Anthropic's Claude Fable 5.1 Cuts Agent Costs 45 Percent

A New Model With an Old Problem in Mind

Anthropic's latest Claude release is less about chasing benchmark supremacy and more about fixing what has been bothering its paying customers. On Sept. 1, the company launched Claude Fable 5.1 for general use and a more permissive counterpart, Claude Mythos 5.1, for vetted cybersecurity and life-sciences partners 79. Both share the same underlying model and differ only in the safety guardrails wrapped around them, a structure Anthropic first introduced with Fable 5 and Mythos 5 back in June 9. Axios framed the release as a possible template for how frontier labs will need to ship models going forward: intelligence gains alone are no longer sufficient, and they must be paired with fixes to cost, reliability and safety friction that enterprise customers have been complaining about 1.

What Actually Improved

Anthropic says Fable 5.1 is meaningfully better at long-running software engineering, multi-step research, debugging across large codebases and scientific work involving experiments and data tables 78. On Terminal-Bench 4.0, a test of agentic coding, Fable 5.1 scored 55.8%, up from Fable 5's 42.0%, while the more permissive Mythos 5.1 hit 60.9% — both ahead of GPT-5.6 Sol's reported 37.3% 7. On the Terminal-Bench-Science evaluation, Fable 5.1 more than doubled its predecessor's score, moving from 24.7% to 52.6% 7. Anthropic also reported gains on computer-use tasks, general reasoning benchmarks like Humanity's Last Exam, and business-workflow automation tests, though outlets covering the release cautioned that benchmark improvements do not automatically translate into equivalent real-world performance 7.

The Real Story Is the Cache Pricing

Despite the coding and research gains, the more consequential change is financial. Fable 5.1's headline API pricing is unchanged from Fable 5, still $10 per million input tokens and $50 per million output tokens 8. What Anthropic actually cut is the price of reading cached tokens, slashing it by 75% from $1 to $0.25 per million tokens, while cache-write pricing stays at $12.50 for five-minute storage and $20 for one-hour storage 78. That distinction matters because autonomous agents tend to resend large amounts of repeated context — system instructions, codebases, tool definitions and accumulated task history — every time they act. Anthropic says cached content can make up half or more of total token usage in long-running jobs, and it estimates the new pricing lowers effective costs by roughly 25% for typical workloads and up to 45% for heavily agentic ones 78. In other words, Anthropic didn't make its flagship model cheap; it made the expensive parts of running it as an agent much cheaper.

Fable Versus Mythos, and Loosened Restrictions

The Fable/Mythos split lets Anthropic offer a highly capable model broadly while keeping its most dangerous capabilities gated. Fable 5.1 can now identify software vulnerabilities for the first time, a meaningful expansion, but exploit development and penetration testing are still routed elsewhere or restricted 7. Mythos 5.1 remains limited to U.S. organizations through a Cyber Verification Program and a Life Sciences Verification Program developed with the U.S. government, though Anthropic says it intends to expand access to international partners over time 7. This mirrors the arrangement Anthropic set up with Mythos 5 in June, when the company said the same underlying model with lifted safeguards would initially serve a small group of cyberdefenders and infrastructure providers through Project Glasswing 9. That earlier model briefly had its access suspended entirely in June after a U.S. government export-control directive tied to a reported jailbreak technique, before being restored — a reminder of how closely these releases are tied to government oversight 910.

Enterprise Trust as a Product Feature

Alongside the model, Anthropic introduced Enterprise Frontier Safeguards, which store customer data on the customer's own cloud infrastructure rather than Anthropic's 7. This addresses friction created by Anthropic's earlier decision to require 30-day data retention on Mythos-class model traffic, a policy meant to help detect jailbreak attempts but one that created real hesitation among security-conscious enterprise customers 9. Fable 5.1 and Mythos 5.1 are also the first Claude models to embed invisible watermarks in text and file outputs, a step Anthropic had already committed to for models released after Aug. 2, 2026, in line with the EU AI Act 8.

