPT brought computer use to Chrome users: in Chat it can look across your open tabs, and in Work it can control your Chrome tabs to get things done (sending messages, filling out forms, and so on). @nickbaumann_ called it a “subtle but huge release.”

Why it matters: Voice in Codex plus browser control in the main ChatGPT product show OpenAI pushing agent capabilities into everyday entry points.

OpenAI: GPT-Live-1.1: https://x.com/OpenAI/status/2110797830361436604
ChatGPT: Chrome tabs and Work mode: https://x.com/ChatGPT/status/2110823738531037651
nickbaumann_: “subtle but huge”: https://x.com/nickbaumann_/status/2110849181829185833

8. Magnitude: an open-source local inference engine claiming 2× over Ollama

Key points: @tomgreenwald released Magnitude, an open-source inference engine that spends up to about a minute tuning kernels on your actual chip when you download a model, keeping the fastest combination that doesn’t change output, and pairs each model with the best speculative-decoding drafter. Official numbers, averaged over 11 popular local models: on an M4 Pro vs Ollama (MLX), 1.32× faster prefill and 2.19× faster decode; on DGX Spark vs Ollama (llama.cpp), 2.17× faster decode. It runs on Apple Silicon, NVIDIA, AMD or just a CPU, connects in one click to agents like Pi, OpenCode, Hermes and Codex, roughly halves per-agent memory with a quantized KV cache, and lets agents share prefix caches. The post has 1.7k bookmarks, more than its likes.

Why it matters: Speed and memory are what limit local-model agents; per-machine automatic kernel tuning is a way forward that needs no new hardware. Numbers are the team’s own benchmarks.

tomgreenwald: Magnitude launch: https://x.com/tomgreenwald/status/2110718297255670071

9. Editors for the agent era: T3 Code leaves alpha, Mayoi, IntelliJ comes to Cursor

Key points: @jullerino announced that T3 Code’s Orchestrator V2 is on the stable channel and that the Alpha branding is gone, with this release called 0.1.0 Beta; mobile builds 2.0.0 are on the App Store and Play Store. @theo quoted it: “T3 Code is finally out of alpha.” @chenchengpro open-sourced the editor Mayoi, built for reviewing diffs, reviewing agent code and launching the next agent: written in Go with a GPU-drawn native UI and no Electron, a 24 MB app (vs. 970 MB for Cursor by his count), about 180 MB of memory, a Ghostty-based terminal, MIT-licensed. @intellijidea announced that Java and Kotlin by IntelliJ IDEA, built with @cursor_ai, has reached 1.0 — an LSP-based extension that takes JVM tooling beyond the IDE.

Why it matters: As people hand-edit less code, the editor’s center of gravity is shifting from writing to reviewing and orchestrating agents.

jullerino: Orchestrator V2 / 0.1.0 Beta: https://x.com/jullerino/status/2110874152186040662
theo: T3 Code out of alpha: https://x.com/theo/status/2110907433656938681
chenchengpro: Mayoi open-sourced: https://x.com/chenchengpro/status/2110699165051453771
intellijidea: Java/Kotlin extension 1.0: https://x.com/intellijidea/status/2110737206767665224

10. pi-durable: from a Python IDE on iOS to an in-browser programming teacher

Key points: The pi-durable agent from the pi ecosystem showed up in several projects over the past day. @mitsuhiko said that with the new models, in an afternoon (mostly passive) plus one flight, he built a Python IDE for iOS that runs Pyodide, has a debugger, linter and formatter, embeds a pi-durable-powered coding agent, and supports iCloud Drive; he also recorded the game he builds with his kids, which embeds a pi-durable agent right in the browser as a programming teacher. @badlogicgames also shared kitze’s personal-agent build on it.

Why it matters: Embedding a small, durable agent runtime straight into an app, or even the browser, is lighter than building another big IDE, and it’s a concrete example of agents as components.

mitsuhiko: iOS Python IDE: https://x.com/mitsuhiko/status/2110680120843444589
mitsuhiko: agent teacher in a game: https://x.com/mitsuhiko/status/2110726538622636420
badlogicgames: kitze picks up pi-durable: https://x.com/badlogicgames/status/2110741895374647367

Meta ships the 816B open-weight Muse Spark Open, Argon lands in Antigravity, Anthropic says it's profitable