Moonshot releases Kimi K3.1 Pro weights, Cognition launches SWE-2 Flash, RoboColosseum evaluates robots in real-world conditions

Moonshot AI delivers on its promise and releases the weights of Kimi K3.1 Pro, a 2.8T-parameter MoE, on Hugging Face under an MIT-style license with conditions for the largest resellers. Cognition launches SWE-2 Flash, a coding model built on Inkling-Small and five times cheaper than SWE-2, and RoboColosseum evaluates robotic policies on more than 100 real robots. OpenAI is set to acquire Pierre, whose tools already power Codex, while VS Code 1.145 brings Copilot, Claude and Codex together in a single session and GitHub Copilot joins Xcode’s coding assistant.

Moonshot AI releases the weights of Kimi K3.1 Pro, a 2.8T-parameter MoE, on Hugging Face as promised

November 4 — When Kimi K3.1 Pro launched on October 23, Moonshot AI promised the full weights by November 4 at the latest: it has delivered. The moonshotai/Kimi-K3.1-Pro repository, created on November 2 and announced by @Kimi_Moonshot on November 4 at 16:00 UTC, is freely accessible, with no form to fill out: 96 safetensors files, approximately 1.56 TB in total.

The model card details what the launch announcement summarized. Kimi K3.1 Pro, which Moonshot AI describes as its most capable “3T-class” model, is natively multimodal (text and image) and post-trained from Kimi K3. One architectural change from Kimi K3 is highlighted: its Gated MLA full-attention layers give way to 24 Gated DeepSeek Sparse Attention (Gated DSA) layers, which are more efficient over long contexts. The weights ship in MXFP4, with MXFP8 activations, a format learned during training (quantization-aware training).

Technical specification	Value according to the model card
Total and active parameters	2,8T, including 104B active
Network layers	93, including 69 KDA and 24 Gated DSA
MoE experts	896, including 16 per token and 2 shared
Context window	1 048 576 tokens
Weight format	MXFP4, activations in MXFP8
Vision encoder	MoonViT-V2, 401M parameters
Files released	96 safetensors, approximately 1,56 To

Moonshot AI recommends vLLM, SGLang and TokenSpeed for serving the model, which is also available through the platform.kimi.ai API (identifier kimi-k3.1-pro). Together AI, listed among the page’s inference providers, already serves it and offers post-training (see the Briefs). The model card stresses one integration requirement: in multi-turn conversations or those involving tool calls, the assistant’s complete message, including its reasoning, must be sent back to the model unchanged. It repeats the scores presented at launch: 65.7 on Terminal-Bench 4.0, just behind Claude Opus 5.5 (66.4), and 51.7 on SWE Marathon v1.1, the highest score in the table, with Kimi K3.1 Pro evaluated using Claude Code as its harness.

The license, named “Kimi K3.1-Pro License,” follows the framework of an MIT license (permitting use, modification, redistribution and sale) with two conditions. A company operating a Model as a Service business, giving third parties access to inference or fine-tuning, whose cumulative revenue exceeds $20 million over twelve consecutive months must sign a separate agreement with Moonshot AI before any commercial use. A commercial product with more than 100 million monthly active users, or more than $20 million in monthly revenue, must prominently display “Kimi K3.1-Pro.” Neither internal use nor use through Moonshot AI’s official products or its certified inference partners is affected.

🔗 Kimi K3.1 Pro on Hugging Face · Moonshot AI’s announcement on X

Cognition launches SWE-2 Flash, a coding model built on Inkling-Small and five times cheaper than SWE-2

November 4 — Cognition launches SWE-2 Flash, a fast and efficient coding model, available immediately in Devin Desktop and the Devin CLI, and free until November 30 (for Pro, Max and Teams subscribers, the announcement thread specifies). A change of base: whereas SWE-2, released on September 10, was post-trained on Kimi K3, SWE-2 Flash starts from Inkling-Small, Thinking Machines’ open-weight model (276B parameters, including 12B active). Cognition describes it as a fully American coding model. According to the company, it costs five times less than SWE-2 and runs three times faster, with no absolute pricing published; the speed demonstration (200 versus 60 tokens per second) is a simulated replay of a task.

Introducing SWE-2 Flash, our new fast & efficient coding model. On leading evals, it outperforms all American open-weight models, even ones 3x larger. To balance transparency and cyber risks, we are building a program to securely share the weights with partners & researchers. — @cognition on X

The main innovation lies in the recipe. For the first time, Cognition inserts a mid-training phase before reinforcement learning: several trillion tokens of scientific papers, blogs and code, in three stages that save the rarest and most carefully curated data for last. Large-scale RL follows, with gains coming primarily from more varied and better-verified tasks. The reported improvement over the base model: Terminal-Bench 4.0 rises from 0.9% to 20.0%, FrontierCode 2.0, its in-house benchmark, from 3.2% to 24.0%, and ExploitBench from 11.3% to 34.8%.

