Model Pulse #16 · Saturday, Nov 14, 2026
LangChain raises $300M at a $3B valuation, Antigravity agents start writing their own workflows, and llama.cpp hits 1.0
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 13, 2026
DEALS: LangChain raises $300M at a $3B valuation, led by CapitalG, to build out LangSmith for running and governing agents
LangChain announced a $300M Series C at a $3B valuation on Nov 13. Existing investor CapitalG, Alphabet's independent growth fund, led the round, with significant participation from new investors GIC and DFJ Growth and existing investors IVP and Sapphire Ventures. Salesforce Ventures, Insight Venture Partners, QuantumLight and MongoDB Ventures joined, and Sequoia, Amplify and Benchmark put in more. Co-founder Harrison Chase says more than 7,000 organizations now use LangSmith, the company's framework-agnostic platform to run, improve and govern agents, from startups like Harvey and Rippling to NVIDIA, Toyota and Salesforce. The company says Workday cut the time to deploy a new agent from five months to under a day with it. The money goes into three areas: agent runtime (LangGraph, Deep Agents, LangSmith Agent Server and sandboxes), agent improvement (observability, evals, and an in-platform agent called LangSmith Engine that proposes code and prompt fixes), and governance through model and tool gateways. LangChain engineer Jacob Lee posted that the team is growing past 450 people this year. That's quite a run for an open-source project started four years ago. For AI coders it means the open-source LangChain and LangGraph libraries have a well-funded owner. New LangSmith features are due at a release event on Friday, Nov 20.
Nov 13, 2026
CODING TOOLS: Google Antigravity's agent now writes Python scripts to run teams of subagents, a day after Gemini 4 Hydrogen arrived as its fast coder
Google Antigravity began rolling out Dynamic Workflows on Nov 13. Type /run-workflow plus a task, and the agent writes a short Python script that starts subagents in parallel or one after another, each in its own conversation, and returns a structured result checked against a JSON Schema. Per the docs, you approve the script before it runs, can keep working while it runs in the background, and can save a script that works and run it again later. Good fits: one review per changed file, one audit per package, or write a failing test and then fix the code. File reads and test logs stay in the subagents' chats, so your main context stays clean. It is on Google AI Pro and Ultra plans in the Antigravity app and CLI, and Google says the older markdown workflows are deprecated. The model underneath is new too: Google released Gemini 4 Hydrogen on Nov 12, a fast model for everyday coding and agent work, priced flat at $1.50 per million input tokens and $7.50 per million output tokens across its 1M-token context. Google says it scores 51 on the Artificial Analysis Intelligence Index, edging its bigger Gemini 4 Argon (50), and 57% on Terminal-Bench 4.1. Google's own model card puts it behind Claude Sonnet 5.5 on that test (63.9%). A new Orchestrator mode has Gemini 4 Carbon or Argon plan while Hydrogen writes code and runs tests. Google says this comes close to Carbon's success rate while using 24% less quota and finishing 19% faster (company figures).
Source: https://x.com/antigravity/status/2121327669534933025
Nov 13, 2026
AGENTS: DeepSeek puts its app's web search, file reading, URL fetching and voice into the API as switches on its low-cost Flash model
DeepSeek opened the tools behind the DeepSeek App to developers on Nov 13. The deepseek-flash model (V4.1-Flash) now runs web search with citations, search during multi-step reasoning, PDF and Office file parsing, and automatic reading of links in user messages on DeepSeek's servers. You turn them on with parameters such as deepseek_web_search and deepseek_read_url, and they work in the Messages, Responses and Chat Completions formats. Search, fetch, OCR, speech-to-text and text-to-speech are also sold as standalone APIs. DeepSeek also published an open-source chat front end you can fork. Tools are billed per use on top of tokens. AI Tamer read the pricing page as $0.75 per 1,000 searches or fetches and $0.225 per 1,000 pages for file parsing and OCR, with Flash tokens from $0.15 per million input and $0.60 per million output off-peak. DeepSeek requires public-facing apps to turn on its safety check, and says the developer stays responsible for moderation and for any registration and AI-labeling rules that apply. For AI coders building agents or chat features in games and tools, this skips the slowest part of the job, the search and file plumbing. Weigh it against your data-handling requirements first, since requests go through DeepSeek's own servers.
