Anthropic's Claude Fable 5.2 lands, Epic ships an Unreal preview built mostly by agents, and Meta prices an agent workhorse at $0.10

Nov 9, 2026

MODELS: Anthropic ships Claude Fable 5.2, 30% faster and 20% cheaper, with Mythos 5.2 for vetted orgs

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on Nov 9. They are the same underlying model: Fable carries cybersecurity and biology safeguards and is open to all paid users and developers, while Mythos is limited to organizations in Anthropic's Cyber and Life Sciences Verification Programs. Anthropic says both generate output up to 30% faster than the 5.1 generation and cost $8 per million input tokens and $40 per million output tokens (CNBC puts that at 20% below 5.1). Per the model docs, you get a 1M-token context, 128K max output and the model ID claude-fable-5-2 on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry. An optional fast mode runs up to 2.5x quicker at double the price. In Anthropic's prerelease testing, the company says the model finished software migrations its engineers had considered infeasible and formalized the proof of Fermat's Last Theorem in 30 hours. Those are company claims. For game devs, an early outside test is worth a look: Wharton's Ethan Mollick says he had Fable 5.2 run 25,000 agents that built more than 450 portal-linked 3D worlds in code, and the demo and repo are public. At $40 per million output tokens, long agent runs add up quickly, so set a budget before you make it your default.

Source: https://www.anthropic.com/claude-fable-and-mythos-5-2

Nov 9, 2026

GAME DEV: Epic's Unreal Engine 5.9 MegaJam Preview was built mostly by agents, and it puts your coding agent inside the editor

Epic posted the Unreal Engine 5.9 MegaJam Preview on Nov 9, calling it "the first Unreal Engine release built primarily with agentic workflows." According to Epic, a handful of people directed fleets of agents for about a week to port AI tooling from the UE6 branch and to bring back a web target, starting from UE4's retired HTML5 code. Engineers directed, reviewed and tested the work. The practical changes for indie game devs: Unreal MCP now launches with the editor and covers rendering, source control, profiling, debugging, gameplay testing, audio, UI and level creation. A new Fab toolset lets agents search the catalog and install assets. The experimental Unreal Assistant panel runs Claude Code, Codex, Cursor or OpenCode right inside the editor. Python APIs now reach audio, MetaSounds, Niagara, materials, PCG and networking. Unreal Canvas generates images, meshes and video using your own model API keys. The experimental WebAssembly plus WebGPU target runs projects in a browser, but for now it needs a source build and a C++ project, so Blueprint-only projects can't use it yet. Epic says QA was shorter than usual, so test on a copy of your project. The preview is timed for the Epic AI MegaJam, Nov 12-15.

Source: https://forums.unrealengine.com/t/unreal-engine-5-9-megajam-preview/2852840

Nov 9, 2026

MODELS: Meta's Muse Ember is a $0.10 workhorse for the small steps agents repeat all day

Meta released Muse Ember on Nov 9, its smallest and most efficient model, built for tool calls, coding subtasks, routing, extraction and long-document parsing. It costs $0.10 per million input tokens and $0.50 per million output tokens, has a 1M-token context and up to 512K output tokens, and takes text, image, audio and video input. You can use it now as muse-ember-1.0 on the Meta Model API and in Muse Code. Meta says it matches Claude Haiku 5.5's best Artificial Analysis Intelligence Index score at about a third of the cost per task (about $0.14 vs $0.41) and beats every GPT-6 Luna setting. Those are Meta's numbers, and The Next Web notes Ember still trails Haiku on one coding benchmark. Meta pitches it as the execution layer under its bigger Muse Wonder model: "Muse Wonder plans, Muse Ember handles the volume," as TechCrunch quotes it. For AI coders running subagents, this is a cheap candidate to swap into your high-volume steps. Try it on your own eval set first.

