How Claude Fable 5.1 Cuts Agent Costs and Boosts Long‑Running Research Tasks

Model variants

Claude Fable 5.1 and Claude Mythos 5.1 are built on the same underlying model. The distinction lies in safety limits and access scope: Fable 5.1 is open to all users, while Mythos 5.1 is restricted to verified cybersecurity and life‑science professionals.

Cache‑read price reduction and agent cost impact

Input token price remains $10 per million tokens and output token price $50 per million tokens. Cached reads of already‑processed context are now $0.25 per million tokens, a 75 % discount.

Anthropic’s internal usage data (week of 8 September) shows overall workload cost can drop to roughly 75 % of Fable 5, and heavily agent‑oriented workloads can fall to about 55 % (up to a 45 % reduction).

Artificial Analysis measured that Fable 5.1 consumes ~1.7 × more output tokens than Fable 5, causing some tasks to cost ~20 % more. The same tests reported an average per‑task saving of $1.40, and on the “xhigh” inference tier a single task costs $2.72, $1.04 less than the highest tier.

Research benchmark improvement

On the Terminal‑Bench‑Science 0.1 benchmark (Stanford‑led), Fable 5.1 achieved a score of 52.6 %, more than double Fable 5’s 24.7 % and well above Opus 5 (29.0 %) and GPT‑5.6 Sol (22.4 %). Scores are zero for tasks that trigger safety rules, illustrating the effect of active security guards.

Long‑task case studies

Millennium traced a rare crash bug that appeared once per 1 M runs after Fable 5.1 decompiled an external library and compared core dumps.

MongoDB used Fable 5.1 to autonomously explore internal services, design, and implement a complex prototype over several hours without human supervision.

Ramp ran a 38‑hour autonomous experiment in which the model identified a mislabeled dataset, corrected the experiment, launched six parallel runs, and returned results with next‑step recommendations.

Mythos 5.1 research capabilities

In molecular design, Mythos 5.1 generated binders with ~10 × higher affinity than the best solutions in the Adaptyv Bio competition and achieved a ~50 % hit rate across 12 targets (typical rates 10‑15 %).

Using 30‑year‑old Magellan radar data, Mythos 5.1 produced a Venus elevation map with spatial resolution improved from 10‑20 km to 2‑3 km and height accuracy gains up to 25 %.

In computational biology, Mythos 5.1 wrote custom GPU kernels and cached intermediate results, accelerating inference of seven open‑source protein‑genome models by up to 2.5 × and potentially reducing whole‑genome GPU costs by 30‑60 %.

Security mechanism evolution

Mythos 5.1 passes the latest biological safety tests but remains below the next risk tier, so access restrictions persist.

Network‑security testing shows Mythos 5.1 is Anthropic’s strongest released model, yet it stays in a low‑risk compliance bracket. Compared with Claude Code, the new protection reduces average security‑intervention counts by ~60 %.

Reward‑hacking attempts and successful cheats are lower than in the previous generation, though occasional bypasses still occur in ultra‑long contexts, multi‑agent collaboration, or theoretically impossible tasks.

Model governance and provenance features

New anti‑distillation safeguards prevent newly created API accounts from editing historical multi‑turn context while retaining internal reasoning records, closing a known distillation attack vector. Existing accounts are unaffected for now.

The “dual‑version” approach (Fable vs. Mythos) separates capability from permission boundaries, offering stricter network‑ and bio‑security to verified professionals.

Text watermarking embeds a non‑visible numeric marker in generated text without affecting quality or revealing user identity. Detection APIs are being rolled out to regulators, media, fact‑checkers, academia, and compliant enterprises.

Using Claude Fable 5.1

Fable 5.1 is available on Anthropic’s platform and via AWS, Google Cloud, and Microsoft Azure. Specify the model name claude-fable-5-1 in the Claude API.

Developers should monitor:

Additional output tokens consumed relative to Fable 5.

Whether cache hits translate into real cost reductions for long tasks.

If the “xhigh” inference tier can cover typical workloads without resorting to the highest tier.

Frequency of security‑mechanism false positives on testing, scripting, and network tasks.

Stability of goal‑keeping in extended autonomous runs.

For developers in China, configure a compatible base URL, e.g. set ANTHROPIC_BASE_URL to https://code.ai80.vip, to avoid payment and endpoint issues.