A Rapidly Layered Claude Lineup

Fable 5.1 arrives at the end of a fast sequence of Claude releases built around a tiered pricing structure. Fable 5 and Mythos 5 launched in June at $10/$50 per million input/output tokens as Anthropic's top capability tier 9. Claude Sonnet 5 followed on June 30 as a mid-tier, agent-oriented model priced initially at $2 input/$10 output per million tokens before rising to $3/$15, positioned to undercut Opus while closing much of the performance gap — its agentic coding score of 63.2% sat between Sonnet 4.6's 58.1% and Opus 4.8's 69.2% 11. Claude Opus 5 launched July 24, marketed as reaching near-Fable-level capability at roughly half the price, a pitch aimed squarely at businesses whose AI budgets were being strained by heavy token use 1213. Together, these releases sketch a routing architecture: cheaper models handle routine steps while costlier ones are reserved for the hardest parts of a task.

OpenAI Is Fighting the Same Battle

OpenAI's GPT-5.6 family — Sol, Terra and Luna — launched with a nearly identical logic: give developers tiers to match cost against capability 1415. At general availability, pricing stood at $5/$30 for Sol, $2.50/$15 for Terra and $1/$6 for Luna per million tokens 14. On July 30, OpenAI cut Luna's price by 80% to $0.20/$1.20 and Terra's by 20% to $2/$12, while leaving Sol untouched, and said Luna had become the default model for background agent automation at several partner companies 16. OpenAI's own benchmarking claims Terra and Luna can match or beat Fable 5 on certain evaluations at a small fraction of the cost, an argument that, regardless of how directly comparable the tests are, signals the same shift Anthropic is making: competition is less about raw intelligence and more about cost per completed task 14. Separately, OpenAI has said it plans to limit release of a more capable model, referred to as Astra, due to concerns about its cyber capabilities following a prior security incident involving Hugging Face, underscoring that safety-driven restraint is not unique to Anthropic 6.

Google's Flash Strategy Pulls in the Same Direction

Google has pursued cost-performance gains largely through its Gemini Flash line. Gemini 3 Flash launched with pricing of $0.50 input/$3 output per million tokens, reporting 78% on SWE-bench Verified, 90.4% on GPQA Diamond and roughly three times the speed of Gemini 2.5 Pro while using about 30% fewer tokens on average 17. Google followed with Gemini 3.6 Flash, said to cut output token use by up to 17%, Gemini 3.5 Flash-Lite for high-volume agent and document workloads, and a security-focused Gemini 3.5 Flash Cyber model initially limited to governments and partners through Google's CodeMender platform 18. Most recently, Gemini 3.7 Flash launched targeting coding and agentic work with an introductory price of $0.75 input/$3.75 output per million tokens through the end of 2026, rising to $1.50/$7.50 in 2027 1920. On raw token price, Google's Flash tier undercuts Anthropic's Fable line by a wide margin, though the two are aimed at different jobs — Flash targets high-volume, latency-sensitive work, while Fable is pitched as a costlier but more capable option for hard, long-horizon tasks.

Why the Shift Matters

The throughline across Anthropic, OpenAI and Google's recent releases is that the competitive question has moved from "which model scores highest" to "which model finishes a task for the least total cost." A cheap model that requires retries, produces excess output or loses track of context can end up costing more than an expensive model that finishes correctly in one pass. That is precisely the logic behind Anthropic's cache-price cut, OpenAI's tiered Sol-Terra-Luna routing, and Google's Flash lineup. It also reflects mounting pressure from real-world agent usage: heavy, automated consumption of models through coding agents and tool-calling systems strains subscription pricing that was never designed for that volume, pushing providers toward pricing structures that reward efficient, repeated use rather than penalizing it. Anthropic's decision to simultaneously loosen some safety-driven false-positive blocks — while keeping tighter restrictions around exploit development and advanced biology work — suggests providers increasingly view unnecessary refusals not just as a safety issue but as a reliability problem that can break automated workflows and drive customers elsewhere 7.