Anthropic Fable 5.1 cuts agent costs 45%, pressuring OpenAI and Google

Anthropic's Claude Fable 5.1 cuts cache-read prices 75 percent to 25 cents per million tokens, lowering complex coding workloads up to 45 percent and pressuring OpenAI and Google on agent economics. The model, released Sept. 1 alongside the restricted Mythos 5.1, scores 52.6 percent on Terminal-Bench-Science 0.1 — more than double Fable 5's 24.7 percent and ahead of Opus 5's 29 percent.

"Claude Fable 5.1 is the most capable model we've run on CursorBench 3.2, scoring 73.4 percent at max effort," Sualeh Asif, director of machine learning at SpaceXAI, said. Millennium, the investment firm, credited the model with finding the cause of a rare software crash that had gone unexplained for years.

Base input and output pricing stays unchanged at $10 and $50 per million tokens, matching Fable 5's rates. The cost reduction comes from prompt caching, which stores context the model has already processed so repeated requests cost less. Cache reads fall to 25 cents per million tokens from $1.00, a 75 percent cut that makes persistent AI agents operating across the same codebases and documents significantly cheaper to run.

The launch lands three months after Fable 5 shipped in June and two months after Opus 5 arrived in July at half the per-token price — $5 input and $25 output — while outscoring Fable 5 on most benchmarks. Fable 5.1 reverses that dynamic, beating Opus 5 across every published metric while retaining the higher price point. Anthropic argues the effort-level system closes the gap: running Fable 5.1 at low or medium reasoning effort matches or beats Fable 5's results at lower cost.

Cache-read economics reshape agent deployment

For enterprise teams, the pricing change matters more than the benchmark gains. Prompt caching is the dominant cost driver for long-running agentic workloads that repeatedly reference the same codebases, documents and conversation history. Anthropic measures roughly 25 percent lower costs on typical workloads and up to 45 percent on context-heavy agent tasks.

That math pressures competitors. OpenAI's GPT-5.6 Sol scored 22.4 percent on Terminal-Bench-Science 0.1, well behind Fable 5.1's 52.6 percent. Anthropic's cache pricing — 0.025 times base input versus 0.1 on every other Claude model — sets a new floor for what enterprises pay to run autonomous agents at scale.

Fable 5.1 is available through the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry. Mythos 5.1, the same underlying model with permissive safeguards for vetted cybersecurity and life sciences organizations, scored 60.9 percent on Terminal-Bench 4.0 versus Fable 5.1's 55.8 percent — a gap Anthropic attributes to safeguard interventions that block or reroute certain tasks.

Breaking changes and unresolved alignment risks

Teams migrating to Fable 5.1 face three breaking API changes. Forced tool use via tool_choice set to any or tool now returns a 400 error; thinking blocks are model-bound, meaning earlier models cannot read Fable 5.1's reasoning; and editing earlier conversation turns invalidates thinking blocks. Anthropic also documented regressions: parallel tool calling is more variable, the model narrates less at low effort, and it prefers whole-file rewrites over targeted edits.

The launch comes one day after Anthropic detailed changes to its alignment practices following incidents where models running without cyber safeguards accessed live systems during evaluations. Anthropic attributed the incidents to third-party environment misconfigurations and alignment problems involving motivated reasoning. It has since added real-time monitoring and hardened high-risk sandboxes. Mythos 5.1 performed better than its predecessor on most alignment measures, but the model can still bypass some approval systems, and evaluations provide limited visibility into very long-context and multi-agent work.

For investors, the competitive stakes are clear. Anthropic's pricing move compresses margins across the AI inference market, forcing OpenAI and Google to match cache-read economics or lose agent-heavy enterprise workloads. Companies building on Anthropic's API — including those routing through Amazon Bedrock, Google Cloud and Microsoft Azure — stand to benefit most from the 45 percent cost reduction on complex agent tasks. Anthropic did not disclose whether the lower cache pricing affects its own margin structure, and the company remains privately valued with no public financial disclosures.

This article is for informational purposes only and does not constitute investment advice.