Anthropic Launches Claude Fable 5.1
Anthropic has released Claude Fable 5.1, focusing on long-term complex programming, agent research tasks, and cost reduction in caching; the model is now available via Anthropic API, AWS, Google Cloud, and Microsoft Azure.
Fable 5.1 is aimed at long-duration software engineering work across files and modules: Anthropic claims it can execute software project development, code reviews, and complex application code modifications, achieving results similar to or better than Fable 5 under low or medium "thinking intensity" settings while reducing actual operating costs. The company has not positioned it as a complete replacement for Fable 5 in all programming tasks.
In the Terminal-Bench 4.0 agent programming test disclosed by Anthropic, Fable 5.1 scored 55.8%, higher than Fable 5's 42.0% and GPT-5.6 Sol's 52.3%; during the same period, Mythos 5.1 scored 60.9%. These figures come from Anthropic's own benchmarking, reflecting specific task sets and evaluation configurations, and cannot be directly equated to production efficiency across all enterprise codebases.
Scientific tasks have also become a key focus of this upgrade. Fable 5.1 scored 52.6% in the Terminal-Bench-Science 0.1 agent research test, surpassing Fable 5's 24.7% and GPT-5.6 Sol's 29.0%; Anthropic states that the model can assist researchers in reading complex charts and data tables, simulating experimental results, designing experimental plans, and invoking tools in multi-step research processes.
The pricing structure has not lowered the base input and output unit prices: Fable 5.1 remains at $10 per million input tokens and $50 per million output tokens. The changes focus on caching reads, specifically reusing processed context input, with prices dropping from $1 per million tokens in Fable 5 to $0.25, a 75% reduction; 5-minute cache writes remain at $12.5, and 1-hour cache writes at $20.
Anthropic expects that the reduction in caching prices will lower the total cost of typical token-based workloads by about 25% compared to Fable 5; for complex agent tasks that frequently call long contexts, savings can reach about 45%. The extent of savings depends on cache hit rates, output token amounts, cache write durations, and the scale of task repeat reads, rather than all requests automatically receiving a 45% discount.
On the safety front, Anthropic has simultaneously launched the Enterprise Frontier Safeguards framework, allowing eligible enterprises to keep security monitoring data within their own infrastructure while executing relevant security reviews through mechanisms deployed by the enterprises. The company has also restricted Mythos 5.1 to two types of audit projects: the Cyber Verification Program and the Life Sciences Verification Program, for use by defensive cybersecurity and life sciences organizations.
In market mechanisms, the reduction in caching prices directly attracts enterprise agent buyers who need to repeatedly read codebases, enterprise knowledge bases, documents, and tool definitions; funding will shift from manual code maintenance, outsourced development, and repetitive document analysis to model calls, cloud platforms, and agent orchestration. Anthropic, along with AWS, Google Cloud, and Microsoft Azure, will benefit from the growth in high-frequency long-context reasoning; development outsourcing and low-end code maintenance services that rely on hourly billing and highly standardized work content will face pricing pressure.
Source: Public Information
ABAB AI Insight
Previously, Anthropic integrated the model into command line, desktop, and mobile development processes through Claude Code, and entered non-developer knowledge work with Claude Cowork; Fable 5.1 is not merely an expansion of chat model capabilities but is strengthening the infrastructure for "continuous task execution by the model." Traditional code completion tools primarily address function-level generation, while Claude Code places the model into repositories, terminals, and tool invocation environments; this reduction in long context caching prices specifically targets the cost bottleneck of agents needing to re-read codebases, system prompts, and task histories in each round.
The change in capital pathways is reflected in the billing units. While the base input and output pricing of Fable 5.1 remains unchanged, the caching read price has dropped from $1 per million tokens to $0.25: Anthropic is actively lowering the marginal cost of processed context to encourage enterprises to continuously host larger codebases, more knowledge bases, and longer task states within the Claude workflow. The continued charge for cache writes indicates the company's bet that once customers build persistent contexts, subsequent high-frequency calls will increase, raising the engineering costs of migrating to other models.
This is similar to how AWS creates customer stickiness through S3, databases, and network services, and how Snowflake and Databricks build computational consumption around data residency: the deeper the foundational storage or context is embedded, the easier subsequent computations are to occur on the original platform. Unlike GitHub Copilot, which provides coding assistance based on user seats, Fable 5.1 is closer to an agent economy that charges based on task states, context scales, and reasoning cycles. The improvement in Terminal-Bench scores indicates that the model is expanding from "generating code snippets" to "completing multi-step terminal tasks," but benchmark performance does not equate to capabilities in production environments regarding permissions, test coverage, code auditing, and accountability.
Essentially, this represents a shift in pricing power: software development has traditionally been priced mainly based on engineer hours, outsourced projects, and SaaS seats; as long-term agents mature, part of the budget will continuously flow towards model platforms based on context reuse, tool invocation, and task completion. The importance of caching prices lies in their determination of whether long-context workflows are economically viable: if each inference incurs the full cost of reading the entire codebase, agents can only perform short tasks; when repeated reads approach low-cost status, model vendors may compete for ongoing project-level workloads.
ABAB News · Cognitive Laws
The cost of agents lies not in answering but in repeated understanding.
The more persistent the context, the more real the platform switching costs.
Lowering marginal inference prices is essential to compete for long-term workflows.