Tuesday, November 10, 2026

Anthropic’s New Models Push Further

Anthropic’s New Models Push Further

Today’s Overview

Good morning, Anthropic says its latest models have tackled work ranging from software migrations to mathematical proofs, while Meta is betting on a smaller model built for the many repeatable steps agents take. Researchers are also finding new ways to make visual agents use what their tools reveal, rather than merely calling those tools. Let’s dive in.

Top Stories

Meta Debuts Muse Ember for Agent Work

Meta introduced Muse Ember, a compact reasoning model aimed at agent tasks such as tool use, coding, routing, and long-document parsing. Meta says it matches Claude Haiku 5.5’s top reasoning score at roughly one-third the cost and beats GPT-6 Luna across reasoning settings.

  • On AutomationBench-AA, Muse Ember scored 37.8%, compared with 20.2% for GPT-6 Luna and 13.9% for Claude Haiku 5.5.
  • The model’s context window reaches 1 million tokens, with up to 512,000 output tokens.
  • Meta reports a score of 27.2 on FrontierCode 2.0, below Haiku 5.5’s 34.9 but above GPT-6 Luna’s 25.2.

OpenAI Opens GPT-6 Cyber to Select Security Teams

OpenAI introduced GPT-6 Cyber for approved security customers, saying the model powered its recent Linux kernel research. Security teams at Mozilla, Palo Alto Networks, Calif, Cisco, and OpenAI are already using it, with access set to expand over time.

  • The model is currently available to select Daybreak Red customers.
  • OpenAI describes Daybreak Blue as a way for verified defenders to validate vulnerabilities and fixes with fewer refusals.
  • The company says it aims to expand access to organizations of every size, independent researchers, and open-source maintainers.

Anthropic Upgrades Fable and Mythos

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, models with improvements in coding, reasoning, and vision. They share the same underlying model, with stronger safeguards on Fable, and generate output up to 30% faster than their prior versions at lower prices.

  • On FrontierCode 2.0, Fable 5.2 scored 49.1%, up from 32.4% for version 5.1 and ahead of Opus 5.5’s 45.5%.
  • A team of 1,000 Fable 5.2 agents formalized Fermat’s Last Theorem in 30 hours, producing a proof seven times shorter than an earlier attempt.
  • In one cybersecurity test, Mythos 5.2 found 10.2% more vulnerabilities than Opus 5.5 and 14.4% more than Mythos 5.1.

Research & Analysis

Sage-Taiyi Uses Metacognition to Guide Scientific Discovery

Sage-Taiyi keeps a running account of evidence, assumptions, and unresolved questions, then uses that understanding to choose research targets and methods. Its authors report validation through five expert-guided collaborations across chemistry, materials science, and engineering.

  • The Taiyi-Bench evaluation drew on 265 tasks from 50 papers across ten disciplines.
  • Sage-Taiyi scored 54.0 overall on Taiyi-Bench, compared with 47.9 for AI Scientist-v2.
  • The reported chemistry results include allylic amination yields rising from 13% to 69% after reaction-condition optimization.

Reframe Trains Visual Agents to Use Their Observations

Reframe is a two-stage post-training framework designed to stop visual agents from ignoring tool-returned evidence and to repair deficient tool actions before distillation continues. It outperformed Mix-RL and standard multi-teacher on-policy distillation across image and video benchmarks, without requiring task-specific supervised fine-tuning first.

  • Its Reframe-100K training set contains 63,000 image and 37,000 video question-answer pairs.
  • The paper reports observation dependence of 75.53 and tool gain of +5.23 for Reframe in its comparison.
  • During training, the system drops the original visual input after a successful tool call with 50% probability, while keeping the tool-returned observations.

Recursive Rewriting Boosts a 27B Terminal Agent

Terminal-RSI combines recursive self-rewriting with policy refinement to improve an agent’s ability to complete terminal tasks. The paper reports a threefold increase in terminal task success rate for a 27B model.

  • Terminal-RSI-27B reached a 50.58% four-suite mean success rate, 29.90 percentage points above the base model.
  • On Long-Horizon-Terminal-Bench, the system achieved 5 strict completions out of 46.
  • The paper’s revised reward method reduced the share of reward going to failed trajectories from 50.09% to 28.72%.

Flide Speeds Up Decision-Model Inference

Flide runs the shared state and multiple decision questions together while preserving each question’s standalone computation. Its ForkAttention mechanism uses GPU-resident metadata to keep branches separate across both standard attention and recurrent layers.

  • For a request with a shared state of P tokens and multiple questions, Flide reduces token computation from nP plus the question tokens to P plus the question tokens.
  • An optional batch-invariant mode aims to keep answers bitwise identical across server loads.
  • The engine can resume expanding agent histories from checkpoints saved during earlier inference passes.

Trending AI Tools

  • OpenRouter Auto Router and Auto Code A general model router and a coding-focused option weigh request details alongside live pricing, provider health, and cache status.

  • Claude Personal Finance Connect bank, card, loan, and investment accounts to ask questions about up to 24 months of transactions.

  • Ace Collaborate with agents and teammates on local machines without exposing work to the public internet.

Quick Hits

Keep reading for free

Enter your email. If you're already subscribed, we'll send a sign-in code. If not, you'll subscribe in the next step.

Free access. Subscribe once, then use the same email on future issues.

Free to read. Subscription just unlocks the full issue.