AI & Agents Daily Briefing

Claude Fable 5.2 ships 30% faster at a lower price, Meta's small Muse Ember, and Google's TPU 8tx and 8i Max

Mon, 09 Nov 2026 22:30:00 GMT

Here's today's news. Monday (US time) was busy. Anthropic shipped its top models, Claude Fable 5.2 and Claude Mythos 5.2, and the same day Meta announced a new model and Google announced new TPUs.

Fable 5.2 generates output up to 30% faster than Fable 5.1, and input and output prices both drop 20%. Meta's Muse Ember is the smallest Muse model, built to handle repetitive agent work cheaply at $0.10 per million input tokens.

Google introduced TPU 8tx for confidential inference and TPU 8i Max for long contexts. It also open-sourced TPU Sync, a KV-cache tiering layer, and put a TPU agent suite for coding tools into preview.

Claude Fable 5.2 and Mythos 5.2: Anthropic's top models, up to 30% faster than Fable 5.1 and 20% cheaper (Anthropic, GA)
Meta Muse Ember: the smallest Muse model, for tool calls and coding subtasks at $0.10 per million input tokens (Meta, GA)
Google unveils TPU 8tx and 8i Max for confidential inference and long contexts, and open-sources TPU Sync (Google, Preview)

Tencent's DSpine speeds up Qwen3.8-27B about 3x, and Anthropic explains how Claude's agent memory works

Sun, 08 Nov 2026 23:30:00 GMT

Here's today's news. This window covered Saturday and Sunday, and no new models or major products shipped. So we picked one thing you can download and run today, plus one write-up that teams building agents should read.

Tencent posted DSpine to Hugging Face, a speculative-decoding draft model that speeds up generation on Qwen3.8-27B. It plugs into SGLang with one command. On replayed agent workloads Tencent measured 2.7 to 3.8 times the throughput of plain decoding.

We also added a late catch: Anthropic's engineering post on agent memory, published early on November 7 (KST). It fell inside our previous issue's window, and we missed it.

Tencent DSpine: a speculative-decoding draft model that runs Qwen3.8-27B about 3x faster (Tencent, GA)
[Late catch] Anthropic on agent memory: letting Claude reread its work logs and fix its notes with "dreaming" (Anthropic, Research)

Meta brings Muse Wonder-trained models and spaces to Muse, Ai2 opens ToyBlocks, Ripple-65B ships open weights

Sat, 07 Nov 2026 22:45:00 GMT

Here's today's news. This window covered a Friday and a Saturday, and we didn't find a major new product launch, so today's picks lean toward updates you can try or download right now.

Meta swapped the models behind Muse and Meta AI for specialized versions trained with Muse Wonder, and launched spaces: describe a tracker or a game and Muse builds it for you. Ai2 released ToyBlocks, a simulated world with one robot arm and some blocks, along with code, data and models.

We also have one late catch. On November 4, Ricursive Intelligence released Ripple-65B Preview, a 65B dense model under Apache 2.0. It slipped past our earlier issues.

Meta puts Muse Wonder-trained models into Muse and Meta AI, and adds build-your-own spaces (Meta, GA)
Ai2 ToyBlocks: an open robot-arm-and-blocks simulation that generates its own demonstrations (Allen Institute for AI (Ai2), Research)
[Late catch] Ripple-65B Preview: a 65B dense model tuned for coding agents, with Apache 2.0 weights (Ricursive Intelligence, Preview)

Thinking Machines ships Inkling 2 Small, Qwen3.8-27B runs locally for Copilot, Mistral Studio gets managed agents

Fri, 06 Nov 2026 22:20:00 GMT

Here's today's news. Thinking Machines Lab released Inkling 2 Small, its second-generation model. It's a mixture-of-experts model with 594B total and 17B active parameters, available only on Tinker for now, and the company says Apache 2.0 weights follow next week.

On the developer side, Microsoft showed how to run Qwen3.8-27B on your own GPU with Foundry Local and wire it into GitHub Copilot, and Mistral added managed agents with skills, sandboxes and subagents to Studio.

One enterprise item too: Anthropic added a security center to Claude Enterprise and made inference hooks, which route prompts and tool calls through your DLP tools, generally available.