Source: https://research.meta.ai/blog/muse-ember

Nov 9, 2026

DEALS: ClickHouse buys Xata to give agents instant Postgres branches; Nominal triples to $3B and Mavria hits $1B

Builders kept getting funded and bought on Nov 9. ClickHouse announced it is acquiring Xata and made ClickHouse Managed Postgres generally available on AWS. Xata's copy-on-write storage adds instant branches and clones, automatic idling and scale-to-zero, so developers and AI coding agents can build and test against isolated copies of production data. ClickHouse also plans a dedicated tier for agent sandboxing and agent memory. Price terms were not disclosed. Nominal raised a $240 million Series C at a $3 billion valuation, led by Lightspeed Venture Partners and co-led by Insight Partners. TechCrunch notes that triples its valuation in eight months. Nominal's AI platform serves hardware engineering teams at more than 80 organizations, and it is building agents for the full machine lifecycle. Mavria came out of stealth with $130 million raised across pre-seed and seed, at a $1 billion valuation in a round led by Paradigm. Its factory-built Terrain data center platform is already in production in Houston, and the founders, who previously worked at the Department of Energy, are betting that American AI needs infrastructure built like a product rather than a custom construction job.

Source: https://clickhouse.com/blog/managed-postgres-ga-xata-acquisition

Nov 7, 2026

TOOLS: OpenCode Browse lets your agent drive your real browser, logins included, and hand the tab back for 2FA

OpenCode shipped Browse on Nov 7. It is an MCP server plus a Chromium extension that lets agents open and drive tabs in your own browser through Playwright, using your existing logins and extensions (product page). Agent tabs open in the background in collapsed groups. Agents can take screenshots, record the screen and capture network traffic as a HAR file. When a login, 2FA prompt or payment comes up, the agent hands the tab back to you and carries on after you click Continue. It works with Chrome, Edge, Brave, Arc, Vivaldi, Opera and Chromium, and with any MCP client, including OpenCode, Claude Code, Cursor, Codex and Zed. Per-session allowlists limit which sites an agent can touch. OpenCode's dax notes on X that the agent and the browsers can run on different machines. For AI coders testing web apps, this replaces brittle headless-login scripts. Use a separate browser profile and tight allowlists, because the agent acts with your real sessions.

Source: https://opencode.ai/product/browse

Nov 9, 2026

ROBOTICS: NVIDIA's open GR00T 2 learns from human video and exposes an API agents can call

NVIDIA introduced GR00T 2 at the Conference on Robot Learning on Nov 9. It is an open world action model built on Cosmos that jointly predicts future video and robot actions. NVIDIA reports it completed 40 of 50 evaluated tasks at least once with no task-specific demos, averaged 71% success with one demo per task and 95% with five, and transfers skills between different robot bodies. For AI coders, the notable part is agent control: agents can call GR00T 2's API to run skills in sequence and recover from errors. NVIDIA also shipped Isaac Lab 3.0, with up to 2.5x training throughput using the Newton physics engine, and agents can now set up training environments for you. The research page says weights and training code are planned for December, so treat this as a preview until they land. If you build sims or physics-heavy games, Isaac Lab and Newton are the open pieces you can use today.

Source: https://blogs.nvidia.com/blog/robotics-simulation-advancements-corl/

Nov 7, 2026

FREE TIER: Qoder makes its mystery coding model Qoduck free each day for individual users

Qoder started a limited-time promotion on Nov 7: every individual Qoder International user, including Free accounts, gets a daily allowance of Qoduck calls that use no credits, even at the highest thinking setting. Qoduck is a "stealth" model whose identity Qoder has not disclosed. Qoder is Alibaba's agentic coding platform, but which model sits behind Qoduck is unconfirmed, and the X speculation about it is just that. After the daily allowance runs out, Qoduck bills at 0.5x credits. Teams and Enterprise accounts are excluded, and for now the promotion covers only Qoder International, not the Qoder IDE, JetBrains plugin or CLI. No end date has been set. For AI coders on a budget, it is a free daily slot to test a new model on real tasks. Keep sensitive code out of it until the provider says what the model is.

Source: http://docs.qoder.com/events/qoduck-free

Ship faster this week

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