Benchmark (%)	SWE-2 Flash	Inkling-Small (base)	Muse Spark Open	GLM 5.3 Flash	SWE-2
Size (total-active)	276B-A12B	276B-A12B	816B-A42B	320B-A18B	2,8T-A104B
FrontierCode 2.0	24,0	3,2	6,2	14,2	23,8
Terminal-Bench 4.0	20,0	0,9	8,0	41,8	27,3
DeepSWE 1.1	60,1	17,4	60,0	63,4	73,0
SWE-Atlas QA	50,0	27,4	54,0	56,5	56,3
ExploitBench	34,8	11,3	—	36,0	36,1

The blog post’s table qualifies the headline claim. Meta’s Muse Spark Open, three times larger, retains the advantage on SWE-Atlas QA (54.0 versus 50.0) and is on par on DeepSWE 1.1. Against Chinese open models, the results are mixed: SWE-2 Flash outperforms GLM 5.3 Flash and DeepSeek V4.1 Flash on FrontierCode 2.0, but remains far behind them on Terminal-Bench 4.0 (20.0 versus 41.8 and 42.3). It narrowly beats SWE-2 on FrontierCode 2.0 without catching up on any other row.

The weights have not been released for now. Cognition targeted cybersecurity during training and points out that these capabilities also serve attackers, as a recent wave of cyberattacks has shown, according to the company. It is preparing a program to securely share the weights with industry partners and trusted research labs, which could fine-tune the model and run it in sensitive environments; details are promised in a few weeks.

🔗 Introducing SWE-2 Flash (Cognition)

Coding agents: Claude Code 2.1.318, VS Code 1.145, Copilot in Xcode, Vibe CLI 3.1.0 and Antigravity 2.29.0

Claude Code 2.1.318 flags expiring logins, and Haiku 5.5 writes subagent status lines

November 4 — Claude Code 2.1.318 has 84 entries, including 47 fixes. It arrived on npm on November 3 at 21:55 UTC, with its release notes published on GitHub shortly after midnight UTC; npm’s stable channel remains on 2.1.311. The most useful addition concerns stored claude.ai logins, which expire: their expiration date now appears in claude auth status (loginExpiresAt) and in the JSON results of claude -p (credential_expires_at), and claude -p issues a warning on stderr ten days before the deadline. An automated task no longer discovers the expiration when it fails.

Area affected	What’s new in 2.1.318
Cloud sessions	No repository option: a session in an Anthropic-hosted environment, without selecting a repository
Background subagents	A shell command moves to the background after 3 minutes so that the prompt cache does not expire
Administration	`deniedMcpServers` also blocks Anthropic-provided servers whose upstream URL matches, `assume_role` on `mantle` upstreams
IDE extension	No longer installed at startup in an IDE’s terminal: use `/ide`
Self-hosted runner	New orchestrator log format; scripts that search for `retryable=true:` need reviewing

On the same day, @ClaudeDevs announced a change that none of the release notes, from 2.1.316 to 2.1.318, describes. The status line of a background subagent, refreshed every 30 seconds, previously reread its entire conversation; it is now written by Haiku 5.5 from short notes, while the subagent keeps its own model. According to Anthropic, each update costs about 99% less, and sessions with background subagents consumed about 5% less of the usage limit, allowing roughly 6% more prompts. The change applies to most Pro, Max and Ultra subscribers and all Team and Enterprise subscribers, with Claude Code 2.1.316 or later; Lydia Hallie, from the Claude Code team, clarifies that it only affects the CLI.

🔗 Claude Code 2.1.318 release notes · @ClaudeDevs’ thread on Haiku 5.5

VS Code 1.145 combines Copilot, Claude and Codex harnesses in a single session

November 4 — VS Code 1.145 continues its shift toward agents. The main addition, in preview: a single session can bring together nested conversations running on different harnesses, for example Claude reviewing a plan written by a Copilot conversation, or Codex being assigned the same task, side by side. Each conversation retains its model, agent and permission levels. The idea builds on Compare Models, which arrived in preview with VS Code 1.143.

Local automations now support scheduling through cron expressions and can use a custom workspace agent; cloud automations cannot. For remote agents, the Agents window’s integrated browser connects through the remote host, allowing previews of the application launched by the agent on that machine (preview). The Copilot harness remains the default choice for new sessions, with a migration guide from the Local harness, which loses three tools (#installExtension, #runCommand and #newWorkspace) as well as the /new command: custom agents and prompt files that reference them need updating. For enterprises, the organization-managed list of allowed MCP servers now takes precedence over the registry restriction.

