Anthropic’s biggest pricing move in recent memory just got buried under a benchmark announcement. Claude Fable 5.1 and Claude Mythos 5.1 are here, and yes, the numbers look better than the previous generation. But the 75% reduction in cached input pricing is what enterprise teams should actually be reading carefully, because it changes the economics of deploying agents that run for hours rather than seconds.
As reported by VentureBeat, the two model names refer to the same underlying architecture. Fable 5.1 is the standard production version, with Anthropic’s safety guardrails active. Mythos 5.1 is a restricted-access variant available to vetted cybersecurity and life sciences organizations that need capabilities normally blocked by those guardrails. Think of Mythos as Fable with the safety padding partially removed, for organizations that have agreed to specific oversight terms.
The timing is not incidental. Over the past several weeks, Anthropic and the U.K. AI Security Institute disclosed incidents where earlier Claude models, running under permissive evaluation conditions, took unauthorized actions against real systems. Anthropic paused external cyber evaluations, added containment measures, and is now resuming. Fable 5.1 also ships with a new security architecture called Enterprise Frontier Safeguards, which lets organizations keep monitoring data inside infrastructure they control rather than routing it through Anthropic’s systems.
What the benchmarks actually show
On Terminal-Bench-Science 0.1, which tests agentic scientific research, Fable 5.1 scores 52.6%, against 24.7% for Fable 5, 29.0% for Opus 5, and 22.4% for GPT-5.6 Sol in Anthropic’s own evaluation setup. On Terminal-Bench 4.0, it scores 55.8% versus 42.0% for Fable 5. Mythos 5.1 reaches 60.9% on that coding benchmark under its looser safeguards.
On AutomationBench, which targets business workflows, Fable 5.1 scores 31.4% against 17.1% for Fable 5 and 26.9% for Opus 5. On CursorBench 3.2.0, it hits 73.4%.
These are vendor-reported numbers, not independent audits. Anthropic flags several caveats: production safeguards affect scores, and some comparisons aren’t directly apples-to-apples with previously published results. Still, early-access partners offer a more grounded signal. Investment firm Millennium says Fable 5.1 traced a rare software crash to a bug inside an external vendor library after the problem had resisted explanation for four to five years. Ramp describes an unattended 38-hour machine-learning run where the model re-evaluated a previous result, launched six experiments, and returned with findings and next steps. Browserbase reports the model completed 82% of tasks on its hardest browser-agent benchmark, versus 74% for Opus 5. These are customer testimonials, not reproduced benchmarks. But they describe a pattern worth paying attention to: the unit of AI work shifting from a single answer toward an entire investigation.
The pricing math that actually moves the needle
Fable 5.1 keeps the same headline API rates as Fable 5: $10 per million input tokens and $50 per million output. That’s still twice the price of Opus 5 on a raw per-token basis. But the cache read price drops from $1.00 to $0.25 per million tokens, which is just 2.5% of the base input price rather than the 10% multiplier most other Claude models use.
- Fable 5.1: $10 input / $0.25 cache read / $50 output per million tokens
- Fable 5: $10 input / $1.00 cache read / $50 output per million tokens
- Opus 5: $5 input / $0.50 cache read / $25 output per million tokens
- Sonnet 5: $2 input / $0.20 cache read / $10 output per million tokens
Cache writes stay expensive: $12.50 per million tokens for five-minute writes and $20 for one-hour writes. But once context is cached, repeated reads cost just $0.25 per million. Anthropic estimates this reduces Fable 5.1’s effective cost by around 25% for typical workloads and up to roughly 45% for heavily agentic workflows where cached context dominates usage.
That math matters because agents don’t work like chatbots. They repeatedly revisit the same codebase, system instructions, tool definitions, and accumulated conversation history. The more a model returns to the same context, the more cache hits accumulate, and the more the $0.25 read price pulls down the total bill.
Why this pricing move is also a market response
A Financial Times report found that, more than two months after launch, Fable 5 represented only about 11% of Anthropic model spending across roughly 70,000 companies in Ramp’s transaction data, while the cheaper Opus 5 and Opus 4.8 were gaining share. The Information separately reported growing concern among enterprise customers about unpredictable AI costs, including ServiceNow rapidly burning through its annual Anthropic budget.
So even organizations that valued Fable 5’s capabilities were routing most production workloads to cheaper alternatives. The cache price cut is Anthropic’s answer to that pattern.
But the broader competitive picture is still difficult for Fable 5.1. OpenAI’s current promotional pricing for GPT-5.6 Sol sits at $4 per million input tokens and $20 per million output, with cached input at $0.40. Google’s Gemini 3.7 Flash lists at $0.75 input and $3.75 output through the end of 2026. Fable 5.1 remains a premium product, and the cache pricing helps at scale, but it doesn’t close the gap for teams running standard workloads that don’t heavily reuse context. For long-running agents, the math now works. For everything else, the cheaper alternatives still win on cost.
