Skip to content
Agentdex

Model Pulse #18 · Wednesday, Nov 18, 2026

OpenAI ships GPT-6.2 Astra and Sol, Cognition launches Devin Bots, Microsoft releases MAI Code 2, and Genesis AI raises $500M

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 16-17, 2026

DEALS: Genesis AI raises a $500M Series A at a $3.5B valuation to put its dexterous Eno robot to work, and Stanford's Noeri launches with $110M

Genesis AI announced a $500M Series A at a $3.5B post-money valuation on Nov 17. Premji Invest, Khosla Ventures and Eclipse co-led the round. NVIDIA, AMD Ventures, Bpifrance, Samsung Next, Ericsson, Zoom, T. Rowe Price-advised funds and Durable Capital joined, along with former Google CEO Eric Schmidt, Zoom CEO Eric Yuan, Jerry Yang and Xavier Niel. The Paris and San Carlos, California company builds its AI, its hardware and its training systems together. Its flagship robot, Eno, rides on a wheeled base, has human-proportioned hands and has no conventional head. CEO Zhou Xian says the money goes to more data collection and training, harder hand manipulation, hiring, and deploying Eno with customers. Manufacturing, logistics and labs come first, then hospitality and healthcare. TechCrunch reports that the training data comes from sensor gloves and headset cameras worn by more than 100 real workers, and that the robot wears the same gloves. The company says it has built 30 to 40 Enos and picked a contract manufacturer to start mass production in the US next year. Crunchbase News puts total disclosed funding at $605M since 2024. Also in the last 48 hours, Noeri came out of stealth on Nov 16 with $110M led by Menlo Ventures. Madrona, Kindred Ventures, Wing, NEA, Cisco Investments, NVIDIA and Snowflake took part. Noeri is run by Stanford professors Yejin Choi and Carlos Guestrin as co-CEOs. It builds a self-improving harness that uses a deployed agent's own logs, tool calls and test results to keep tuning that system, which the customer owns. Madrona notes that coding is where this kind of in-production learning took hold first, because tests show when an agent got it right.

Source: https://www.prnewswire.com/news-releases/genesis-ai-raises-500m-series-a-at-3-5b-valuation-to-master-human-level-dexterity-in-general-purpose-robots-and-deploy-at-scale-302941249.html

Nov 17, 2026

MODELS: OpenAI ships GPT-6.2 Astra and Sol to ChatGPT Work, Codex and the API, with Astra's token prices cut 20%

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on Nov 17. Astra is the frontier model for the hardest projects. Sol is the cheaper everyday model. Both are live for Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and in the API as gpt-6.2-astra and gpt-6.2-sol, with Ultrafast mode available at launch. Astra costs $8 per million input tokens, $0.40 cached and $40 output, which OpenAI says is 20% below GPT-6 Astra, with a 60% cut on cached input. Sol keeps GPT-6.1 Sol's rates of $2, $0.10 and $10. OpenAI says Astra writes cleaner code and builds better frontends, "particularly for websites, gaming, and 3D design." It also says Astra scores 5.5 points above Claude Fable 5.2 on Agents' Last Exam at 5% of the cost, and that on Cognition's FrontierCode 2.0, Sol lands within 0.5 points of Astra at about a fifth of the cost (company benchmarks). In Codex, OpenAI's release notes suggest Astra for tracing a bug across services and Sol for reviewing pull requests. Pick either with /model in the CLI after updating. Two things to know before you switch. The API changelog says GPT-6.2 changes how the Responses API keeps reasoning when you edit a conversation's prefix, and how it handles the top-level instructions field, so read the migration guide first. Tool calling needs the Responses API. OpenAI also launched both models with tighter safeguards on higher-risk cybersecurity requests, and says this will block some legitimate security testing.

Source: https://openai.com/index/introducing-gpt-6-2-astra-sol/

Nov 17, 2026

AGENTS: Cognition launches Devin Bots, persistent agents that own a goal and run teams of Devins, and says one built its Android app in 44 days

Cognition introduced Devin Bots on Nov 17. A Bot is a persistent agent with its own memory and tools. You give it a goal rather than a ticket, such as "keep Android at parity with iOS" or "manage our releases." It investigates, starts Devin sessions in parallel, checks the results and decides what to do next. It brings you only the decisions that need a human. Cognition's main example: an engineer set up an "Android Bot" on Sept 28, and 44 days later Devin had an Android app, now in open testing on Google Play. The company reports 1,134 merged pull requests, 344,651 lines of Kotlin and 2,841 screenshot tests. The engineer sent 183 messages over that time and none at all on 22 of the days (company figures). Fortune reports that a Bot joins Slack as a team member, can be reached over iMessage, and runs on separate cloud computers, away from keys stored on engineers' laptops. Admins choose which repos and tools each Bot can touch, and network controls aim to stop data leaving. Teams decide which changes need human review. CEO Scott Wu said the safeguards rely on rules companies set "rather than relying on models themselves being smart enough to recognize a risky or malicious prompt." Cognition says usage grew tenfold in two weeks across early customers including Decagon, Fireworks AI and Reducto. Bots are available at devin.ai/bots. Cognition has not published pricing or said which plans include Bots.