🔗 VS Code 1.145 release notes

GitHub Copilot joins Xcode’s built-in coding assistant in public preview

November 4 — GitHub Copilot now works in Xcode’s built-in coding assistant, in public preview starting with Xcode 27.3 beta, on a Mac running macOS Tahoe 26.6 or later. The integration is configured in Xcode’s settings, under Intelligence, and works with all Copilot plans. It offers a second entry point for Apple developers: the separately distributed GitHub Copilot for Xcode extension remains available. GitHub does not specify which models are offered or which features are available through Apple’s assistant, and is collecting feedback in the extension’s repository or through Apple’s reporting tool (Feedback Assistant).

🔗 GitHub Copilot in Xcode’s coding assistant

Vibe CLI 3.1.0 adds /context, a monthly usage gauge and image reading

November 4 — Mistral releases Vibe CLI 3.1.0, the first minor release since 3.0.0 on October 19, with no announcement on X: 67 entries (11 additions, 25 changes, 29 fixes and 2 removals). The /context command shows what fills the context window (system prompt, tool definitions, skills, conversation, free space and auto-compaction reserve), with an estimated token count for each category. The footer displays the share of monthly Vibe Code usage consumed, first in yellow and then in red, or PAYG when pay-as-you-go billing takes over. The read_file tool also opens images with a model that can read them, Ctrl+R searches prompt history (voice recording moves to Ctrl+T), and an experimental Mistral MCP connector arrives.

Several fixes target Mistral Large 4: a turn started without a thinking level no longer fails with a 400 error, the selector offers Low, High and Max, and smart approve mode no longer requests approval for every tool call with models that cannot disable their reasoning, such as GLM 5.3. The smart_approve_default setting is removed in favor of default_agent = "smart-approve".

🔗 Vibe CLI 3.1.0 release notes

Antigravity 2.29.0 begins rolling out a Gemini Live voice mode and remote agents in the cloud

November 3 — Version 2.29.0 of the Antigravity app (22 improvements, 13 fixes) begins gradually rolling out two features. Voice mode, powered by Gemini Live, lets users speak to Antigravity and give the agent requests by voice. Remote agents (remote agents) are arriving for Google AI Ultra and Pro subscribers: conversations in a cloud project run in the cloud rather than on the computer, and remain accessible from antigravity.google.com/remote. Antigravity already supported remote operation with Remote Control since August, but the session ran on the user’s machine; this time, execution itself moves off the computer.

The release also adds a native Jupyter notebook viewer, a redesigned automations page (scheduled, always active, paused) and, in the built-in browser, downloads and browsing history. Among the fixes, Git worktrees created for sub-agents now carry over uncommitted local changes.

🔗 The Antigravity changelog

RoboColosseum evaluates robotics policies on a fleet of more than 100 real robots

November 4 — RoboColosseum opens its doors: an open evaluation initiative that measures robotics policies outside laboratories and simulators, on a fleet of robots deployed in real-world conditions (in-the-wild), such as Unitree G1 humanoids in homes or YAM arms in restaurants. Its website says it was founded by NUS MAGIC Lab, BitRobot, EAIC Singapore and Hugging Face, which describes itself as a founding partner. The launch thread announces more than 100 robots, more than 140 scenes and 10 000 hours of free evaluation, allocated through applications.

Robotics has an evaluation problem. Most benchmarking efforts have measured progress in sim or labs. — @RoboColosseum on X

Two tracks run side by side. The Open Track pits two anonymous models against each other, giving them the same previously unseen instruction in the same scene: a human evaluator chooses the better execution or declares a tie. The Fine-tune Track fine-tunes each model on the same task using the same 100 demonstrations, then gives it 25 attempts. The first campaign pits five vision-language-action (VLA) policies against three frontier models.

Evaluated policy (Franka, Open Track)	Model publisher	DCR score (±1 standard deviation)	Blind comparisons
GPT-6 Astra	OpenAI	1 925 ±38	194
MolmoAct2	Ai2	1 869 ±24	1 859
G0.5	Galaxea	1 744 ±27	1 626
Claude Opus 5.5	Anthropic	1 636 ±47	206
GR00T N1.7	NVIDIA	1 350 ±36	1 797
π0.5	Physical Intelligence	1 285 ±43	1 845
LingBot-VLA v2	Robbyant	1 233 ±56	1 603
Grok 4.7	SpaceXSI	1 160 ±128	193

On the Franka arm, the Open Track’s online leaderboard (data from November 4, updated hourly, 4 662 blindly judged comparisons) puts GPT-6 Astra in the lead, ahead of Ai2’s MolmoAct2, but the gap remains within the error bars, and the frontier models have only around 200 comparisons each, compared with more than 1 600 for each VLA. Ai2 emphasizes that MolmoAct 2 finishes second without additional training for these tasks.