Inkling 2 Small: Thinking Machines' second-gen model, with Apache 2.0 weights promised next week (Thinking Machines Lab, Preview)
Foundry Local runs Qwen3.8-27B on your own GPU and plugs it into GitHub Copilot (Microsoft, Preview)
Mistral Studio adds managed agents with skills, sandboxes and subagents, on-prem included (Mistral AI, Preview)
Claude Enterprise gets a security center, and inference hooks for DLP checks are now GA (Anthropic, Beta)

OpenAI's GPT-6 Cyber, Gemini Omni 1.2 Flash stretches to 15 seconds, and Images 2.6 adds GIF and SVG

Thu, 05 Nov 2026 21:50:00 GMT

Here's today's news. OpenAI released GPT-6 Cyber, a model built only for security research. It finds vulnerabilities, proves whether they can actually be exploited, and was trained to stay inside the scope a researcher sets. For now, only a select group of approved Daybreak Red customers can use it.

There are two generative-media updates as well. Google made Gemini Omni 1.2 Flash generally available in the Gemini API, with 15-second clips, conversational editing and a 30% price cut. OpenAI shipped Images 2.6, which adds GIF creation and SVG export and splits the API into two models, Flare and Sunburst.

OpenAI GPT-6 Cyber: a security model that finds and proves exploits while staying in scope (OpenAI, Pilot)
Gemini Omni 1.2 Flash goes GA: 15-second video, conversational edits and a 30% price cut (Google DeepMind, GA)
OpenAI Images 2.6: GIFs, SVG export, and two API models, Flare and Sunburst (OpenAI, GA)

Tencent ships Hy4, Copilot lands in Xcode's built-in assistant, Claude plugins connect at login

Wed, 04 Nov 2026 22:30:00 GMT

Here's today's news. Tencent released the full version of Hy4. It adds image understanding and computer use, neither of which the August preview had, and it's available now through Tencent Cloud and OpenRouter. Tencent says the weights are coming soon but gave no date.

There are two developer-tool updates too. GitHub Copilot is now a public preview option in Xcode 27.3 beta's built-in coding assistant. Anthropic opened a beta of Claude-issued auth, which lets Team and Enterprise members reach company plugins and MCP connectors just by signing in to Claude, without logging in to each service separately.

Tencent Hy4: a 770B model that can see the screen and drive a computer, with weights promised soon (Tencent, GA)
GitHub Copilot comes to Xcode's built-in coding assistant, in public preview from Xcode 27.3 beta (GitHub, Preview)
Claude-issued auth beta: sign in to Claude and your company's plugins and MCP connectors are already connected (Anthropic, Beta)

Apple opens Siri AI to every app, Anthropic open-sources FlashProve, Claude's watermark reaches older models

Tue, 03 Nov 2026 21:50:56 GMT

Here's today's news. No big new model shipped, but three releases are worth a developer's time. Apple opened a Siri AI developer preview in the iOS and macOS 27.3 betas, so any app can now hand its actions and content to Siri AI through App Intents. Anthropic open-sourced FlashProve, a system that uses cryptography to prove a reply really came from the model a provider committed to.

Claude's release notes also say the text watermark now covers older models such as Opus 4.7 and Sonnet 4.6. There's a catch for Sonnet 4.6: change the sampling parameters and the watermark may not be there.

Apple Siri AI developer preview: any app can hand actions and data to Siri AI through App Intents (Apple, Preview)
Anthropic FlashProve: open-source cryptographic proof that a reply came from the promised model (Anthropic, Research)
Claude's text watermark reaches older models, but Sonnet 4.6 can drop it if you change sampling (Anthropic, GA)

Claude API message threads in beta, Gemini 2.5 Pro and Flash auto-reroute on Dec 1, Vercel agent-desktop 1.0

Mon, 02 Nov 2026 23:10:00 GMT

Here's today's news. No big launch today, so this issue sticks to updates developers can try now or should prepare for. Anthropic opened message threads on the Claude API in beta: the server holds the conversation briefly, and after the first request you send only the new messages. Google's changelog says that starting December 1, calls to gemini-2.5-pro and gemini-2.5-flash will be routed to newer models and billed at their rates.

We also missed Vercel Labs' agent-desktop 1.0, released on 10/28. It hit #1 on GitHub's daily Trending on 11/02, so it's in today as a late catch.

