Anthropic and Meta released new models on Monday that point to the same way of building AI agents: let an expensive model plan and judge, and hand repetitive steps to a cheaper one. Anthropic launched Claude Fable 5.2, which it says can act as an advisor or orchestrator for its smaller models. Meta launched Muse Ember, a low-cost model it says is built for "the work agents repeat all day."
"Muse Wonder plans, Muse Ember handles the volume," Meta wrote in its announcement, referring to its frontier model. Anthropic's post describes an Advisor role, where smaller Claude models consult Fable 5.2 on harder problems, and an Orchestrator role, where Fable 5.2 delegates simpler tasks to smaller models.
Why agents split the work
Agents spend most of their tokens on many small, structured steps, according to Meta, such as tool calls, coding subtasks, routing, extraction and long document parsing. TechCrunch described how that could look in a customer-support agent: a heavier model sets the rules for handling an issue, then a lighter model applies them across tens of thousands of cases.
The price gap explains the appeal. Muse Ember costs $0.10 per million input tokens and $0.50 per million output tokens, according to Meta. Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, according to Anthropic, which CNBC reported is 20% less than the 5.1 generation.
Meta said Ember scored 36 on the Artificial Analysis Intelligence Index at about $0.14 per task. In Meta's chart, Anthropic's Claude Haiku 5.5 reached a similar score at its highest setting for about $0.41 per task.
What each company is claiming
Meta's benchmark table shows Ember ahead of Haiku 5.5 and OpenAI's GPT-6 Luna on AutomationBench-AA, a test of end-to-end business workflows, at 37.8% against 13.9% and 20.2%. Ember trailed Haiku on FrontierCode 2.0-AA, which judges whether code changes would be merged, at 27.2% against 34.9%. Cline founder Saoud Rizwan said in Meta's post that Ember "is purpose-built for, and excels at, subagent work like coding and document analysis."
Anthropic's pitch for Fable 5.2 centers on running large teams. The company said a team of 1,000 Fable 5.2 agents fully formalized the proof of Fermat's Last Theorem in 30 hours with minimal human intervention. In a separate experiment, Anthropic said, the model coordinated more than 1,000 subagents over eight days to build a prototype operating system that boots on a laptop.
Anthropic also said Jump Trading tested Fable 5.2 in a harness similar to the one the firm uses for its own research. "In one run, Fable 5.2 effectively played all the necessary roles," Anthropic wrote, where earlier models had needed people to rebalance roles by hand.
What builders should check
The benchmark figures come from the companies, and several of Meta's are internal evaluations. Meta's footnotes say its Terminal-Bench 4.0 comparison ran each model in its best published harness, with Ember in mini-swe-agent and Haiku in Claude Code, so the results reflect the harness as well as the model.
Anthropic does not position Fable 5.2 as a default. It recommends most users start with Claude Opus 5.5 and move up to Fable or down to Sonnet or Haiku as the work requires. Fable 5.2 is available on all platforms, according to Anthropic, while Claude Mythos 5.2, the same underlying model with fewer cybersecurity and biology safeguards, is limited to organizations in its verification programs.
Ember is available through the Meta Model API and Muse Code, Meta said. Meta has kept the weights of its Muse models private, The Next Web reported. The company said it has more efficient models in training.
Source: Anthropic, Anthropic, “Introducing Claude Fable 5.2 and Claude Mythos 5.2”
Additional sources: Meta, Meta Superintelligence Labs, “Introducing Our Smallest, Most Efficient Model: Muse Ember”; TechCrunch, Russell Brandom, “Meta releases Muse Ember, a new lightweight model for agent execution”; The Next Web, “Meta's Muse Ember matches Claude Haiku at a third of the cost per task”; CNBC, Ashley Capoot, “Anthropic releases Claude Fable 5.2 and Claude Mythos 5.2 ahead of IPO”