The paper accompanying the launch, posted on arXiv on November 2, qualifies this picture based on 5 699 comparisons conducted across 124 sites: frontier agents compete with specialized VLAs, but no system reliably completes open-world tasks. After fine-tuning, GR00T N1.7 takes the lead on SO-101 and π0.5 on Franka and the bimanual YAM station, and Open Track rankings are poor predictors of rankings after fine-tuning. On the G1 humanoid, no policy completes any of the five tasks, with average progress ranging from 13.3% to 23.2%. All executions, including failures, are published on Hugging Face in LeRobot format.

🔗 The RoboColosseum launch thread · The RoboColosseum leaderboard · The arXiv paper 2611.02497

OpenAI to acquire Pierre Computer Company, whose Trees and Diffs tools already power Codex

November 4 — OpenAI announces that it will acquire Pierre Computer Company, a small developer tools company. Its commercial product, Code Storage, is an API that lets applications store code and track its history with Git; its open source libraries Trees, for browsing project files, and Diffs, for viewing changes, are already integrated into the Codex app.

Today we’re announcing that OpenAI will acquire @pierrecomputer to accelerate our vision for making Codex a place where teams can build software together with AI, from first idea to the version they ship. Pierre’s open-source developer tools already power parts of Codex today. Together, we’ll build on that foundation for the agentic era. — @OpenAINewsroom on X

OpenAI links the deal to Pando, the Git backend for its monorepo unveiled two days earlier: the company wants to set the standards for software development, from the first branch created by an agent through deployment, and advance version control (source control) across the industry once the acquisition is finalized. Pierre’s open source projects will remain open, and Code Storage will remain available to its customers while its features are integrated into Codex. The deal still requires customary regulatory approvals; neither the price nor the timeline has been disclosed.

🔗 OpenAI to acquire Pierre Computer Company

AI for science: a bacteriophage vault with Claude, NeuralGCM+ at Ai2 and FVCOM ported to Rust

Anthropic describes the 3D structure of a bacteriophage-derived vault with Claude

November 4 — Anthropic’s science team publishes the 3D structure of a vault produced from a bacteriophage protein, a first in prokaryotes according to the company. Vaults are large, hollow protein shells, roughly three times heavier than a ribosome, discovered in 1986 and whose function remains unknown. This one, named L05 and produced in the bacterium E. coli, contains 96 copies of the same protein, compared with 78 in mammals, weighing around 9.5 megadaltons and occupying around 25% more volume. A small unidentified molecule, nestled inside each of the 96 copies, resembles coenzyme A in shape: an initial clue to a function, if confirmed.

Claude contributed at every stage: searching for relatives of the protein among roughly 15 billion predicted proteins with Claude Science, reducing purification from around three days to around 90 minutes, and producing an initial structural model in 34 hours instead of several weeks, according to the team. Anthropic also details its mistakes: a protocol containing an additive that hinders DNA entry into cells cost two weeks, and a skipped standard step in cryo-electron microscopy produced a map mixed with its mirror image. The preprint has been submitted to bioRxiv; Leonard Rome, a co-discoverer of vaults, advised the team as a paid consultant.

🔗 Investigating vault particles with Claude

Ai2 launches NeuralGCM+, an open climate model funded by Google.org over three years

November 4 — Ai2 launches a climate modeling program funded by Google.org with several million dollars over three years (the exact amount has not been published), with technical support from Google Research. It is intended to produce NeuralGCM+, an open source model that extends Google Research’s NeuralGCM: a hybrid model that solves physical equations for large-scale dynamics and delegates finer-scale processes, such as cloud formation and precipitation, to AI. In a public benchmark, an atmosphere-only version of NeuralGCM simulated one year of the global atmosphere in eight minutes on a single TPU.

Ai2 aims to couple it with the ocean and sea ice to sustain stable simulations over several decades, while reducing regional temperature and rainfall errors over the past 50 years. The roles are divided: Google Research focuses on forecasts weeks or months ahead, while Ai2 targets trends over years and decades, such as changes in rainfall and water availability by mid-century in the southwestern United States. Models and modeling frameworks will be published with documented workflows.

🔗 Ai2’s post on the partnership with Google

Google ports the core of FVCOM, a 174 000-line Fortran ocean model, to Rust with Gemini 4 Argon

November 3 — Another way to equip ocean modeling: on the Antigravity blog, Dan Gruhl and Prashant Nair from Google’s AI R&D team describe how Gemini 4 Argon modernized FVCOM (Finite-Volume Community Ocean Model), a coastal ocean model comprising 174 774 lines of Fortran 90 and C across 264 files. From a single prompt, the agent ported its core to fully memory-safe Rust in 59 minutes, under four constraints: no unsafe code, no allocations in computation loops, files under 400 lines and results identical to the Fortran version across 100 random seeds.

The audit uncovered six decades-old bugs in the original code, including an integer overflow on grids with more than 1 300 nodes. The 100 simulations match the Fortran results to 12 decimal places.