Claude API message threads (beta): keep the conversation server-side, send only new turns (Anthropic, Beta)
Gemini 2.5 Pro and Flash deprecated: from Dec 1, calls reroute to newer models at their prices (Google, GA)
[Late catch] Vercel agent-desktop 1.0: a CLI that lets agents read and drive desktop apps through the accessibility tree (Vercel, GA)

DeepSeek open-sources a super-agent, ChatGPT and Codex merge usage limits, Karpathy ships HackerTube

Sun, 01 Nov 2026 22:40:00 GMT

Here's today's news. It was a weekend with no major launches, so we picked things developers can try now or should plan for. DeepSeek published dsh-superagent, a research prototype of a coding agent that takes requests on GitHub and answers with pull requests. OpenAI's help center says Plus, Pro and Business plans will share one token-based allowance across ChatGPT and Codex starting in December.

Andrej Karpathy launched HackerTube, a video recommendation site built by having Claude filter 43 million Hacker News comments, and released the analysis data as a zip. We also added Google's Connected Apps Platform from October 28 as a late catch, since earlier issues missed it.

DeepSeek dsh-superagent: an open-source coding agent that takes GitHub requests and replies with PRs (DeepSeek, Research)
ChatGPT and Codex share one limit: token-based allowance for Plus, Pro and Business from December (OpenAI, Preview)
Karpathy's HackerTube: 6,454 videos picked from 43M HN comments, with the data released (Andrej Karpathy, Beta)
[Late catch] Google Connected Apps Platform: one MCP server puts your service inside the Gemini app and Search (Google, Waitlist)

OpenAI's Agents API adds command auto-review, Claude Code fences writes to the working folder, and NVIDIA's open transcription model

Sat, 31 Oct 2026 22:30:00 GMT

Here's today's news. No big launches this time, so we picked updates developers can try right away. OpenAI's Agents API picked up Auto Review, which screens risky shell commands before they run, along with background compaction and an experimental sense of time. Claude Code 2.1.314 through 2.1.316 adds a permission that blocks file writes outside the working folder and a default effort setting for subagents.

NVIDIA's Nemotron 3.5 Transcribe came out on 10/28 and slipped past our last two issues, so it's here as a late catch. It's an open-weight speech model that handles transcription, speaker labels and translation in one checkpoint.

OpenAI Agents API update: command auto-review, background compaction and time awareness (OpenAI, GA)
Claude Code 2.1.314–2.1.316: block writes outside the working folder, default subagent effort, idle compaction (Anthropic, GA)
[Late catch] Nemotron 3.5 Transcribe: NVIDIA's open-weight model for transcription, speaker labels and translation in 40 languages (NVIDIA, GA)

Mistral opens Large 4's weights, Gemini Computer Use goes GA, and one robots.txt line to opt out of AI training

Fri, 30 Oct 2026 22:30:00 GMT

Here's today's news. Mistral put the weights for Mistral Large 4 (nicknamed Le Chonk), the 1-trillion-parameter model it previewed earlier this month, on Hugging Face. It says three more weeks of reinforcement learning lifted coding and defensive-security scores. Google made the Gemini API's Computer Use tool generally available, so you can now build agents that read screenshots and decide where to click and type on a supported, non-preview feature.

OpenAI, Google and others said they will honor a shared robots.txt signal (__usage_ai-training) that lets a site opt out of training by several AI companies at once. The draft was submitted to the IETF by an Anthropic engineer. We skipped Oct 30, so we're also adding the Gemini API CLI (beta, Oct 27), which the Oct 28 and Oct 29 issues missed, as a late catch.

Mistral Large 4 weights are out: a 1T-parameter open model tuned harder for coding and cyber defense (Mistral AI, GA)
Gemini API Computer Use is GA: screen-reading, clicking agents become a supported feature (Google, GA)
A shared robots.txt signal for AI training: one group tells several AI companies 'don't train on this' (IETF · Anthropic · Google · OpenAI, Preview)
[Late catch] Gemini API CLI beta: coding agents can call Gemini with one shell command (Google, Beta)

Google unveils Gemini 4 Carbon, Copilot Fusion goes GA, and Genie 4 learns to talk

Wed, 28 Oct 2026 22:20:00 GMT

Here's today's news. Two weeks after Gemini 4 Argon, Google announced its next frontier model, Gemini 4 Carbon, built for finding security flaws and for engineering tasks that run for days. For now it is limited to a group of cyber defenders through the Fairwind Program, and there's no date yet for developer API access.

