Fable 5.1, touted as the world's most powerful, has also begun to increase cache prices.

Author: Xiaojing, Tencent Technology

On September 1st, US local time, Anthropic simultaneously released Claude Fable 5.1 and Claude Mythos 5.1.

These two models use the same underlying model, but have different security restrictions. Fable 5.1 is widely available to Pro, Max, Team, and Enterprise users, and is provided through the Claude API and major cloud platforms.

Mythos 5.1's approval mechanism has been opened to a limited number of cybersecurity and life science professionals, but access remains very limited. Fable 5.1's input and output prices remain unchanged, but cache read prices have decreased by 75%. According to Anthropic's calculations, the overall cost for typical workloads is reduced by approximately 25% compared to Fable 5, with up to 45% reduction for complex coding and highly proxyed tasks. Anthropic also launched Enterprise Frontier Safeguards (EFS), allowing enterprises to store relevant data in cloud infrastructure under their control. Eligible customers can also use Fable 5.1's zero-data-retention mode before EFS is officially available. In terms of model capabilities, Fable 5.1 outperformed Fable 5 in tests of coding, scientific research, knowledge work, and business automation. Anthropic also released real-world use cases from early customers, including long-running machine learning tasks, software troubleshooting, and browser agent tasks. A scientific research agent test scored 2.3 times higher than GPT-5.6 Sol. Anthropic has significantly changed its focus on Fable 5.1 this time: the model needs to handle longer and more complex tasks independently, continuously checking results, calling tools, and adjusting paths during execution. Anthropic's own test results show that in Terminal-Bench-Science 0.1, Fable 5.1 scored 52.6%, more than double that of GPT-5.6 Sol (22.4%), Fable 5 scored 24.7%, and Opus 5 scored 29.0%. This test examines the AI ​​agent's ability to perform scientific research tasks; the model needs to complete continuous research work in a terminal environment. In Terminal-Bench 4.0, Fable 5.1 scored 55.8%, higher than Fable 5's 42.0% and Opus 5's 52.3%. If security restrictions are relaxed, Mythos 5.1, using the same underlying model, further improved to 60.9%, becoming the highest-scoring version in this test. Similar improvements were also seen in knowledge work and business automation tasks. Fable 5.1 scored 1853 points in GDPval-AA v2, higher than Opus 5's 1824 points and Fable 5's 1723 points; it achieved 31.4% in AutomationBench, compared to 17.1% for Fable 5 and 26.9% for Opus 5; and reached 73.4% in CursorBench 3.2.0. These test results reflect the changes that the model can maintain its functionality in longer task chains. Anthropic's early customer case studies further illustrate this point. Investment firm Millennium encountered a rare software crash issue, occurring approximately once every million runs. This problem had plagued the team for four or five years, and they hadn't found the cause, including with previous models. Fable 5.1 eventually disassembled the external vendor's software library and compared the results with the core dump file left after the system crash, ultimately pinpointing the problem to a bug in the software library. The Ramp test is a continuously running machine learning task. Fable 5.1 ran unattended for 38 hours, re-examined the previous results, determined that there were errors caused by labels, and further launched 6 experiments. After running the experiments overnight, the model returned new results and next steps suggestions. There are also some tasks that are closer to the daily use cases of enterprises. Browserbase tested Fable 5.1's browser agent tasks, achieving an 82% completion rate in its most challenging test, surpassing Opus 5's 74% and Fable 5's 57%. Jane Street Capital stated that in internal testing, Fable 5.1 solved more coding problems than both Fable 5 and Opus 5, and its output remained relatively easy to understand even in long, multi-step tasks. These examples demonstrate that Fable 5.1 handles more than just a single problem with a single answer. A task might begin with data retrieval, followed by code analysis, tool calls, experimental verification, reassessing previous results, and finally, delivering a deliverable outcome. Scientific research tasks were another key application showcased by Anthropic. Protein design is particularly noteworthy among these. In modern drug development, many drugs require first identifying proteins that can bind to specific targets in the human body. Anthropic used Mythos 5.1, an open-source protein design and folding tool, to complete the design, and then submitted the model-generated results to two external institutions for experimental verification. Of the 12 targets, nearly 50% of the designs generated by Mythos 5.1 were ultimately validated as effective binders. Anthropic states that the typical hit rate in protein design is currently around 10% to 15%. On three of the targets, the model-designed binder achieved a binding capacity 10 times greater than the best design in the previous Adaptyv Bio protein design competition. The significance of this test lies in the fact that the model's work has already involved the design phase before experimentation. It can generate schemes based on the targets, then use existing protein design and folding tools for calculations, and finally provide candidate results to experimenters for verification. Anthropic also released a computational modeling case study. Fable 5.1 used radar images acquired more than 30 years ago by NASA's Magellan mission, along with existing maps covering approximately one-fifth of Venus, to train a neural network and regenerate a high-resolution elevation map covering approximately one-third of Venus. The new map shows details at a scale of approximately 2 to 3 kilometers, while the original map had a resolution of approximately 10 to 20 kilometers, and the accuracy of altitude measurement is improved by up to 25%. Anthropic plans to release this map under the Creative Commons license before commencing follow-up work on NASA's VERITAS mission and ESA's EnVision mission. In computational biology, Mythos 5.1 focuses on computational