Anthropic just made running AI agents up to 45 percent cheaper and doubled a key science benchmark score in the same release. On September 1, 2026, the company launched Claude Fable 5.1 alongside a gated companion model, Claude Mythos 5.1, and the pairing is quickly becoming the new baseline for Claude Fable 5.1 AI agents used in coding, research, and long-running business workflows. Three months after Fable 5 shipped, Anthropic says this update is built specifically for agentic work that runs for hours instead of minutes.

For anyone building or buying AI agents, the timing matters. Token costs have been the biggest complaint about agentic workflows all year, and Fable 5.1 is Anthropic’s direct answer. Below, we break down what actually changed, what the benchmarks show, and how businesses should think about adopting it.

What’s New in Anthropic Fable 5.1

Fable 5.1 and Mythos 5.1 are technically the same underlying model, split apart only by safeguard level. Fable 5.1 is generally available to everyone, while Mythos 5.1 is reserved for vetted cybersecurity researchers and life sciences organizations enrolled in Anthropic’s trusted access programs. Both carry a 300k token context window with a 1M token maximum, and a knowledge cutoff of June 2026. Full details are in Anthropic’s official announcement.

The headline numbers come from Terminal-Bench, Anthropic’s testbed for real coding and system tasks. Fable 5.1 jumped from 42.0 percent to 55.8 percent on Terminal-Bench 4.0, while Mythos 5.1 reached 60.9 percent. The bigger leap came on Terminal-Bench-Science, a benchmark covering data analysis, model fitting, and theorem proving, where the score more than doubled from 24.7 percent to 52.6 percent. Anthropic frames this as an early signal that its models are becoming useful research collaborators, not just coding assistants.

Safeguards also loosened in useful ways. Fable 5.1 is now cleared to help identify software vulnerabilities for defensive security work, something earlier models often refused. Anthropic reports roughly 60 percent fewer cybersecurity safeguard interventions per session in Claude Code, and an 85 percent drop in biology and medical fallback refusals, meaning researchers spend less time arguing with the model and more time getting work done.

AI Agent Cost Reduction 2026: What Actually Got Cheaper

Base pricing for Fable 5.1 held steady at $10 per million input tokens and $50 per million output tokens, the same as Fable 5. The real change is in prompt caching, which matters enormously for agentic workflows that repeatedly reread the same context, tool definitions, and conversation history.

Cache reads dropped from $1.00 to $0.25 per million tokens, a 75 percent cut. Cache writes stayed at $12.50 per million tokens, with 1-hour cache writes at $20 per million tokens. Anthropic estimates this translates to roughly 25 percent savings on typical token-billed workloads and up to 45 percent for highly agentic tasks that lean hard on cached context.

This directly addresses a problem we covered in our piece on AI agent costs in 2026: agentic workflows can consume 5 to 30 times more tokens than a simple chatbot exchange, because agents reread context on every step of a multi-step task. Cache-read pricing is exactly the lever that determines whether that repeated context reading bankrupts a project or barely registers on the bill.

How Businesses Can Put Claude Fable 5.1 AI Agents to Work

For teams already running Claude-based agents, the upgrade path is mostly automatic, but a few practical moves make the difference between noticing savings and missing them entirely.

First, revisit prompt caching configuration. Workflows that were not caching aggressively before now have a much stronger reason to start, since cache reads are a quarter of their old cost. Second, long-horizon coding and research tasks are where Fable 5.1 shows the clearest gains, so teams running brownfield debugging, multi-step data analysis, or extended agentic coding sessions similar to what we described in our agentic coding trends coverage should see the most immediate benefit. Third, if your work touches defensive cybersecurity or biomedical research, expect noticeably fewer safeguard interruptions, which means less time spent rephrasing requests or building workarounds.

Smaller teams evaluating whether to switch from Fable 5 or a competing model should weigh the 1M token context ceiling and the science benchmark jump against their actual workload. A support-ticket agent probably will not notice the science gains, but a research or engineering team running long agentic sessions almost certainly will.

What Fable 5.1 Signals for the Next Wave of AI Agents

Fable 5.1 lands in the middle of an unusually crowded few weeks. OpenAI is preparing to release Astra, which reportedly crosses the company’s highest cybersecurity capability threshold and is headed for a tightly controlled rollout rather than a broad launch. That contrast is telling: Anthropic loosened safeguards on defensive security work while OpenAI is tightening control around offensive capability in its next model, suggesting the two labs are placing different bets on how much autonomy AI agents should have in sensitive domains.

The cost story is likely the bigger deal for most businesses. As agentic AI shifts from single-prompt tools to systems that operate across several business functions at once, cache-read pricing will increasingly decide which vendors are affordable to run at scale. Expect competitors to respond with their own caching discounts before the end of the year, following the same pattern we saw with Claude Sonnet 5 pricing earlier this summer.

Key Takeaways

Claude Fable 5.1 delivers three things that matter for anyone running AI agents today. It cuts real-world agentic costs by up to 45 percent through cheaper prompt caching, without raising base token prices. It more than doubles Anthropic’s science benchmark score, pointing toward AI agents that can meaningfully assist with research, not just code. And it removes a meaningful share of the safeguard friction that has slowed down legitimate defensive security and biomedical work.

If you are building or evaluating AI agents, Fable 5.1 is worth testing against your actual workload rather than the benchmark headlines alone. Explore more tools, comparisons, and deployment guides at BigAIAgent to see how the latest agentic AI releases stack up for your use case.

What would a 45 percent drop in your agent’s token costs let you build that was not worth building last month?