At GitHub Universe, Copilot Fusion became generally available: you pick an effort level and Copilot decides which models to use and how to combine them. Google DeepMind also opened Genie 4, a world model that generates sound and dialogue alongside real-time video, as a research preview.

Gemini 4 Carbon: Google's new frontier model for security and long-running work, cyber defenders first (Google DeepMind, Pilot)
GitHub Copilot Fusion is GA: choose an effort level, let Copilot pick and combine the models (GitHub, GA)
Genie 4: Google DeepMind's world model now generates sound and dialogue, in research preview (Google DeepMind, Research)

GPT-6.1 Terra ships, Liquid AI opens the 2B-class LFM3 family, and Copilot gets server-to-server auth

Tue, 27 Oct 2026 21:49:59 GMT

Here's today's news. OpenAI released GPT-6.1 Terra, a mid-tier model built for reading screens and working across apps. Standard API pricing is half that of GPT-5.6 Terra and GPT-6.1 Sol, and it's already live in ChatGPT Work, Codex and the API.

Liquid AI published open weights for six LFM3 models, covering text, vision, audio and omni at roughly 2B parameters each. GitHub opened a public preview of server-to-server authentication, so services and agents can call Copilot on an organization's behalf without anyone signed in.

GPT-6.1 Terra: a computer-use model at half the API price (OpenAI, GA)
LFM3: six 2B-class open models from Liquid AI that run on phones and a Raspberry Pi (Liquid AI, GA)
Copilot server-to-server auth: services and agents call Copilot as the organization, no user login (GitHub, Preview)

Codex CLI 0.164 adds presets and keeps your cache, Block's speech-only Dessa, and the Mercury 3 diffusion LLM

Sat, 24 Oct 2026 21:56:46 GMT

Here's today's news. It was a quiet stretch heading into the weekend with no major launches, so today's picks lean practical. OpenAI shipped Codex CLI 0.164.0 and changed Codex so that switching reasoning effort mid-conversation no longer throws away your prompt cache.

Block unveiled Dessa, a full-duplex voice model trained from scratch on human speech alone, with no text. For now it's a waitlist aimed at safety researchers. We also missed Inception's diffusion LLM Mercury 3 (announced October 19) in earlier issues, so it's included today as a late catch.

Codex CLI 0.164: ten prompt preset slots, and changing reasoning effort keeps the cache (OpenAI, GA)
Block Dessa: a full-duplex conversation model trained only on human speech, no text (Block, Research)
[Late catch] Inception Mercury 3: a 1,000-tokens-per-second diffusion LLM with 1M context (Inception, GA)

Kimi K3.1 Pro, MiniMax M3.1 Flash weights, and Anthropic-managed Claude on Google Cloud goes GA

Fri, 23 Oct 2026 22:01:24 GMT

Here's today's news. Moonshot AI shipped Kimi K3.1 Pro, a 3T-class model with a 1M-token context window, to its API and apps, and says the weights will follow by Nov 4. MiniMax put the weights for MiniMax M3.1 Flash, its agent and coding model, on Hugging Face.

Anthropic made Claude Models on Google Cloud (Anthropic managed) generally available, bringing the full Claude Platform to Google Cloud accounts. We also include a late catch from earlier in the week: Claude API version 2026-10-21 and the major SDK releases that come with it.

Kimi K3.1 Pro: a 3T-class 1M-context model at $2 per million input tokens, weights due by Nov 4 (Moonshot AI, GA)
MiniMax M3.1 Flash weights: a 1M-context agent and coding model with 23B active parameters (MiniMax, GA)
Claude Models on Google Cloud (Anthropic managed) is GA: the full Claude Platform on your Google Cloud commitment (Anthropic, GA)
[Late catch] Claude API 2026-10-21: the first new API version in three years, new major SDKs, and no more sampling parameters (Anthropic, GA)

OpenAI's Decisions API goes GA, Recursive releases Flint, Agent Skills moves to the AAIF, 1Password launches agent sign-in

Thu, 22 Oct 2026 22:40:00 GMT

Here's today's news. OpenAI made the Decisions API generally available. It doesn't write a response; it picks from answers you define. Recursive released Flint, a 40B coding model that runs on a laptop, under Apache 2.0.

Anthropic handed the Agent Skills standard to the Agentic AI Foundation, which already hosts MCP. 1Password launched 1Password for Agents, which signs agents in without showing them your password. There was no major new frontier model, so today leans toward developer updates you can try now.