efficiency. Researchers often need to repeatedly run specialized machine learning models on GPUs, and computational speed directly impacts experimental progress. Mythos 5.1 improves the inference speed of seven open-source deep learning models by up to 2.5 times by writing custom GPU kernels and caching intermediate results, while maintaining consistent output. These models include ChromBPNet, Flashzoi, Enformer, Profluent-E1, ProGen2, and Evo2 at various scales. According to Anthropic's estimates, in genome-wide analyses, the optimized models can reduce GPU costs by 30% to 60%. Anthropic stated that such optimizations typically require performance engineering teams to spend weeks completing, while Mythos 5.1 only took a few days, using only publicly available source code. The optimization plans will be open-sourced later. Same Underlying Model, Two Security Mechanisms While Fable 5.1 and Mythos 5.1 use the same model, their application scope differs. Fable 5.1 is available on all platforms, including AWS, Google Cloud, and Microsoft Azure, and developers can also call claude-fable-5-1 via the Claude API. In terms of security strategy, Anthropic has focused on adjusting restrictions in the cybersecurity and life sciences fields. Regarding cybersecurity, Fable 5.1 can now help users discover software vulnerabilities for defensive security work. Anthropic states that compared to the cybersecurity protections previously used by Fable 5, the new protection mechanism reduces the average number of interventions in the Claude Code by approximately 60%. The life sciences field, however, remains more stringent. Ordinary users involved in life science research and development tasks will still be directed to the Opus model, while professionals can apply for access to Mythos 5.1 through the Life Science Validation Program. This program has partnered with the US government and already has initial participants. Mythos 5.1 will also be made available to vetted cybersecurity personnel through the cybersecurity validation program. Claude Security is now also powered by Mythos 5.1, used to scan for vulnerabilities in codebases and propose remediation solutions for human review. Meanwhile, Anthropic has also strengthened its data privacy and enterprise security mechanisms. The new Enterprise Frontier Safeguards (EFS) allows enterprises to store data in their own cloud infrastructure, rather than in Anthropic's systems. By default, human review is also handled by the enterprise itself. EFS will be rolled out in phases starting this fall; before official support, eligible enterprises can use Fable 5.1's zero data retention mode. Anthropic states that EFS was developed in partnership with over 100 companies across sectors including finance, healthcare, manufacturing, telecommunications, legal, retail, and the public sector. From this release, the upgrades to Fable 5.1 focus on several specific areas: stronger code capabilities, greater stability for long tasks, real-world applications in research tasks, lower cache read costs, and clearer data control for enterprises. For ordinary users, the most direct change is likely that the model can more easily and consistently complete tasks when handling complex problems; for enterprises, lower costs and data retention mechanisms lower the barrier to long-term use. Anthropic also delivers enhanced cybersecurity and life science capabilities to vetted professional users through Mythos 5.1. Fable 5.1 has thus become one of Anthropic's main models for production environments, while research and high-risk professional scenarios are separately placed within Mythos 5.1's controlled access system. The cache price has dropped by 75%. Fable 5.1's adjustments go beyond just model capabilities and security mechanisms; they also affect usage costs. Anthropic retains Fable 5's previous standard API price: $10 per million tokens for input and $50 per million tokens for output, without directly reducing the base price. However, the cache read price has dropped from $1 per million tokens to $0.25, a decrease of 75%. According to Anthropic's calculations based on actual usage in August 2026, the overall cost of a typical workload can be reduced by approximately 25%. For highly agent-driven tasks with a high proportion of cache reads, the cost reduction is even greater. However, tests by the third-party evaluation agency Artificial Analysis show that a decrease in cache prices does not necessarily mean a decrease in the actual cost of every task. Their data shows that at the highest inference strength, the average cost for Fable 5.1 to complete an Intelligence Index task is $3.76, 20% higher than Fable 5. The main reason is that Fable 5.1 uses approximately 1.7 times more output tokens than Fable 5. The reduction in cache prices still brings significant cost savings. Artificial Analysis estimates that the reduction in cache read prices saves an average of about $1.40 per task. This change is mainly reflected in agent testing, as these tasks repeatedly read previously processed context. If the inference intensity is adjusted from the highest level to xhigh, Fable 5.1's Intelligence Index score is 65, and the cost per task drops to $2.72, $1.04 lower than the highest intensity. However, this cost is still higher than Opus 5's $2.34 at the highest inference intensity. Therefore, the cost change in Fable 5.1 is not a simple "price reduction." Anthropic reduces the cost of cache reads, but the model consumes more output tokens in complex tasks, so the cost per task ultimately depends on the specific invocation method and inference intensity. Furthermore, Fable 5.1 supports batch processing. For asynchronous tasks, the input and output prices can be reduced to $5 and $25 per million tokens, respectively. The price factor for dedicated inference within the US is 1.1x, and the price for web searches is $10 per 1000 searches; web crawling is not charged separately. Anthropic also provides a price comparison between different models. Opus 5's input and output prices are $5 and $25 respectively, with cache reads costing $0.50; Sonnet 5's prices are $2, $10, and $0.20 respectively. The Fable 5.1's normal input and output prices are still higher than these two models, but its cache read price is lower than the Opus 5, only $0.05 higher than the Sonnet 5.