Claude Fable/Mythos 5.1: The SOTA Model That Changes Agent Economics

Anthropic just shipped new frontier models that dominate every benchmark while cutting cache prices 75% — but at a cost. This is the week agent economics reshuffled.

The Launch

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 as its new flagship models for coding and knowledge work, directly via @claudeai. The announcement drew over 12 million views and immediately dominated every AI news channel. The benchmark table speaks for itself: Fable 5.1 leads across every evaluation, including 52.6% on Terminal-Bench-Science 0.1 and 55.8% on Terminal-Bench 4.0 — the coding agent benchmarks that matter most right now.

The Pricing Twist

List pricing stayed the same: $10/$50/$12.5 per million tokens for input/output/cache write. But Anthropic cut cache read prices by 75% to $0.25/MTok — fantastic for long-running sessions and large context windows. The catch? Observations from Artificial Analysis show 1.7x more output tokens per task, resulting in a net 20% per-task cost increase. The cache discount is real, but the output tax is real too.

This matters enormously for anyone building agent systems with promptengines: the math of your agent infrastructure just shifted. If your agents generate verbose output (which good agent harnesses do), your costs go up despite the headline price cuts.

What Was Crossed

Beyond the raw numbers, Fable 5.1 is positioned explicitly around "complex, multi-step work that runs on its own" — Anthropic's product and engineering voices framed it specifically for autonomous, multi-step reasoning: coding, knowledge work, and long-running problem solving. This is the agent-native positioning.

A key community finding emerged: Fable and Mythos 5.1 may share the same underlying weights, with different safety and routing behavior, not actually different base models. If true, this is a significant design choice — one model, two personalities, with the "knowledge work" variant (Mythos) potentially having different guardrails than the "coding" variant (Fable).

The Wave

This launch didn't happen in isolation. It's part of a round-robin of frontier model releases: Grok 4.7 and Gemini Flash 3.8 are also on the way. And alongside the model launches, World Labs Astra — Fei Fei Li and Justin Johnson's world model — launched what the Latent.Space newsletter calls "by far the most impressive world model launch we've ever seen." The convergence of frontier coding models and generative world models in the same week is not a coincidence; it's the AI industry showing its full hand.

What It Means for PromptEngines Projects

• Agent harness benchmarks just moved. Fable 5.1's Terminal-Bench dominance means existing agent harness benchmarks need recalibration. If you're evaluating agent performance, the baseline just changed.
• Cache economics flipped. The 75% cache read price cut changes the optimal architecture for long-running agents. Context-heavy agent loops just became cheaper to run but more expensive to complete.
• Output verbosity is now a cost variable. The 1.7x output increase means agent tracing, planning, and reflection loops need tighter prompt engineering to control token spend.
• One model, two personas. If Fable and Mythos share weights, prompt engineering becomes even more about system prompt design and routing logic than model selection.
• The multi-model wave is here. Grok 4.7 and Gemini Flash 3.8 are coming. The prompt engineer's toolkit just expanded — but so did the evaluation complexity.

The Signal

This week's Latent.Space AI News roundup checked 12 subreddits, 544 Twitter accounts, and zero Discords. The Fable/Mythos 5.1 story dominated every channel with over 12 million views. When a benchmark-shattering model launch gets this level of attention, it's not incremental — it's a terrain shift.

The pattern is becoming clear: the bottleneck in AI has shifted from "can the model do it?" to "how do we build systems that use it well?" — and then again to "how do we make it economical?" Fable/Mythos 5.1 answers the first question. The prompt engineering and infrastructure challenge is now the hard problem.