OpenAI Decisions API is GA: classify, route and pick next actions with no output-token charge (OpenAI, GA)
Recursive Flint: a 40B open coding model for laptops that uses 2.5B parameters per token (Recursive, GA)
Agent Skills joins the Agentic AI Foundation: after MCP, the skills standard gets a neutral home (Anthropic, GA)
1Password for Agents: an approval layer that signs agents in without showing them your password (1Password, Preview)

Claude Duet opens in preview, Mistral ships Voxtral 3, DeepSeek open-sources DeepInfer, OpenAI adds a 12-hour prompt cache

Wed, 21 Oct 2026 22:10:00 GMT

Here's today's news. Anthropic opened a waitlist for Claude Duet, an agent app that comes to you first with what needs attention. It starts with Pro and Max users in the US. Mistral released Voxtral 3: four speech recognition and synthesis models plus a voice agent SDK.

DeepSeek open-sourced DeepInfer, the inference engine behind its own API, under Apache 2.0. OpenAI added an option to keep API prompt caches for at least 12 hours.

Claude Duet: Anthropic's agent app that flags what matters and asks before it acts (Anthropic, Waitlist)
Voxtral 3: four Mistral speech models and a voice agent SDK (Mistral AI, Preview)
DeepInfer: DeepSeek open-sources the inference engine behind its production service under Apache 2.0 (DeepSeek, Beta)
OpenAI's 12-hour prompt cache: cached-input pricing even when requests are hours apart (OpenAI, GA)

Hugging Face × TorchTPU, Claude Code 2.1.305, [Late catch] Timestral

Tue, 20 Oct 2026 22:20:00 GMT

Here's today's news. There was no big launch today, so I picked practical updates. Hugging Face showed how to run Transformers and Diffusers models on TorchTPU, the PyTorch backend for TPUs.

Claude Code added /auto-mode-setup, a guided setup for auto mode. Mistral's time-series model Timestral came out on Oct 15 and was missed earlier, so it appears here as a late catch.

Hugging Face × TorchTPU: run Transformers and Diffusers models on TPU with device_map="tpu" (Hugging Face · Google, Beta)
Claude Code v2.1.303-305: /auto-mode-setup, usage warnings, gateway tool-search retry (Anthropic, GA)
[Late catch] Timestral: Mistral's first time-series foundation models, 5M to 209M, open weights (Mistral AI, GA)

OpenCode Console breach, ai-memory 3.1, and a late catch: Ultramarine 1 Flash

Sun, 18 Oct 2026 22:40:00 GMT

Here's today's news. No big launches today, so we picked practical updates. OpenCode, the open-source coding agent, said on its official X account that external provider API keys belonging to 1.4% of its Console users may have been exposed.

ai-memory, an open-source tool that lets coding agents share one memory, shipped version 3.1 with a supported Windows install. We also add Ultramarine 1 Flash, a knowledge-work model released on October 14 that earlier issues missed, as a late catch.

OpenCode Console breach: provider API keys of 1.4% of users may be exposed (OpenCode (Anomaly), GA)
ai-memory 3.1: shared memory for Claude Code, Codex and other coding agents, now with supported Windows install (akitaonrails (open source), GA)
[Late catch] Ultramarine 1 Flash: a low-cost model for finance, legal and consulting work, weights promised (Paper Instruments, Preview)

DeepSeek V4.1-Flash-1017, Reflection Beam Weights Released, Claude Code v2.1.302

Sat, 17 Oct 2026 22:00:00 GMT

Here's today's news. No big launches today, so I picked practical updates. DeepSeek released the MIT-licensed V4.1-Flash-1017 weights and opened its own search and fetch APIs.

US startup Reflection AI released the weights of Beam, the 501B coding and agent model it previewed in early October, under Apache 2.0. Claude Code v2.1.302 turned fork subagents on by default.

DeepSeek V4.1-Flash-1017: MIT Weights, Built-in Search and Fetch APIs, Higher Agent Benchmarks (DeepSeek, GA)
Reflection Beam Weights Released: A 501B Coding and Agent Model Under Apache 2.0 (Reflection AI, GA)
Claude Code v2.1.302: Fork Subagents On by Default, Pick the Auto Mode Classifier Model (Anthropic, GA)

Claude Code Review goes GA, GPT-Live voices from a prompt, Elorian Glimpse, and Perplexity Decider v1.2

Fri, 16 Oct 2026 22:00:00 GMT

Here's today's news. No big launch this time, so the picks lean practical. Anthropic made Claude Code Review generally available on Team and Enterprise plans, and admins can now set each repo to Quick, Standard, or Deep reviews.