Migration stage	Execution time	Measured speedup	Peak memory
Original Fortran 90 and C (-O3)	883 ms	1,00x	11,84 Mo
Idiomatic safe Rust	682 ms	1,29x	6,39 Mo
Structure of arrays (SoA), with no allocations in loops	548 ms	1,61x	5,83 Mo
LLVM profile-guided optimization	469 ms	1,88x	5,48 Mo

Google presents the method as an open source skill, migrate-to-rust, available in the google/skills repository since October 13.

🔗 Antigravity’s post on FVCOM

APIs and developer platforms: Perplexity’s Decisions API, Grok’s default watermark, Together Compiler and Cohere on OpenRouter

Perplexity’s Decisions API searches the web before answering

November 4 — Perplexity adds web grounding (web grounding) to its Decisions API, the decision-model API launched on October 1, which responds with probabilities rather than text. Until now, the model knew only its training data and the context (state) sent by the developer. With the "mode": "web_grounding" setting, the gateway first runs one Perplexity Fast Search per question (up to 100 per request), adds page titles, URLs and text to the context, and lists the sources in search_results in the response. Each search costs 0.001 dollars in addition to input tokens (0.02 dollars per million for pplx-decider-v1.2). With the Python and TypeScript SDKs, the parameter is passed in the request body.

Metric published by Perplexity	Without web grounding	With web grounding
Internal test, 328 statements	56,4 %	97,3 %
Internal test, median latency	139 ms	366 ms
PWS-Decisions, 100 public questions	51 %	86 %

On the public PWS-Decisions benchmark, Perplexity also measures 370 ms median latency and 685 ms at the 95th percentile with web search.

🔗 Web grounding for the Decisions API

SpaceXSI will watermark text from its Grok models in the API by default starting November 7

November 3 — Version 0.2.8 of SpaceXSI’s TypeScript SDK surfaced a documentation page updated on October 31: starting November 7, 2026, text produced by grok-4.7, grok-4.6, grok-4.5, grok-4.3 and the three variants of grok-4.20 will be watermarked by default in the API to comply with the European AI Act. The invisible watermark slightly steers token selection without adding characters; it covers all text APIs, including structured outputs and tool-call arguments, across all endpoints. The x-watermark header signals a watermarked response, and access to the statistical detection API is available only by application.

Only team administrators can disable it, model by model, with the change taking effect within 5 minutes, by attesting that a legal exception applies and that the team assumes sole responsibility for its marking obligations; some models may remain locked. Anthropic expanded its own text watermark on November 3, while OpenAI’s remains disabled by default in the API.

🔗 Output Watermarking, SpaceXSI documentation

Together AI launches Together Compiler, agents that continuously optimize inference for its open models

November 4 — Together AI entrusts agents with optimizing its inference engine. Together Compiler starts with the model, hardware and traffic pattern: a swarm of agents proposes kernel changes, a persistent engine tests each for speed and accuracy against the original engine, and only verified changes reach production. Together calls this continual compilation (continual compilation), rerun when the model, hardware or workload changes.

The reported gains are Together’s own: up to 2.6 times more tokens per GPU at the same per-user speed on Kimi K3, GLM 5.3, MiniMax M3.1-Flash and DeepSeek V4.1-Flash, after 24 hours of compilation on two nodes per model, compared with the best configurations published on SemiAnalysis’s InferenceX benchmark; up to 50% faster on customer workloads; 1 111 and 1 205 tokens per second at low latency on MiniMax M3.1-Flash and DeepSeek V4.1-Flash. SemiAnalysis says it validated one of the results. The compiler is already enabled by default on DeepSeek V4.1 Flash, GLM 5.3 Flash and GLM 5.3 in serverless mode. Together describes it as the first inference system with recursive self-improvement (recursive self-improvement) at production scale, a claim attributable to the company alone.

🔗 Together Compiler on the Together AI blog

Cohere makes Embed 5 and Rerank 5 available on OpenRouter

November 4 — All four variants of Cohere’s retrieval models (Embed 5 Pro and Fast, Rerank 5 Pro and Fast) are coming to OpenRouter, usable with existing OpenRouter keys and credits, at Cohere API prices: $0.12 and $0.08 per million tokens for embeddings, $0.05 and $0.02 for reranking (reranking). Two limitations specific to OpenRouter: vectors are returned as floats (int8 and binary are available through the Cohere API), and Rerank 5 remains limited to text. The post, a retrieval-augmented generation (RAG) tutorial, notes that according to Cohere’s tests on 40 datasets, querying an index built with Embed 5 Pro using Embed 5 Fast retains 98.4% of the quality of an all-Pro search.