Source: https://devin.ai/blog/introducing-devin-bots

Nov 17, 2026

CODING TOOLS: Microsoft opens Ignite with MAI Code 2 and Code 2 Flash in GitHub Copilot, and Copilot Fusion posts 71.2% on Terminal-Bench 4.1

At Ignite on Nov 17, Microsoft AI released MAI Code 2 and MAI Code 2 Flash, two coding models built for everyday enterprise work rather than the hardest problems. MAI Code 2 targets multi-file changes, deep debugging, repo-wide reasoning and UI. Flash targets high-volume work such as explanations, targeted edits and routine fixes. Microsoft says MAI Code 2 delivers 98% of GPT-6.1 Sol's performance at 40% of the cost, and 96% of Claude Opus 5.5's at 5% of the cost. Those figures are averaged over DeepSWE 1.1, Terminal-Bench 2.1, SWE-Atlas and FrontierCode 1.1, use a 50% promotional price, and were mostly run on Microsoft's own eval platform (company figures). In blind tests covering interfaces, games and 3D scenes, reviewers preferred MAI Code 2 over GPT-6 Sol 64% of the time. Microsoft says plainly that the models are "not yet tuned" for very long-horizon agent work. Both are in GitHub Copilot now. Flash is in limited preview in Microsoft Foundry, with OpenRouter, Vercel, OpenCode and Kilo listed as coming soon. The same day, GitHub said Copilot Fusion, its router that picks and coordinates models for each task, is now listed on the public Terminal-Bench 4.1 leaderboard. GitHub says Fusion Max scored 71.2%, 5.4 points ahead of Opus 5.5 and Fable 5.2, at 88% less cost than Fable 5.2. Fusion now works on github.com, in Visual Studio and in the GitHub Mobile beta apps. GitHub also added Claude Sonnet 5.5 to Fusion Max, GPT-6.1 Terra to Fusion High and MAI-Code-2 to Fusion Standard. GitHub does not explain how it measured cost.

Source: https://microsoft.ai/news/introducing-mai-code-2-and-mai-code-2-flash/

Nov 17, 2026

OPEN WEIGHTS: A new Hugging Face harness adapter, a small LoRA, more than doubles Qwen3.8-27B's hard SWE-bench Pro score inside the pi coding agent

Hugging Face researchers released a LoRA adapter on Nov 17 that tunes Alibaba's open Qwen3.8-27B to work as the model behind pi, the open-source coding agent. It was trained with reinforcement learning inside pi itself. On the 51 hardest SWE-bench Pro tasks, the pass rate rose from 29.4% to 62.7% with 55% fewer tool calls. On Terminal-Bench 2.1 it went from 39.5% to 55.8% with 62% fewer tool calls. The team published intervals and sign tests and checked the score on 68 tasks it did not use to pick the checkpoint (authors' results, not independently verified). The model card gives full setups for vLLM, transformers serve, llama.cpp (a GGUF version is available) and Llama.app on macOS, plus how to merge the adapter into the base weights. It notes the GGUF version has not been re-evaluated, and that thinking must stay off because the adapter was trained without it. Hugging Face's Younes (@yb2698) said more adapters for other harnesses will follow in the coming days. For AI coders running local models, the bigger point is that a small adapter trained for one specific agent can beat swapping to a bigger model, and the recipe is open: the training script is examples/async_grpo_harbor in TRL.

Source: https://huggingface.co/harness-hub/Qwen3.8-27B-pi-adapted

Nov 17, 2026

CODING TOOLS: Google AI Studio now builds full-stack Kotlin Android apps with Firestore, Google Sign-In and secure Gemini calls, then ships them to your own Play library

Google said on Nov 17 that the Google AI Studio Build agent, running Gemini 4 Carbon and Hydrogen, can now go from one prompt to a full-stack Kotlin and Jetpack Compose Android app. When your prompt needs saved data, sign-in or AI features, the agent asks first. Then it sets up a Cloud Firestore database with security rules, Firebase Authentication with Google Sign-In, and Firebase AI Logic with App Check, so Gemini API keys never sit inside the APK. The in-browser emulator now boots faster, streams in 1200p HD and plays live audio, so you can test sound-driven apps and games. A one-click screenshot button sends the emulator screen to the agent for visual fixes. Under the hood, the agent now uses the official Android CLI for scaffolding and Gradle builds. When you're done, AI Studio builds and signs the app bundle and publishes it to your personal Google Play library, so you can install it on your phone without USB debugging or a full developer console account. Third-party coverage reports that personal-library publishing is US-only and that Google Play lists the personal library as a beta. Google gives no figures for build success rates.

Source: https://developers.googleblog.com/build-full-stack-android-apps-with-gemini-in-google-ai-studio/

Nov 17, 2026

GAME DEV: HolyMolly launches a prompt-to-playable-3D-game tool that handles characters, rigging, animation and physics, and lets you keep reshaping it by chat

HolyMolly launched on Nov 17, pitched as "make anything a playable 3D world." You drop in an idea or an image and get a 3D game, then keep chatting to change characters, physics and animations, and post or remix the result. The company says it handles characters, rigging, animation, sound and visual effects. There is a mobile app on Google Play, and creators can join a waitlist for early access to HolyMolly Studio on desktop at holymolly.ai. The company's claim of a "studio-quality 3D game on the first try" is unverified. The examples it shared, such as a supermarket FPS a user built "in under 2 hours" and a basketball game a tester said took one prompt and about 25 minutes, are company-picked showcases, not independent tests. The launch post passed 300,000 views on X in its first day. HolyMolly has not published pricing, technical details or export options for engines like Unity, Unreal or Godot. For indie game devs it is worth an hour for fast prototypes and pitch demos, but don't plan a production pipeline around it until exports and asset licensing are clear.

Source: https://x.com/HolyMolly_AI/status/2122604857127325911

Ship faster this week

Grab the free Agentdex starter kit with prompts and agent configs for Cursor, Claude Code and Codex.

Get the free kit

Model Pulse

Daily AI model drops, tools, and what they mean for AI coders.

We store your email with Resend and use it only for Model Pulse. See privacy. Missed the last issue? Read Model Pulse #17.