On — Nov 10

Anthropic releases Claude Fable 5.2 and Claude Mythos 5.2 for hard technical work, with gains in coding, reasoning, and vision. Fable 5.2 is in the Claude Platform and Claude Code, runs up to 30% faster than Fable 5.1, and costs $8/$40 per million input/output tokens, down from $10/$50.

Why it spiked. Builders are posting cost and time comparisons across effort levels to decide which setting is worth paying for.

• @claudeai Announces Fable 5.2 and Mythos 5.2 with improvements in coding, reasoning, and vision.
• @ClaudeDevs States availability in the Claude Platform and Claude Code, the 30% speedup, and the new $8/$40 pricing.
• @0xRhinoX Runs one prompt at four effort levels: $24.31 in 1h 05m at High, up to $642.42 in 10h 21m at Ultra.

Meta releases Muse Ember, its smallest and most efficient model, for agent subtasks such as tool calls, routing, extraction, and coding subtasks. It is on the Meta Model API and Muse Code. Artificial Analysis scores it 36 on its Intelligence Index, matching Claude Haiku 5.5 (max) at about a third of the cost per task.

Why it spiked. It undercuts Haiku 5.5 and GPT-6 Luna on cost per task at similar scores.

• @alexandr_wang Says Ember offers close to Muse Spark 1.3 (max) intelligence at a 17x lower price and is the first model trained using Muse Wonder Full.
• @ArtificialAnlys Reports a score of 36, matching Claude Haiku 5.5 (max) at a third of the cost per task.
• @vercel_dev Adds Muse Ember to AI Gateway as meta/muse-ember.

Theo Browne, who leads T3 Code, posts a chart of Codex versus Claude usage in the app over time and says Opus 5.5 now dominates. He also posts a setup that has Grok Bot hand coding tasks to T3 Code running Opus 5.5 on a Claude plan, to stay under Grok Bot usage limits.

Why it spiked. Developers are comparing which coding model they use day to day and how to spread work across subscriptions.

• @theo Shares the Codex vs Claude chart for T3 Code and calls Opus 5.5's lead unlike anything he has seen.
• @theo Lists four steps to route Grok Bot tasks to T3 Code with Opus 5.5 on a $100 Claude plan.
• @imvincz Reports running 12 Opus 5.5 subagents plus five T3 Code threads with little of the $200 plan's limit used.

Expo co-founder Charlie Cheever previews Exact, a pre-alpha, open-source app framework in Rust designed for AI agents to write. One codebase targets web, macOS, iOS, and Linux. UI is written in a small declarative language called Contract, and logic in Rust or TypeScript.

Why it spiked. It comes from the head of the main React Native tooling company, and it is built for agents rather than people.

• @ccheever Shares the preview and describes Exact as an agent-native Rust framework built for speed.
• @SadCreatorTalks Says Cheever designed it to be written by agents and checked design choices with Codex and Claude daily.
• @AndrejDrats Notes Exact is an alternative to Expo's own stack and Cheever has already said Expo is not going away.

Epic releases the Unreal Engine 5.9 MegaJam Preview, which it says is the first engine release built mainly with agentic workflows. It adds experimental features: expanded Unreal MCP toolsets, AI coding agents inside the editor, Unreal Canvas for generated assets, and a browser target.

Why it spiked. Game developers can now drive the editor with Claude Code, Codex, or Cursor ahead of the November 12–15 MegaJam.

• @UnrealEngine Announces the preview and its experimental agent and MCP features.
• @XclusiveInsight Says a small team directed agents for about a week, starting the browser target from UE4's retired HTML5 code.
• @alexchristou_ Shares a one-shot Call of Duty-style remake built with Fable 5.2 through the Unreal 5.9 MCP.

NVIDIA releases GR00T 2, an open world-action model for robot manipulation, at CoRL 2026 in Austin, alongside Isaac Lab 3.0 and new open benchmarks. The model predicts future video frames and robot actions together. A Turbo version runs on a humanoid with Jetson Thor.

Why it spiked. It is open and learns skills from human video that transfer to other robot bodies.

• @yukez Explains that a reasoning model plans through the WAM API while GR00T 2 executes, and agents can point to what to grasp.
• @yukez Shows GR00T 2 Turbo on Jetson Thor driving a humanoid's whole body to serve popcorn and clear tables.
• @sphinx_thesis Says it is built on Cosmos, trained on robot data plus 100,000 hours of first-person human video.