OpenAI opened prompt-to-voice to every API customer: describe a voice in plain text and GPT-Live will build it. A Korean voice, Noeul, is among nine new ones, and voice mode in ChatGPT moved to GPT-Live-1.1.

Startup Elorian released Glimpse, its first model, which keeps going back to the image or video while it reasons. Perplexity shipped Decider v1.2, open-weight models that pick one answer from a fixed set and attach a probability to it.

Claude Code Review is GA, with Quick, Standard, and Deep review levels per repo (Anthropic, GA)
GPT-Live prompt-to-voice is GA: describe a voice in text, plus a new Korean voice, Noeul (OpenAI, GA)
Elorian Glimpse: a 'visual thinking' model that keeps looking back at the image while it reasons (Elorian, Preview)
Perplexity Decider v1.2: 27B and 4B open-weight models that pick an answer and give you its probability (Perplexity, GA)

Meta opens up Muse Spark Open, NVIDIA ships Nemotron 3.5 Super VL and Bolt, Grok Voice 3.0, and OpenTelemetry for the Claude API

Thu, 15 Oct 2026 22:40:00 GMT

Here's today's news. Meta released Muse Spark Open, an open-weight version of Muse Spark. The 816B-parameter model activates only 42B per token and is tuned for coding and agent work. Two Mac Studios can run it.

NVIDIA put out Nemotron 3.5 Super VL, a 120B multimodal agent model, along with Bolt 74B-A9B, a local model that fits in 24GB. xAI shipped Grok Voice 3.0 for phone support agents, at the same API price as before.

Anthropic opened a beta that exports Messages API and Managed Agents requests as OpenTelemetry traces. Claude calls now show up in the observability tools you already run.

Muse Spark Open: Meta's 816B open-weight model for coding and agents (Meta, GA)
Nemotron 3.5 Super VL and Bolt: a 120B server model and a 24GB local agent model (NVIDIA, GA)
Grok Voice 3.0: a speech-to-speech model for support calls, at the same price (xAI, GA)
OpenTelemetry export for the Claude API (beta): trace Messages API and Managed Agents calls in your own tools (Anthropic, Beta)

Liquid's laptop-sized agent model, Fish Audio's 16-voice Drama 3 TTS, and idle compaction in Claude Code

Wed, 14 Oct 2026 22:20:00 GMT

Here's today's news. No big launches today. Instead, this issue covers open-weight models and tool updates you can download and try right now.

Liquid AI released LFM2.5-24B-A2B, a 24B-parameter model that activates about 2B per token and is aimed at tool-calling agents running on a laptop. Fish Audio put out Drama 3, a text-to-speech model that performs multi-character scripts with emotional direction, weights included.

Claude Code 2.1.299 compacts long conversations that have sat idle before the prompt cache expires, so picking a session back up costs less.

LFM2.5-24B-A2B: Liquid AI's on-device agent model wakes just 2B parameters per token (Liquid AI, GA)
Fish Audio Drama 3: open-weight TTS that voices up to 16 characters in one script (Fish Audio, GA)
Claude Code v2.1.299: idle long conversations get compacted before the cache expires (Anthropic, GA)

Google ships Gemini 4 Argon, and ColGrep-Agent moves coding-agent search onto your laptop

Tue, 13 Oct 2026 22:20:00 GMT

Here's today's news. The big one is Gemini 4 Argon, Google's new frontier model. It went GA in the Gemini API on launch day at an introductory $2 per million input tokens. Independent benchmarks put it near the top, but not first.

For developers, LightOn open-sourced ColGrep-Agent, a local 2B model that handles code search for your coding agent so it spends fewer API tokens. And Google Play now lets you run Android apps you vibe-coded in AI Studio on your own phone without a developer account.