Nov 13, 2026
OPEN SOURCE: llama.cpp hits v1.0.0 after 3.5 years, with semantic versioning, FP8 KV caches and a MoE expert cache
The ggml-org team at Hugging Face released llama.cpp v1.0.0 and ggml v1.0.0 on Nov 13, 3.5 years after Georgi Gerganov started it as a MacBook experiment. The main change is a promise: from now on, a breaking change to the public C API means a major version bump, new features and model support are minor versions, and fixes are patches. That makes it safer to build an app or game on libllama. Version 1.0 is itself a breaking release: the release notes list 35 deprecated functions removed from llama.h, a renamed function and a dropped --defrag-thold flag, so code written for 0.x needs one migration. New features include FP8 KV caches, a host-memory cache for mixture-of-experts layers that lets big MoE models run with part of the work on the CPU, and new architectures including Kimi-K3.1-Pro and Ripple-65B. llama-server gets security hardening, an OpenAI-compatible /v1/audio/speech endpoint and a redesigned web UI. The team says it supports more than 150 architecture variants. AI Scoop News points out that the versioning rule covers the C headers only, not llama-server's HTTP endpoints or command-line flags.
Nov 13, 2026
CODING TOOLS: YC startup Reason Machines launches a bring-your-own-keys agent harness that it says cuts the cost per coding task
Reason Machines launched its Reason harness on Launch YC on Nov 13. It runs coding agents on cloud Linux or Mac machines that keep their state as you move between devices, runs hundreds of sessions in parallel, and works from the web, a Mac app, a CLI, Slack or a REST API. You plug in your own API keys, open-weight models or subscriptions such as Grok, Copilot or Kimi, and work moves to the next account when one hits its limit. The two-person Fall 2026 team says the harness, not the model, sets most of the cost. It claims 30-80% lower cost per task than Codex, Cursor or Devin. Its own benchmark write-up says that on the same model it scored 65.1 against Codex's 62.5 on the Coding Agent Index, at 22% lower cost per solved task. Treat the headline numbers with care: CellCog's review of the launch found that the "up to 20x lower cost" in the co-founder's X post compares different models on each side, and that Artificial Analysis has not reviewed or endorsed the results (company claims, unconfirmed). Plans are $50 Pro and $200 Max a month, with a 3-day trial. The company says it has 20,000 users four weeks after launching and is raising a seed round.
Source: https://www.ycombinator.com/launches/XUY-reason-the-most-efficient-agent-harness
Nov 13, 2026
AGENTS: Deepnote's open benchmark drops 12 models into a fake company's Slack and finds the hard part is knowing when to stay quiet
Deepnote Research published the Agent Initiative Benchmark on Nov 13. It tests whether an always-on agent sitting in a company Slack helps when it should, reaches the right person, and stays silent when the matter is already handled. Twelve models faced the same 50 situations inside PageMantle, a synthetic software company with six years of records across 11 simulated services, including Slack, Notion, Linear, HubSpot, Stripe, BigQuery and Snowflake. Every model got the same tools and a 180-second budget. OpenAI's GPT-6 Astra led at 8.97 out of 10, with GPT-6.1 Sol at 8.65 for $0.11 per task, ahead of Claude Fable 5.2 at 8.17 for $1.87. Claude Opus 5.5 was the fast "sniper," answering in about 29 seconds. Gemini 4 Argon (5.60) and Grok 4.7 (4.51) made the most tool calls and often ran out of time before replying. The pattern across models: they sent help 9 times out of 10 when it was needed, but held back only 6 times out of 10 when silence was right. Scores come from two LLM judges on a 50-task set at one effort level, so Deepnote says to read them as tendencies. The code is on GitHub under Apache 2.0 and the PageMantle dataset is on Hugging Face, so you can test your own agent against it.
Nov 13, 2026
GAME DEV: Melty Gym turns games like GTA V into training grounds for robots, and Pearl Abyss expands AI across its whole pipeline
Melty founder David Miller introduced Melty Gym on Nov 13 with the line "Train robots in any game at massive scale." The demo shows simulated humanoid robots gathering and driving in GTA V, then working at server racks in a custom data center. According to Melty's Launch YC page, researchers can drop their own models and robots into custom games and change the mechanics, objectives and team dynamics to build reinforcement-learning and evaluation tasks. Melty is building these tasks with teams now, on request, so it's not self-serve yet. It's the same engine behind Melty's mod tools for players, which include 3,500 game mashups played by 170,000 people (company figures), 256-player multiplayer for mods, and an MCP server that lets a coding agent build a custom game from one prompt. In studio news, GamesRadar reported on Nov 13 that Crimson Desert maker Pearl Abyss said in its Q3 earnings letter that it is expanding AI "throughout our entire development pipeline," covering repetitive tasks, content support and test optimization. The studio says core gameplay and content stay human-made, including for its creature-collecting game DokeV, now targeted for late 2028. That lines up with GDC's November 2026 survey of more than 2,600 game workers, which found 40% now use generative AI at work, up from 36% last year.
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 #15.