🔗 Cohere retrieval models on OpenRouter

Voice: Gemini 3.8 Flash TTS voice remixing and ElevenAgents in contact centers

Gemini 3.8 Flash TTS gains remixing, which creates a new voice from an existing one

November 4 — Google is adding a third voice creation mode to Gemini 3.8 Flash TTS, following design (Voice design) and replication, launched on September 23. Voice remixing (Voice remixing) starts with an existing voice, either one of the more than 2,000 voices in the library or a custom-designed voice, and modifies it through natural-language instructions: timbre, pitch, pace, accent, or age. Google’s examples include a slight Southern US accent or the voice of a fire-breathing dragon.

On the API side, the POST /v1beta/voices call with gemini-3.8-flash-tts creates a new stored voice, independent of the original, usable with Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS, which can itself serve as the basis for another remix. In AI Studio, the Remix this voice button offers three takes to listen to side by side. A project on a paid tier (Tier 1 or higher) is required; each project stores up to 200 voices, and a voice is retained for one year after its last use.

🔗 Google’s announcement on X · The Voice remixing guide

ElevenLabs natively integrates ElevenAgents with NiCE CXone, Genesys Cloud, Amazon Connect, and Five9

November 4 — ElevenAgents, ElevenLabs’ voice agent platform, now integrates natively with four major contact center platforms, adding voice agents to existing call flows without changing telephony, routing, or CRM. For any other platform, ElevenAgents connects over SIP (Session Initiation Protocol), with telephony providers such as Twilio, Vonage, Telnyx, Plivo, Bandwidth, or Exotel.

Contact center platform	Announced integration method
NiCE CXone	Native integration, CXone retains telephony, routing, and reporting
Genesys Cloud	Native Audio Connector, without a SIP connection (SIP trunking)
Amazon Connect	The Amazon Connect AI agent brings in ElevenAgents, then takes back control
Five9	SIP through Five9 AI Agent Connect, a Five9 add-on

The post highlights more than 17,000 voices in over 90 languages and cites AMC Health, which chose ElevenLabs partly because Genesys was already its telephony platform. It provides neither pricing nor performance figures and builds on the partnership with Genesys announced on September 3.

🔗 ElevenAgents integrates with top contact center platforms

Generative video and music: Kling 4.0 Flash for everyone and Suno starting from a MIDI file

Kling makes Kling 4.0 Flash available to everyone, on its website and through the API

November 4 — Kling 4.0 Flash, the fast variant of Kling 4.0, in limited early access since September 28, is now open to everyone on the Kling AI website and through the API. Kling says it can also be called through its MCP server, CLI, and plugins, including those for Adobe Photoshop, Premiere Pro, and After Effects. The announcement provides no maximum duration, resolution, or pricing, and Kling’s release notes stop at the launch of Kling 4.0 on October 15. During early access, Kling 4.0 Flash produced shots lasting 3 to 20 seconds in 720p; there is no indication whether these limits have changed.

🔗 Kling’s announcement on X

Suno lets users start a song from a MIDI file

November 4 — Suno now accepts a MIDI file as the starting point for a song: drag a .mid or .midi file onto the Create page, or select it from the “+” menu, then describe in natural language what Suno should do with it: play the notes exactly as written, or take a motif and develop it or assign it to other instruments. Unlike Remix and Cover, which start from an audio file (WAV or MP3), MIDI conveys the exact notes and rhythm, providing more precise control. The post gives examples of arranging a MIDI file for a full orchestra, a gospel choir with piano, or a string quartet faithful to the chords. At launch, the feature uses Simple Create mode and the v6-wild model and is limited to Pro and Premier subscribers; it complements the MIDI export added on July 26.

🔗 Start Songs with MIDI

ChatGPT dots become generally available for Enterprise, Edu, and Healthcare, enabled by default on November 18

November 4 — Dots, ChatGPT’s persistent agents launched on September 29 for Pro and Business Premium plans with an Enterprise beta, are becoming generally available for eligible ChatGPT Enterprise, Edu, and Healthcare workspaces. An admin-only preview runs until November 18, 2026: until then, dots remain disabled unless an administrator enables them; afterward, they become available by default unless the workspace opts out. Administrators choose which members have access, view personal dots and their connections, pause or restart them, and monitor usage; according to the help article, enabling them also requires compatible Agent Security policies. For compliance, dots support data residency and Enterprise Key Management, with inference hosted in the EU or the United States; dots created before general availability migrate to the required region within 30 days, and they are HIPAA-eligible in covered workspaces.