Gemini 4 Argon: Google's new frontier model launches with GA API access (Google DeepMind, GA)
ColGrep-Agent-2B: hand your coding agent's code search to a local 2B model and save a third of the tokens (LightOn (open source), Beta)
Google Play personal apps: run your AI Studio-built Android app on your own phone, no developer account (Google, GA)

Meta launches Muse Wonder for agents; GPT-6.1 Sol's default context jumps to 500k

Mon, 12 Oct 2026 22:30:00 GMT

Here's today's news. The lead is Meta's new frontier model, Muse Wonder. It's built for agents around two ideas, Active Memory and Active Focus, and the standard version is available through an API today. Independent scores put it near the top, but not in first place.

OpenAI raised GPT-6.1 Sol's default context in Codex and ChatGPT Work from 272k to 500k tokens. Sakana AI published MASS, a research method where one model works as a team and generates its own training data, along with code.

Mistral Large 4 (1T parameters, open weights promised) landed on Oct 6, just before our first issue's window. Its weights are due by month's end, so we've added it here as a late catch.

Meta Muse Wonder: a frontier model for agents that remembers and keeps pushing (Meta, GA)
GPT-6.1 Sol's default context in Codex and Work goes from 272k to 500k, with no long-context fee on subscriptions (OpenAI, GA)
Sakana AI's MASS: one model forms a team and writes its own training data (Sakana AI · UC Berkeley, Research)
[Late catch] Mistral Large 4 preview: a 1T-parameter model with open weights on the way (Mistral AI, Preview)

REA 6.4 brings reverse engineering to coding agents; Copilot for JetBrains gets admin controls

Sun, 11 Oct 2026 22:20:00 GMT

Here's today's news. There were no major launches this time, so we picked practical updates developers can try today. It was a weekend, and none of OpenAI, Google or Anthropic shipped a new model or product inside the 48-hour window.

Microsoft-Decision-1 (Oct 9, US time) and Codex's Dots and Composer predictions (Oct 9) got plenty of attention, but both turned out to land just before the window, so they're left out of this issue. You'll find them in the research log.

Google AI Edge Foresight, an offline Mac meeting-notes app released Oct 8, belonged in our Oct 9 issue. It slipped through, so we've added it here as a late catch.

REA 6.4: a reverse-engineering MCP for coding agents adds Android and JEB support (morluto (open source), GA)
Copilot for JetBrains: admin-set default models, one-click fixes from diagnostics, and an off switch for MCP autostart (GitHub, GA)
[Late catch] Google AI Edge Foresight: an offline meeting-notes app that runs on your Mac (Google, Preview)

Claude Dashboards and Motion in beta, Claude Code 2.1.296

Sat, 10 Oct 2026 21:55:00 GMT

Here's today's news. Again, no major new model came out. The news is Anthropic's workplace features (Dashboards, Motion) and a routine Claude Code release.

The official OpenAI, Google, Meta and Apple sites, plus GeekNews and TechCrunch, would not load in this run. So this issue leans toward Anthropic news.

Claude Dashboards and Motion in beta: live data dashboards and animated explainers (Anthropic, Beta)
Claude Code v2.1.296: earlier subagent compaction, one model for workflows (Anthropic, GA)

Claude dynamic workflows and the Anthropic Cyber Mission

Fri, 09 Oct 2026 21:50:00 GMT

Here's today's news. No major model launch landed in the past 48 hours. Instead, Anthropic shipped a developer feature for agents and a security program.

Google's Gemini work agent (Oct 8) was covered in the last issue, so we skip it here. OpenAI, Google and Meta official sites could not be opened in this run.

Claude Managed Agents dynamic workflows (beta): agents write their own multi-agent jobs (Anthropic, Beta)
Anthropic Cyber Mission: open-source security scans and critical-infrastructure defense (Anthropic, Pilot)

Claude Haiku 5.5 and GPT‑6 in ChatGPT lead a busy 48 hours

Fri, 09 Oct 2026 12:30:00 GMT

Here's today's news. Two big launches came out. Anthropic released Claude Haiku 5.5, a small model that costs much less, and OpenAI made GPT‑6 with Intelligent UI generally available in ChatGPT.

On the agent side, Google previewed a Gemini work agent that gets its own business account. The most practical update for developers is OpenAI's Ultrafast mode for GPT‑6.1 Sol.

Claude Haiku 5.5: a small model at a much lower price (Anthropic, GA)
GPT‑6 and Intelligent UI roll out across ChatGPT (OpenAI, GA)
Gemini's unified work agent: an AI coworker with its own email account (Google Cloud, Preview)
GPT‑6.1 Sol Ultrafast mode: high-speed inference in the API and Codex (OpenAI, GA)