Anthropic Cuts Agent Costs 45% With Claude Fable 5.1

A new kind of AI launch playbook

Anthropic's rollout of Claude Fable 5.1 and the restricted Claude Mythos 5.1 on September 1 was framed by multiple outlets less as a routine model upgrade and more as a shift in how frontier AI companies compete 12. Rather than simply claiming a smarter model, Anthropic paired the release with cost cuts, altered safety behavior and new enterprise privacy tools, prompting Axios to describe it as evidence of a new release "playbook" in which intelligence alone is no longer enough to satisfy business customers 29.

Fable 5.1 is now generally available across the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry, while Mythos 5.1 remains limited to vetted organizations working in cybersecurity and life sciences through Anthropic's Project Glasswing trusted-access program 1011. Anthropic and several outlets emphasize that the two are technically the same underlying model, differentiated only by which safety restrictions are lifted 7813.

What actually improved

Anthropic says Fable 5.1 sharpens long-running agentic coding, multi-step research, and document- and spreadsheet-heavy knowledge work, while getting better at diagnosing root causes of software problems instead of patching symptoms 713. The model keeps a 1-million-token context window and supports up to 128,000 output tokens, with adaptive thinking always active 711.

Anthropic's own benchmark citation shows Fable 5.1 scoring 52.6% on Terminal-Bench-Science 0.1, versus 24.7% for Fable 5 and 29.0% for Opus 5 11. Partner-reported results cited by The Neuron include Fable 5.1 completing 82% of Browserbase's hardest agent tasks compared with 57% for Fable 5, and one launch partner running a 38-hour unattended machine-learning experiment 12. Anthropic also points to a case involving investment firm Millennium, where the model reportedly identified the cause of a system crash that had gone unexplained for years 13. These figures come from Anthropic or its partners rather than independent testing, and should be read as indicators of intended capability rather than verified, apples-to-apples comparisons.

The real cost story is caching, not sticker price

Fable 5.1's base API pricing is unchanged from Fable 5, at $10 per million input tokens and $50 per million output tokens 1014. The actual price cut applies to cached-context reads, which drop 75%, from $1 to $0.25 per million tokens 101114. Cache writes remain $12.50 per million tokens for a five-minute cache and $20 per million for a one-hour cache 14.

Because agentic workflows repeatedly reread the same code, instructions and conversation history, Anthropic says cached tokens can account for half or more of total usage in long-running tasks 14. The company estimates this translates into roughly 25% lower costs for typical workloads and up to 45% savings for heavily agentic ones 11016. Coverage from TechSpot and MarkTechPost both stress that this is a caching discount rather than a blanket reduction in Fable's headline rate, meaning the savings scale with how much repeated context an application actually generates 1114.

Safety tuned for fewer false alarms

Alongside performance and pricing changes, Anthropic adjusted its safety classifiers to reduce what it calls unnecessary interruptions. The company reports roughly 85% fewer interventions on benign biology and medical questions, and about 60% fewer cybersecurity-related interventions per session for Claude Code users 1910. Fable 5.1 can now be used to identify software vulnerabilities defensively, though exploit development still routes to more restricted models 1016.

Axios reports that this responds directly to developer frustration; one routing-company founder said his users wanted to rely less on Claude specifically because of overly aggressive safeguards 9. Anthropic had previously acknowledged, when Fable 5 launched in June, that it deliberately erred toward caution and expected to loosen restrictions over time 915. The Mythos tier preserves that caution for now: Mythos 5.1 keeps expanded biological and cybersecurity capabilities walled off, including a new Life Sciences Verification Program intended to open scientific research access more broadly in the future 1011.

New privacy tools for enterprise buyers

The release also included enterprise-facing privacy measures, described as Enterprise Frontier Safeguards, that would let business customers store data on their own cloud infrastructure while Anthropic monitors for misuse 116. This builds on Anthropic's existing policy of 30-day data retention for Mythos-class model traffic, intended to help detect jailbreak attempts and safety incidents 15. That retention requirement creates tension for regulated industries that want both strong misuse monitoring and minimal data exposure.

Where Fable 5.1 sits in Anthropic's lineup

Fable 5 and Mythos 5 debuted in June as Anthropic's first "Mythos-class" models, pushing capability above the existing Opus tier, with Mythos access deliberately restricted due to its cyber and biological capabilities 915. Weeks later, Anthropic released Claude Sonnet 5 as a cheaper, agent-focused midsize model priced at $2 per million input tokens and $10 per million output tokens, rising to $3/$15 after August 31 17. Sonnet 5 scored 63.2% on an agentic-coding benchmark, positioned between Sonnet 4.6's 58.1% and Opus 4.8's 69.2% 17. Together, the four tiers form a ladder: Sonnet 5 for high-volume agent work, Opus for premium reasoning, Fable 5.1 as the highest broadly available capability with cheaper repeated-context costs, and Mythos 5.1 as the gated frontier tier for trusted partners.

Google presses on price and speed with Gemini

Anthropic's move lands amid an aggressive Gemini release cadence from Google. Gemini 3.1 Pro entered preview in February with a 1-million-token context window, multimodal input support, and pricing of $2 per million input tokens and $12 per million output tokens for prompts under 200,000 tokens 1819. Independent testing cited by index.dev found Claude Opus 4.8 leading on SWE-bench Verified at 88.6% versus Gemini 3.1 Pro's 80.6%, while Gemini undercut Claude sharply on price 20.

The Wall Street Journal has reported that Google is preparing Gemini 3.8 Flash, internally known as "Skimaki," for release this week, with internal testing suggesting meaningful gains in coding ability, an area where Google has trailed Anthropic and OpenAI 35. Other coverage notes Google, alongside Anthropic and World Labs, has also been active in adjacent categories such as cybersecurity-focused models developed with partners like Nvidia and CrowdStrike 6.

Why the comparison matters

The emerging pattern across both companies is that raw intelligence gains are being paired with aggressive cost and reliability engineering aimed specifically at agentic workloads. Anthropic is targeting the expense of long-running, tool-using agents through cache pricing, while Google has leaned on a series of lower-cost Flash-tier Gemini models to compete on throughput and price. As AI systems increasingly run for hours across many tool calls rather than answering single prompts, the practical battleground is shifting from which model tests best on a single benchmark to which one can complete reliable, multi-step work at the lowest total cost. Separately, regulatory pressure is also building around AI-generated content: New York's new AI advertising disclosure law has already drawn its first complaints over undisclosed synthetic media, a sign that scrutiny of AI outputs is intensifying alongside the capability race 4.