🔗 ChatGPT Enterprise and Edu release notes

Briefs

Agents and coding tools

Zed 1.27 and preview 1.28 — Stable version 1.27 incorporates the October 28 preview (a more resource-efficient Agent panel, removal of restore checkpoints, macOS 12 minimum); preview 1.28 follows the latest version of the Agent Client Protocol, truncates lines read by the agent to 2,000 characters, and fixes three context compaction issues. 🔗 source
Delta 0.24 — Changes from a subthread merge into the parent thread (/merge-into-parent, /pull-from-subthread), an agent reads a shared thread using its link alone and writes to the agent in another thread with delta thread send; MCP servers accept bearer tokens and custom headers. 🔗 source
Copilot CLI 1.0.106-7 to 1.0.106-9 — Three previews, still without a stable 1.0.106: a /code command opens the current session in VS Code, the Fusion level is configured in copilot config, and 79 fixes neutralize new escape sequences, among other issues. 🔗 source
Gemini CLI v0.67.0 and v0.68.0-preview.0 — The stable release promotes the previously covered 0.67.0 series, and the preview bundles the last eight days of nightlies; the only new addition is the November 4 nightly, which makes gemini hooks migrate recognize Claude Code’s SubagentStop event and adds --userns=keep-id under rootless Podman. 🔗 source
Qwen Code v0.26.7 — Hooks, permissions, and language servers supplied by a repository no longer run without explicit trust granted to the folder; qwen serve opens up to remote connections, and forced tool selection is fixed for Claude Opus 5.5, Sonnet 5.5, Fable 5.1, and Mythos 5.1. 🔗 source
Kimi Code 2.5.0 — A session can start in an SSH, Docker, or command-defined environment (--environment), and /model or /effort can be used while the agent works; several REST and WebSocket server endpoints are removed, although these changes are listed as fixes. 🔗 source
mistral CLI 0.17.0 — mistral agents init creates an agent project, and an application can embed an agent that mistral dev serves during development and mistral deploy publishes, behind the mistral_cli_agents flag; the mistral login --device connection is now rejected. 🔗 source
DeepSeek Harness 0.2.2-rc.1 — The first release candidate in the 0.2.2 series and the 37th prerelease without a stable release: it mostly incorporates the previously covered alphas, aligns the TypeScript and Python SDKs with Web mode, and moves Ralph loop and session search out of experimental status. 🔗 source
Zach Lloyd’s take on the software factory — On Warp’s blog, he argues that a software factory (software factory) is not built by chaining prompts and cloud agents, but as a stack: a factory defined in code, durable orchestration, evaluations and replay of past tasks, with API and MCP integrations first. 🔗 source

Platforms, APIs, and administration

Together AI and Kimi K3.1 Pro — Together AI serves Kimi K3.1 Pro on the day its weights are released, as is or post-trained with RL or SFT using Together Post-Training; no pricing in the announcement. 🔗 source
OpenAI SDKs — openai-python 3.45.0 and openai-node 7.50.0 support 12-hour response caching and per-project service tiers, announced in October; for Codex CLI, no stable release since 0.166.0 on October 29, only prereleases with empty notes. 🔗 source
Copilot usage metrics — The usage metrics API now separates local BYOK (bring your own key) usage, reported as a lower bound, and identifies custom enterprise models with the is_custom_enterprise_model indicator; existing totals remain unchanged. 🔗 source
grok-imagine-image-quality retired — Since November 2, SpaceXSI has served requests for this image model through grok-imagine-image-2.0 at low quality, without code changes, at $0.04 per 1K image and $0.06 per 2K image, $0.01 less than before. 🔗 source
Claude Marketplace — 14 partner agents and products join the Claude Marketplace, from Chainguard to Zoom, including Datadog, Netlify, Runway, and Glean, bringing the total to more than 25, payable using a portion of eligible companies’ Anthropic spending commitments. 🔗 source

Security

Warp’s GPG key — Warp is renewing the GPG subkey that signs some of its Linux packages as a precaution, following an unspecified security incident with no observed misuse; on Fedora 42 and earlier, RHEL, and openSUSE, the old key must first be removed. 🔗 source
GPT-5.6 Cyber in AWS GovCloud — Approved Daybreak Red customers can access GPT-5.6 Cyber in the AWS GovCloud (US-East) region through Amazon Bedrock, following approval by OpenAI and then activation by AWS; AWS had already published this on October 30. 🔗 source
Mistral Large 4 tackles a Django vulnerability — In a video sped up 90 times, 133 agents map the code, find and confirm a recent Django vulnerability, then Le Chonk fixes it; no CVE number, cost, or actual duration. 🔗 source
CodeQL 2.27.4 — Fewer false positives in Java and Kotlin, Rust, and C/C++, the CWE-730 tag removed from ReDoS queries in six languages, and files linked from outside a pack’s root are excluded when it is created; no AI component this time. 🔗 source
Perplexity’s MCP guide — Three categories and eleven risks in MCP systems, from indirect prompt injection to a compromised local server, along with their mitigations; no Perplexity product or figures. 🔗 source

