AgentdexModel Pulse #10 · Thursday, Nov 5, 2026
AI coding tools get bought up, Kimi K3.1 Pro goes open, and Unreal Engine 6 bets on Verse and AI
The seven AI stories since yesterday's issue that matter most for AI coders and indie game devs. Each one links to its primary source.
Sponsored Sponsor this issue: reach AI coders and solo founders in Model Pulse.
Nov 4, 2026
DEALS: OpenAI buys Pierre, Meta buys Conductor and Factory buys HumanLayer, all on the same day
Three AI coding acquisitions landed on Nov 4. OpenAI announced it will acquire Pierre Computer Company, the maker of Code Storage, a Git-based service that lets apps store code and its history, and of the open-source Trees and Diffs libraries already used to browse files and review changes in the Codex app. OpenAI says Pierre's open-source projects stay open source, Code Storage customers can keep using it while its features move into Codex, and the deal still needs regulatory approval. Terms were not disclosed (Axios). Meta bought Conductor, the Mac app that runs a team of coding agents in parallel workspaces; the team joins Meta Superintelligence Labs, terms undisclosed (TechCrunch, Axios), and promises "something new" early next year. If you use it: new sign-ups closed on Nov 4, existing Pro users keep Pro features for free, Conductor Cloud goes read-only on Nov 13 and starts deleting data on Nov 20, so right-click a cloud workspace and choose "Transfer to local" or export before then. The local app gets new models and critical fixes until early 2027. And Factory acquired HumanLayer, the multiplayer coding-agent IDE from Dex Horthy and Kyle Mistele; Factory will open-source HumanLayer and bring its shared tasks and in-app plan reviews into Factory. Also on Nov 4, Phaze raised an $8.75 million seed led by a16z to give people and AI agents remote access to powerful workstations, from the founders of Parsec. The signal: the tools that sit around coding agents (source control, orchestration, team review) are now worth buying outright.
Nov 4, 2026
MODELS: Cognition's SWE-2 Flash is a fast, cheap American coding model, free in Devin until Nov 30
Cognition introduced SWE-2 Flash on Nov 4, a coding model built on Thinking Machines' Inkling-Small (276 billion total parameters, 12 billion active). The new step in its recipe is a mid-training stage on a few trillion tokens of code, scientific articles and blogs before reinforcement learning, which Cognition says lifted Terminal-Bench 4.0 from 0.9% to 20.0% and FrontierCode 2.0 from 3.2% to 24.0% over the base model. It is available now in Devin Desktop and the CLI at one-fifth the price and three times the speed of its bigger sibling SWE-2, and free for paid Devin Pro, Max and Teams users through Nov 30 (RuntimeWire). Read the benchmarks honestly: Cognition's own table shows Flash edging SWE-2 on FrontierCode 2.0 (24.0 vs 23.8) but trailing it on Terminal-Bench 4.0 (20.0 vs 27.3) and DeepSWE 1.1 (60.1 vs 73.0), so it is a cost-and-speed option for routine work such as automations, code review and docs, not a replacement. Cognition says it beats every American open-weight model, but the weights are not public; it is building a program to share them with vetted partners and researchers, citing cyber risk.
Nov 4, 2026
OPEN WEIGHTS: Moonshot posts Kimi K3.1 Pro's 2.8-trillion-parameter weights, on the day it promised
Moonshot AI published the full weights of Kimi K3.1 Pro on Hugging Face on Nov 4, meeting the deadline it set at the Oct 23 launch. The model card lists a mixture-of-experts model with 2.8 trillion parameters (104 billion active per token), a 1-million-token context window and text-plus-image input, shipped pre-quantized to MXFP4. Moonshot's own numbers put it close to the top closed models on coding: 65.7 on Terminal-Bench 4.0 against 66.4 for Claude Opus 5.5 and 58.2 for GPT-6 Astra; these numbers are company-reported and not yet independently reproduced. It is still not a laptop model: the repository is about 1.56 TB, so self-hosting means a large GPU cluster, and vLLM, SGLang and TokenSpeed are the recommended engines. Per Neural Brew's read of the license, use, modification and commercial use are allowed, with extra terms only for very large API resellers and products, and Moonshot says it removed cybersecurity data from its training. For AI coders, the practical win is that third-party hosts can now serve it, which should push prices down.
