Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 as two access tiers of what the company describes as the same underlying model, differentiated only by safeguard level. Fable 5.1 is generally available across Anthropic's own platform, AWS, Google Cloud, and Microsoft Azure, while Mythos 5.1 is restricted to trusted-access programs and, for now, US organizations only. The launch pairs concrete benchmark gains and a lower cache-read price with an explicit two-tier access model that decides who can actually use the more capable safeguard profile.
What changed
Anthropic released two models built from the same underlying weights but gated differently: Claude Fable 5.1, generally available on Anthropic's own platform, AWS, Google Cloud, and Microsoft Azure, and Claude Mythos 5.1, available only through Anthropic's "trusted access programs" and currently limited to US organizations. The company describes them explicitly as "the same model, but with different levels of safeguards."
On pricing, Fable 5.1 costs $10 per million input tokens and $50 per million output tokens, with cache reads cut to $0.25 per million tokens — a 75% reduction from previous cache pricing. Anthropic states this amounts to roughly a 25% cost reduction versus Fable 5 for typical workloads, rising to as much as 45% for heavily agentic use.
Benchmark deltas are model-specific. On Terminal-Bench-Science 0.1, Fable 5.1 scores 52.6% against 24.7% for Fable 5, 29.0% for Opus 5, and 22.4% for GPT-5.6 Sol. On Terminal-Bench 4.0 (agentic coding), Mythos 5.1 leads at 60.9%, ahead of Fable 5.1 (55.8%), Mythos 5 (42.0%), Opus 5 (52.3%), and GPT-5.6 Sol (37.3%). On the knowledge-work benchmark GDPval-AA v2, Fable 5.1 posts 1853 versus 1723 (Fable 5), 1824 (Opus 5), and 1711 (GPT-5.6 Sol). On OSWorld 2.0 computer-use tasks, Fable 5.1 reaches 77.9% (partial) / 41.7% (strict), up from Fable 5's 72.9%/36.1% and ahead of Opus 5's 75.4%/39.6%. On Humanity's Last Exam, Fable 5.1 scores 60.9% without tools and 65.0% with tools. On CursorBench 3.2.0, Fable 5.1 hits 73.4%, and on AutomationBench it reaches 31.4% — nearly double Fable 5's 17.1%.
Safety changes accompany the launch: cybersecurity safeguards now "block 60% fewer false positives," biology-related safeguards "fire 85% less often for benign requests," and Claude Code defaults to "High" reasoning effort while Claude Cowork and Claude.ai default to "Medium." Fable 5.1 can now discover software vulnerabilities, though Anthropic states it is not built to develop exploits. Enterprise Frontier Safeguards (EFS) begin rolling out in phases this fall, and a Detection API is now available to eligible organizations.
Who this affects
Anthropic API customers running agentic coding, computer-use, or terminal-automation workloads are the direct audience for Fable 5.1's benchmark gains and pricing changes, since the improvements concentrate in Terminal-Bench-Science, Terminal-Bench 4.0, OSWorld 2.0, and AutomationBench rather than across the board. Teams doing high-volume prompt caching benefit most from the 75% cut in cache-read pricing. Organizations already budgeting for Fable 5 should re-check per-workload economics, since Anthropic's stated 25–45% cost reduction is workload-dependent, not universal. Claude Mythos 5.1 matters only to organizations already enrolled in Anthropic's Cyber Verification Program or Life Sciences Verification Program, or otherwise part of a "trusted access" arrangement — it is unavailable to the general API audience and, for now, restricted to US organizations. Security teams evaluating Claude Code's vulnerability-discovery capability, and compliance teams tracking EU AI Act watermarking obligations for models released after August 2, 2026, should also take note. Casual or hobbyist API users see essentially no access change from Mythos 5.1's restrictions.
Verdict
For existing Fable 5 users on agentic coding, computer-use, or terminal-automation workloads, the benchmark gains Anthropic reports — plus a stated cost reduction of 25% to 45% and 75% cheaper cache reads — make Fable 5.1 a reasonable default to switch to now, particularly since it is a straightforward API model-name swap with no new access process. Teams whose workloads sit outside those categories have less documented reason to move immediately, since Anthropic did not publish uplift figures for other task types. Claude Mythos 5.1 is not a decision most readers face: it requires enrollment in a trusted-access program and is currently US-only, so evaluating it means applying to the Cyber Verification Program or Life Sciences Verification Program first, not flipping a model string. Anyone relying on Claude Code's expanded vulnerability-discovery behavior should review its scope internally before enabling it broadly, since Anthropic frames it as detection-only, not exploit development.
Tracked daily from official release feeds and vendor changelogs. Full archive: https://media.patentllm.org
Top comments (0)