Open models and research

Synthetic voice detection — Two researchers at the University of Melbourne release VeriSpoof, a 2.9-billion-parameter audio LLM that detects synthetic voices, attributes their source, and explains its verdict; according to the authors, its VeriSpoof-Plus variant achieves an average equal error rate (EER) of 1.79% on Speech-DF-Arena, with weights under the noncommercial CC BY-NC 4.0 license. 🔗 source
Dharma-AI’s Field Extractor — A 2B model that returns a document’s fields as typed JSON; on its in-house benchmark of 546 documents, 0.834 versus 0.816 for Claude Opus 5, compared against models that are not their developers’ latest releases, and with no weights found on the Hub. 🔗 source
Cribl’s OTAP — A metric that evaluates an agent on the structure of its plan, aligned with reference plans through optimal transport; according to the authors, it agrees more closely with human judges than ROUGE-L across 236 trajectories (0.74 versus 0.64). 🔗 source

Consumer agents

Grok Bot and shopping — On November 3, SpaceXSI’s agent offers a shopping experience: find the best deal, compare items, and track all orders; no countries, plans, or merchants specified. 🔗 source
Grok Bot on the phone — On November 4, the @bot account announces that Grok Bot can make phone calls, with a video demonstration, but no countries, plans, or pricing. 🔗 source
Get Maine Lobster and Computer — In a pilot offered by Perplexity, with a $30,000 test budget supplied by Perplexity, Computer’s Meta campaign reports a 63% higher return on ad spend (ROAS) and 3.5 times the revenue for 2.1 times the spending, compared with the client’s usual ads. 🔗 source
MotionSites in Computer — The skill by @viktoroddy lets Perplexity’s agent build animated WebGL websites, with scroll animations, using 40 reference plans and 52 hosted assets; available to everyone on the web. 🔗 source
Manus in 3D — Manus turns photos into a playable 3D version of the user, ready to enter worlds created with the agent, extending Game Dev; no plan or credit cost specified. 🔗 source

Image, video, and creators

HeyGen Video 1 on mobile — HeyGen’s video model arrives in the HeyGen app on iOS and Android, with no pricing, duration, or resolution details. 🔗 source
Pika H3 Flex — The model officially arrives on the Pika platform, two days after the announced date, with 15 free generations per day until November 15 for Creator and Fancy subscribers. 🔗 source
Luma’s Character Sheet — The day after Expression Sheet, a consistent multiview character sheet, created from references or a description, with eight layouts; no post or pricing. 🔗 source
HyperFrames — HeyGen’s family of video tools claims more than 700,000 monthly users, without explaining its calculation method, and opens a dedicated Discord for its community. 🔗 source

What it means

Open-weight models reach frontier scale, with conditions. Kimi K3.1 Pro, with 2.8T parameters and roughly 1.56 TB of files, is freely available to download, and Together AI serves it and offers post-training from day one. Its license nevertheless shows where openness ends: above $20 million in revenue, a reseller must negotiate with Moonshot AI, and the Mistral Large 4 license already required a commercial license for model-as-a-service use on October 29. Infrastructure is keeping pace: Together Compiler tasks agents with optimizing open models across widely differing architectures, claiming up to 2.6 times more tokens per GPU. Cognition, meanwhile, is keeping SWE-2 Flash’s weights private for now, citing cyber risk, and is changing its foundation: SWE-2 was based on Kimi K3, while SWE-2 Flash is based on an American model.

Coding agents are coming together and gaining tools. VS Code brings Copilot, Claude and Codex into the same session, Copilot joins Xcode’s native assistant, and OpenAI is set to acquire the Git tools already used by Codex, two days after introducing Pando: version control is becoming a strategic battleground as agents multiply branches. Cost remains the other front. SWE-2 Flash costs one-fifth as much as SWE-2, according to Cognition, and Anthropic assigns the subagent status line to Haiku 5.5 to fit roughly 6% more prompts within the same limit.

Evaluation moves beyond the test bench. RoboColosseum evaluates robotic policies on real robots in homes and restaurants, and the results temper the rankings: frontier models rival specialized VLAs, but no system is reliable in open-world settings, and none completes a single task on the G1 humanoid. Perplexity’s Decisions API starts from the same observation on the software side: a model that knows only its training data gets current events wrong, and web search takes pplx-decider-v1.2 from 51% to 86% on a public benchmark. In science, verification is becoming the deliverable: FVCOM in Rust matches the Fortran version to 12 decimal places across 100 simulations, and Anthropic publishes Claude’s errors alongside its results.

Watermarking is becoming a default setting dictated by regulation. SpaceXSI enables it on November 7 for seven Grok models under the AI Act, and only a declaration of a legal exemption allows users to opt out; Anthropic expanded its own on November 3, while OpenAI leaves its watermarking disabled by default. For developers, the difference is visible in the response header and in who bears responsibility: disabling watermarking means taking sole responsibility for meeting labeling obligations.