The new Claude Fable 5.1: Pricing, Coding, Agents, Benchmarks, and Mythos 5.1

Anthropic released Claude Fable 5.1 on September 1, 2026 as its generally available frontier model for coding and knowledge work, alongside Claude Mythos 5.1, a restricted-access version of the same underlying model with different safeguards for cybersecurity and life-sciences use.

The release combines capability gains with a material change in usage economics: standard API pricing remains $10 per million input tokens and $50 per million output tokens, while prompt-cache reads fall to $0.25 per million tokens, 75% below the Fable 5 cache-read price.

Anthropic estimates that the cache change lowers typical usage-based Fable workloads by about 25% and highly agentic workloads by as much as roughly 45%, based on actual August 2026 usage patterns; those figures are vendor estimates and depend heavily on how much context can be reused.

Fable 5.1 is available through Anthropic's own platforms and API under claude-fable-5-1, with availability also through Amazon Web Services, Google Cloud, and Microsoft Azure, making the release immediately relevant to developers and enterprise teams rather than a future roadmap item.

··········

PRICING CHANGES FAVOR REPEATED CONTEXT AND LONG-RUNNING AGENTS.

Cache reuse is where the new economics differ; raw input and output token prices remain unchanged from Fable 5.

........

Pricing or deployment factor	Claude Fable 5.1	Operational effect
API input tokens	$10 per million	Base input rate is unchanged from Fable 5.
API output tokens	$50 per million	Long outputs can still dominate total spend.
Prompt-cache reads	$0.25 per million	75% lower than Fable 5 cache reads.
Typical workload estimate	About 25% lower total cost	Anthropic estimate based on measured usage, not a guaranteed discount.
Highly agentic workload estimate	Up to roughly 45% lower	Largest effect when reusable context and cache reads represent a large share of token consumption.
API model ID	claude-fable-5-1	Available now for developer integration.

........

The lower cache-read rate changes total cost most strongly when an agent repeatedly reuses large system prompts, repository context, tool schemas, retrieved documents, or accumulated working state across many calls.

One-shot prompts with little reusable context receive far less benefit, while output-heavy tasks can still remain expensive because the $50 per million output-token rate has not changed.

For deployment decisions, the relevant variables are cache-hit share, average output length, number of tool iterations, and the length of persistent context; the headline 25% and 45% figures should therefore be treated as workload-dependent estimates rather than universal discounts.

··········

LONG-RUNNING CODING AND AGENT WORK ARE THE PRIMARY PRODUCT TARGET.

Anthropic is optimizing Fable 5.1 around sustained planning, tool use, verification, and recovery across complex workflows.

Anthropic describes Fable 5.1 as a model for work that can run for hours across applications, with emphasis on planning, tool use, recovery from failed steps, verification, and maintaining readable progress over long task chains.

Product defaults reflect that positioning: Fable 5.1 uses High effort by default in Claude Code, while Claude Cowork and Claude.ai default to Medium effort, so benchmark or cost comparisons should specify effort level because higher reasoning effort can change both latency and spend.

In software engineering, the intended workloads include codebase-wide feature work, code review, performance analysis, automated testing, design verification with vision, and multi-day autonomous sessions; Anthropic's launch examples from early-access customers are useful directional signals but remain vendor-selected evidence.

Cybersecurity behavior also changed: Fable 5.1 can be used to identify software vulnerabilities, and Anthropic says the updated safeguards trigger about 60% fewer interventions per Claude Code session than the safeguards used with Fable 5.

The boundary remains explicit: penetration testing, exploit generation, and binary-based vulnerability scanning continue to route to Opus models under Anthropic's current policy, so the release expands defensive usefulness while retaining cyber controls.

Enterprise Frontier Safeguards add a separate deployment dimension. Anthropic says EFS will keep customer data in cloud infrastructure controlled by the customer while supporting misuse detection, with rollout planned in phases beginning later in fall 2026; until then, eligible customers can use Fable 5.1 under zero-data-retention arrangements.

Because EFS is staged for future rollout, it should be treated as a roadmap capability rather than a feature already available across every listed platform.

··········

ANTHROPIC REPORTS BROAD BENCHMARK GAINS, WITH IMPORTANT COMPARABILITY LIMITS.

The launch data show gains across coding, scientific research, automation, and multidisciplinary reasoning, but the scores come from Anthropic's published evaluation setup.

........

Benchmark	Fable 5.1	Fable 5	Opus 5	GPT-5.6 Sol
Terminal-Bench-Science 0.1	52.6%	24.7%	29.0%	22.4%
Terminal-Bench 4.0	55.8%	42.0%	52.3%	37.3%
AutomationBench	31.4%	17.1%	26.9%	19.6%
CursorBench 3.2.0	73.4%	70.5%	70.0%	67.2%
Humanity's Last Exam, no tools	60.9%	57.8%	56.6%	—

........

Mythos 5.1 reaches 60.9% on Terminal-Bench 4.0 in Anthropic's published table versus 55.8% for Fable 5.1, even though Anthropic states that the two versions use the same underlying model; the company attributes the gap primarily to safeguard interventions affecting Fable's permitted task execution.

These figures should be read as vendor-published evaluations because Anthropic notes that production safeguards were enabled and that safeguard interventions affected some benchmark runs, including cases where tasks received zeros or were completed through other Claude models.

Effort settings, tool access, harness differences, and safety routing can all affect both score and cost, so rank order on a launch table should be supplemented with an internal evaluation that uses the same prompt corpus, tool environment, latency limits, and review standards as the production workload.

The practical signal is broad improvement across several agentic categories, while the exact magnitude of the advantage still requires workload-specific validation before procurement, routing, or architecture decisions are changed.

··········

FABLE 5.1 IS MOST COMPELLING WHERE FAILURE AND SUPERVISION ARE EXPENSIVE.

The release is easiest to justify when frontier capability, long context reuse, and reduced human oversight all contribute to completed-task economics.

Fable 5.1 has its strongest economic fit when a task combines expensive failure, long context, repeated tool calls, and enough cache reuse to absorb frontier-level base pricing.

Teams running persistent coding agents, research workflows, incident investigations, multi-application automation, or document-heavy analysis have a clearer reason to test the model because both the capability changes and the cache economics target those usage patterns.

Routine summarization, short drafting, or simple single-turn analysis may still be better served by lower-cost models, since Fable 5.1 retains $10 per million input tokens and $50 per million output tokens and receives limited benefit from cache pricing when most prompts contain fresh context.

Claude Mythos 5.1 belongs to a different access decision: it uses the same underlying model with more permissive safeguards for vetted cybersecurity and life-sciences users, while general Claude customers receive Fable 5.1.

For enterprise procurement, Enterprise Frontier Safeguards should remain separated from today's release decision until the staged rollout reaches the required platform and deployment environment.

The decision rule is operational: benchmark Fable 5.1 on the real workload, record cache-hit share, tool-loop length, output volume, completion quality, and human supervision time, then compare total cost per completed task against the model currently in production.

FOLLOW US FOR MORE.

·····

DATA STUDIOS

·····

[datastudios.org]