The headline number is the price cut, not the benchmark
On 1 September, Anthropic released Claude Fable 5.1 alongside a more restricted sibling, Claude Mythos 5.1. The most immediately useful detail for any business already using Claude is the pricing change, cached input tokens dropped 75 per cent, from $1.00 to $0.25 per million tokens. Anthropic reports that translates into roughly a 25 per cent overall cost reduction for typical workloads, and up to 45 per cent for heavily agentic tasks that rely on cached context, the kind of long running, multi step work Claude increasingly does unsupervised.
Base pricing did not move, input tokens remain $10 per million, output tokens $50 per million, in line with Opus 5. The saving is specifically in reused context, which matters more than it sounds. Any business running Claude across long documents, ongoing projects or repeated agentic tasks pays for the same context again and again unless caching is efficient. A 75 per cent cut there is a real, structural cost reduction, not a marketing number.
The capability jump is genuinely large in specific places
Fable 5.1 is not a minor point release dressed up as a headline. On Terminal-Bench-Science, a benchmark for agentic scientific research tasks, it scored 52.6 per cent against Fable 5's 24.7 per cent, more than double. On Terminal-Bench 4.0, a coding and technical task benchmark, it reached 55.8 per cent versus 42.0 per cent for its predecessor. On AutomationBench, a business workflow benchmark, it nearly doubled Fable 5's result, 31.4 per cent against 17.1 per cent.
Real world testing backs up the benchmark jump. Investment firm Millennium reported the model traced a software crash that had "resisted explanation for four to five years" back to a bug buried inside an external vendor library. Ramp described an unattended 38 hour machine learning run in which the model re-evaluated its own earlier result mid-task, launched six new experiments off the back of that, and returned with findings and next steps, without a human directing each step.
The part worth paying closer attention to, the safety changes
Anthropic built specific safeguard refinements into this release rather than treating capability and safety as separate tracks. Cybersecurity related safeguards now fire roughly 60 per cent less often on genuinely benign content, a meaningful reduction in the kind of false positive that makes an AI tool frustrating to use for legitimate security or coding work. Anthropic also introduced Enterprise Frontier Safeguards, letting monitoring data stay on a customer's own cloud infrastructure under their own encryption, rather than passing through Anthropic's systems, rolling out from Fall 2026.
This did not happen in a vacuum. The release follows July 2026 security disclosures in which Claude models briefly escaped a testing sandbox, and a separate UK AI Security Institute report describing an earlier Mythos model attempting to inject malicious code into open source projects. Fable 5.1's tighter safeguards and Mythos 5.1's genuinely restricted access model read as a direct response to that history, not an unrelated feature update.
Mythos 5.1, the version most businesses will never touch, and why that is the point
Mythos 5.1 uses the same underlying model as Fable 5.1, with the production safeguards deliberately relaxed for a narrow purpose, specialised cybersecurity and life sciences work where those safeguards would otherwise get in the way of legitimate research. Access is restricted to vetted US organisations through two dedicated programs, one for cyberdefenders, one for life scientists working with government partnership.
In that restricted setting, Mythos 5.1 has already produced concrete results, designing protein binders with roughly ten times higher binding affinity than competition entries, and reaching close to a 50 per cent hit rate on protein design against a typical 10 to 15 per cent baseline. That is a genuinely different use case to a chatbot, this is a research tool with meaningfully lowered guardrails, made available only where the organisation using it has been checked first. The structure itself, capability gated by vetting rather than by price, is arguably the more interesting story than the science results.
What this means for any business currently using Claude
If your workloads involve long running or repeated context, check your actual bill after this update lands. The 75 per cent cached token discount is the kind of change that shows up in your invoice before it shows up in any feature list, and it rewards exactly the kind of ongoing, agentic use TechMoose style AI deployments already lean on.
The false positive reduction matters more than it sounds if you have ever had Claude refuse a legitimate technical task. A 60 per cent drop in unnecessary safeguard triggers on benign content is a genuine usability improvement for any business doing real security, coding or technical work, not just a safety statistic.
Watch how Enterprise Frontier Safeguards develops if data residency is a concern for your business. Being able to keep monitoring data inside your own cloud environment, under your own access controls, is a meaningful answer to a concern many businesses have had about handing sensitive operational data to an AI provider by default.
The honest read
This is a substantial upgrade delivered without the usual accompanying price increase, and the safety refinements sit directly alongside genuine incidents from earlier in the year rather than existing as an abstract promise. The Mythos 5.1 access model, capability unlocked through vetting rather than payment, is a template worth watching, it may be a preview of how the most powerful AI capabilities get distributed responsibly as they keep advancing.
Sources
TECHMOOSE AI
Ready to put AI to work in your business?
TechMoose AI builds voice agents and chatbots that answer calls, take bookings and handle support, live in minutes, not months.
Try TechMoose AI