Nov 3-4, 2026
GAME DEV: Unreal Engine 6 drops Blueprints and Actors for Verse, betting on AI-assisted development
At Unreal Fest Tokyo on Nov 3, Tim Sweeney said Unreal Engine 6 will remove Actors, Blueprints and the C++ Gameplay Framework from the start of Early Access, planned for around the end of 2027, rather than phasing them out later as Epic said in June (Game Makers keynote report, Famitsu). Gameplay will be written in Verse, the language used in Unreal Editor for Fortnite, on top of Scene Graph, where entities are built from components; C++ stays for rendering, physics and other low-level engine code. Sweeney tied the change to AI: he said AI tools have raised productivity inside Epic by 1.5x to 4x, and that UE6 will let developers connect any AI model through MCP and Python scripting, while Epic will not build or sell its own models. Epic plans AI-assisted tools to convert Blueprints and C++ gameplay code to Verse, and Sweeney apologized to developers who love Blueprints. In a Nov 4 post he said the goal is an engine "more powerful than UE5, easier to develop in than Unity," and in another that once UE6 Preview 1 ships, UE5 will get only critical security and platform-compatibility fixes. For indie game devs on Unreal, learning Verse now looks like the safe bet; UE5.9 Preview is due next week.
Nov 4, 2026
MODELS: Tencent's Hy4 adds vision and computer use, trained on Blender, CAD and game-dev work
Tencent released Hy4 on Nov 4, the full version of its flagship model after Hy4 Preview. For the first time it can understand images and charts and operate both command-line tools and desktop apps, with specialized training on professional software such as Blender and CAD (Tencent's announcement). For game devs, Tencent says it worked with developers from its games group to integrate Hy4 with Unreal Engine, Unity and Godot for complex game logic and multiplayer code, and reports gains over the preview of 19.1% on its GameDev benchmark and 9.8% on GameLogic. TechNode reports it is a 770-billion-parameter mixture-of-experts model. In Tencent's own table it scores 78.9 on OSWorld 2.1 against 81.8 for Claude Opus 5.5 (company-reported). It is available through Tencent Cloud TokenHub, OpenRouter and OpenCode and inside CodeBuddy and WorkBuddy, which are free from Nov 4 to Nov 15. API pricing stays at the preview's level of about $0.90 per million input tokens and $2.70 per million output tokens, with 50% off on TokenHub through Dec 31. One catch: the weights are not out yet; Tencent says weights and system cards come "in the coming days." A cheaper Hy4 Flash is due soon.
Nov 4, 2026
SECURITY: OpenAI's agent pipeline produced 948 Linux root exploits, and patches for them, for $1.2M
OpenAI published "A thousand fewer paths to root" on Nov 4. Its autonomous pipeline, running an upcoming cyber model in mid-October, used scout agents to pick about 7% of the Linux kernel to audit, produced 2,308 candidate findings, reproduced 1,724 of them in isolated virtual machines, and built 948 distinct local privilege escalation exploits against a stock desktop distribution, each paired with a candidate patch. The run cost $1,213,430 in API credits, about $1,279 per exploit. Human researchers review the results before reporting, and OpenAI says it has already sent hundreds of reports and patches for critical bugs to kernel maintainers. A July run with GPT-5.6 Sol found 166. Two lessons for anyone running agents: in an early run, 20 of 200 "successful" exploits got root through a different zero-day than the one assigned, so check how an agent reached a result, not just the result; and stronger models needed less scaffolding, so re-test your harness when you upgrade. OpenAI's practical advice: on an unpatched machine, a low-privilege account should be treated as equal to root, so keep systems updated and run agents in sandboxes such as Firecracker or gVisor.
Source: https://openai.com/index/a-thousand-fewer-paths-to-root/
Nov 4, 2026
TOOLS: Claude Code's background subagents now burn far less of your usage limit
Anthropic's ClaudeDevs account said on Nov 4 that background subagents in Claude Code now use less of your usage limit. A running subagent refreshes its status line every 30 seconds, and each refresh used to re-read the subagent's whole conversation. Now the smaller Haiku 5.5 model writes the status from short notes instead, which Anthropic says makes each update about 99% cheaper. If you run several long subagents in parallel, those status checks were a hidden cost on your plan, and you no longer need to cut back on background agents to save quota. In the same window, OpenAI brought GPT-Image-2.6 to Codex and the API, with two new API image models, Flare and Sunburst, that it says generate images 50% faster, and with Sunburst keeping details more consistent from one edit to the next, which is useful if you iterate on game sprites or UI art inside your coding agent.
Ship faster this week
Grab the free Agentdex starter kit with prompts and agent configs for Cursor, Claude Code and Codex.
Get the free kitWe store your email with Resend and use it only for Model Pulse. See privacy. Missed the last issue? Read Model Pulse #9.