Claude Fable 5.1 and Mythos 5.1: Anthropic's Cheaper, Sharper Coding Models and the New Frontier Safeguards

In September 2026 Anthropic shipped two models at once: Claude Fable 5.1 and Claude Mythos 5.1. The official framing is refreshingly direct -- they are the same underlying model with different levels of safeguards. Fable 5.1 is generally available; Mythos 5.1 is reserved for trusted access programs in cybersecurity and the life sciences. For developers, the real story is not another benchmark spike but three engineering signals: a meaningful price cut (roughly 25 percent on typical workloads and up to about 45 percent on highly agentic ones), a new enterprise privacy mechanism called Enterprise Frontier Safeguards, and safeguards precise enough to allow vulnerability discovery without enabling exploit development.

1. Two Names, One Model

Naming Fable and Mythos separately turns safeguards into a product dimension instead of an afterthought. Fable 5.1 targets the broad run of coding and knowledge work; Mythos 5.1 carries safeguards designed for cybersecurity and the life sciences, is available only through trusted access programs, and pairs its biology capabilities with an access program developed with the US government. Anthropic is explicit that the performance gap between the two comes mostly from older, less precise safeguards -- and that today's improvements should make the difference much smaller. Translated into engineering terms: one set of weights, with the safety configuration chosen per deployment.

// Fable 5.1 ships with an explicit effort dial. Pick the level
// per task class instead of leaving one default everywhere.
// Fable 5.1 defaults to High in Claude Code, and to Medium in
// Claude Cowork and on Claude.ai.
const effortPolicy = {
  model: "claude-fable-5-1",
  codeAgent: "high",       // interactive coding in Claude Code
  cowork: "medium",        // Claude Cowork and Claude.ai default
  longRunningAgent: "xhigh", // overnight, multi-repo migrations
  quickReview: "low",      // cheap passes over a single diff
};

2. Coding Ability: The Win Is Persistence

On Terminal-Bench 4.0 and CursorBench 3.2.0, Fable 5.1 scores higher per task at a lower mean cost than its predecessor. The more convincing evidence is qualitative: during testing at the investment firm Millennium, Fable 5.1 found the root cause of a rare crash in internal systems that no engineer -- and no other model -- had been able to explain after years of trying. Anthropic also stresses that Fable 5.1 fixes root causes rather than patching around symptoms, and that its output stays readable over long, multi-step tasks. That is precisely the property teams need before handing agents longer, more autonomous assignments.

{
  "model": "claude-fable-5-1",
  "effort": "medium",
  "max_output_tokens": 8192,
  "tools": [
    { "type": "shell", "name": "run_tests" },
    { "type": "read", "name": "read_file" }
  ],
  "cache": {
    "policy": "reuse_when_unchanged",
    "note": "cache reads are priced lower on Fable 5.1; highly
             agentic loops re-read the same context every turn"
  }
}

3. Pricing: Cache Reads Compound

The price cut deserves a closer look: typical token-billed workloads cost about 25 percent less than Fable 5, driven by cheaper reads of inputs that have already been processed and stored. For highly agentic work the savings reach roughly 45 percent, because an agent loop re-reads the same context every turn and cache hits grow linearly with turns. Combined with the effort dial (low through max), teams can finally encode both model quality and per-task budget into policy: low effort for reviews, high effort for cross-repo refactors.

// Agentic savings come mostly from cache reuse: an agent loop
// re-reads the same context each turn, so a cheaper cache-read
// price compounds over dozens or hundreds of turns.
function chooseEffort(task) {
  if (task.kind === "review") return "low";
  if (task.kind === "implement") return "high";
  if (task.depth === "multi_repo" || task.risk === "prod") {
    return "xhigh";
  }
  return "medium";
}

// On Terminal-Bench 4.0, Low and Medium effort land near Fable 5
// high-effort scores at a much lower mean cost per task.
const effort = chooseEffort({ kind: "implement", depth: "single_repo" });

4. EFS: Your Data, In Your Cloud

Enterprise Frontier Safeguards (EFS) is the part architects should care about most: data is stored in cloud infrastructure controlled entirely by the customer, not by Anthropic, delivering complete privacy on par with zero data retention while staying state of the art at preventing adversarial use. EFS rolls out in phases starting later this fall; until then, eligible customers can run Fable 5.1 with zero data retention. For regulated industries, this directly answers the procurement question of whether code snippets end up in training data.

5. Precision: 60 Percent Fewer False Positives

The safeguard upgrade is about precision, not raw strength: cyber safeguards block 60 percent fewer false positives than before. Fable 5.1 can discover software vulnerabilities but will not help develop exploits for them -- a boundary written into the product. Biology access runs through the government-partnered program. For security teams the practical payoff is a scanning agent you can trust on real repositories without worrying that it will be steered into producing weaponizable exploit code.

// Before switching your default model, replay your own tasks.
// Public leaderboard numbers are a floor, not a contract.
const evalSuite = [
  {
    id: "fix-rare-crash",
    prompt: "Find the root cause of the intermittent crash.",
    effort: "xhigh",
    expect: "explain mechanism, then patch",
  },
  {
    id: "review-pr-800",
    prompt: "Review this 800-line pull request for regressions.",
    effort: "low",
    expect: "ranked findings with file:line",
  },
];

6. What to Do Today

First, put the effort dial into task routing instead of sharing one default company-wide. Second, replay your own tasks before switching defaults -- public leaderboards are a floor, not a contract. Third, pick a privacy profile by data sensitivity: standard API for internal tooling, zero retention or EFS for sensitive code. Finally, do not stop at benchmarks: hand Fable 5.1 a years-old bug your team never solved. Real root-cause work beats any chart.

// Data privacy is now a deployment decision, not a checkbox.
// Until Enterprise Frontier Safeguards (EFS) rolls out later
// this fall, eligible customers can run with zero retention.
const privacyProfile = {
  standardApi: {
    training: "opt-out default",
    note: "fine for non-sensitive internal tooling",
  },
  zeroRetention: {
    training: "none",
    note: "available to eligible customers until EFS",
  },
  enterpriseFrontierSafeguards: {
    training: "none",
    storage: "customer-controlled cloud infrastructure",
    note: "complete privacy without losing frontier safeguards",
  },
};