Claude Fable 5 for Coding Agents: What It Is and How to Run It

Published October 13, 2026 · by the AQ team

Claude Fable is Anthropic's frontier tier above Opus: the first Mythos-class model made generally available, launched as Claude Fable 5 on June 9, 2026 with the Claude 5 generation and updated to Claude Fable 5.1 (the current version) on September 1, 2026. It costs $10 per million input tokens and $50 per million output, two and a half times Claude Opus 5.5's $4 and $20, and Anthropic positions it for long-horizon agentic work: multi-hour coding sessions, codebase-wide changes, unattended runs that recover from failed steps. Two things make it unlike earlier Claude launches for agent users: safety classifiers that reroute flagged requests (most often cybersecurity and biology) to an Opus model mid-session, and the fact that it is not the default anywhere, not in Claude Code, not on any plan, not in any harness. You opt into Fable deliberately, usually for the tasks where cheaper models stall.

What Anthropic shipped

Specs below are from Anthropic's model documentation as of October 2026. Effort is the knob that steers how much adaptive thinking the model spends; Fable exposes five levels and defaults to high:

The cache read price is the quiet headline for agent work: agent loops resend a large stable prefix (system prompt, tool definitions, conversation so far) every turn, so long sessions are dominated by cached input. At $0.25 per million cached tokens, Anthropic estimates typical workloads run about 25 percent cheaper than on Fable 5 and highly agentic workloads up to about 45 percent cheaper, at unchanged base prices. The million-token context window is on by default in Claude Code.

The launch history matters if your team depends on model availability. Fable 5's access was suspended three days after the June launch and restored on July 1, 2026 with stricter classifiers that Anthropic said would flag some benign coding and debugging requests; Fable 5.1 moved the needle back, with about 60 percent fewer cybersecurity-safeguard interventions per Claude Code session, per Anthropic as of September 2026.

What the Fable and Mythos names mean

Fable and Mythos are the same underlying model with different safeguards. Claude Mythos 5.1 is available only to organizations verified through Anthropic's trusted-access programs (for example the Cyber Verification Program for defensive security work); Claude Fable 5.1 is the generally available version, with safety classifiers screening requests. A flagged request is rerouted to a less capable model (cybersecurity to Claude Opus 4.8, biology to Claude Opus 5) and not billed at Fable prices; API callers handle this through the fallbacks parameter, and Claude Code switches the session to the fallback automatically with a transcript notice (return with /model, or set it to ask first). For ordinary application development the reroute is rare, but security-adjacent work (penetration testing tooling, exploit analysis, binary vulnerability scanning) should expect to land on Opus, not Fable. One more operational difference: Fable requires 30-day data retention by default for safety monitoring.

What Anthropic says about coding

Anthropic's September 2026 announcement calls Fable 5.1 its most capable model for coding and knowledge work, with large gains on terminal-driven agent benchmarks: 55.8 percent on Terminal-Bench 4.0 (against 42.0 for Fable 5, 52.3 for Claude Opus 5, and 37.3 for OpenAI's GPT-5.6 Sol in Anthropic's comparison), 52.6 percent on Terminal-Bench-Science 0.1, and 73.4 percent on CursorBench 3.2.0. The long-run claims are the ones aimed at agent users: Anthropic says Fable 5.1 stays readable over long multi-step tasks, avoids shortcuts, writes its own tests to check its work, and handles long unattended sessions better (its customer examples include a 38-hour unattended run). It also notes that at low or medium effort Fable 5.1 performs similarly to or better than Fable 5 at much lower cost, worth knowing before leaving the high default on for everything.

The independent read

Independent numbers are messier than usual because Artificial Analysis revised its Intelligence Index between the two Fable releases. At launch, Fable 5.1 at max effort topped the then-current index at 66, the highest score it had measured. On the revised index (v4.3.2, as of October 2026), which measures Fable with its safeguard fallback active, Fable 5.1 at max effort scores 53, fifth overall and below Claude Opus 5.5 at 58, at $7.63 per index task ($2.37 at low effort). Two cautions before reading that as "Opus is better": the with-fallback methodology scores a task at zero when a safeguard intervenes, which Anthropic says likely understates performance, and Artificial Analysis measured an 11x spread in output tokens across the five effort levels, so a single leaderboard number for this model is a range wider than most (the argument of our guide to benchmarking the model or the harness). The way to settle it for your codebase is running Fable against your incumbent on your own repository and judging the diffs.

How to run it in the agent CLIs

Claude Code. Fable 5.1 requires Claude Code v2.1.257 or later (run claude update) and is never the default; the fable alias resolves to Fable 5.1:

# In a session: switch model (Enter saves as default, s = this session only)
/model fable

# At startup
claude --model fable

# Pin Fable 5 specifically
claude --model claude-fable-5

# Effort: low | medium | high (default) | xhigh | max
/effort xhigh
claude --effort max

# Permanent default in settings.json
"model": "fable"

The 1M context window is on by default, and administrators can restrict Fable with the availableModels managed setting.

On a subscription. Fable is plan-covered only at the top tiers, as of October 2026: Max plans and premium Team and Enterprise seats can spend up to 50 percent of their weekly usage limits on Fable at no extra cost (it draws the same weekly limit as other models, just faster). Pro plans and standard seats run Fable on pay-as-you-go usage credits from the first token; Claude Code shows a consent prompt before billing credits, but non-interactive mode (claude -p) bills without asking, which matters for scripted runs. Our guides to Claude Code usage limits and subscription vs API credits cover the mechanics.

Cursor. Cursor's docs recommend Fable 5.1 and keep Fable 5 available, as of October 2026. With Privacy Mode enabled you approve Fable's data retention policy first (Enterprise needs an admin to enable the model), usage bills from Cursor's third-party models pool at Fable's list prices, and guardrail-flagged requests route to Claude Opus automatically.

GitHub Copilot. Fable 5.1 went generally available in Copilot in early September 2026 for Copilot Pro+, Business, and Enterprise, across the pickers in VS Code, Copilot CLI, the coding agent, and github.com; the organization policy is off by default and the model bills at provider list pricing.

Everything else. Any harness that takes an Anthropic API key can call claude-fable-5-1: OpenCode users add it under the anthropic provider, and the model is live on Amazon Bedrock, Google Cloud, and Microsoft Foundry for harnesses wired to those platforms. Mind the default-effort asymmetry when comparing harnesses: Claude Code defaults Fable to high effort while some surfaces default to medium, so the same model can behave differently out of the box.

When to pick it, when not to

Pick Fable 5.1 when the task is long-horizon and the cost of a failed run exceeds the cost of tokens: multi-hour refactors across a large codebase, overnight runs that must recover from failed steps rather than stall, and work where you have measured cheaper models looping or taking shortcuts. Anthropic's own guidance points the same way: start with Opus 5.5, reach for Fable when your evaluations on Opus at higher effort still fall short. The 97.5 percent cache discount shrinks the premium over Opus on long sessions, though output tokens at $50 per million still dominate heavy-thinking runs.

Look elsewhere when the work is routine (Opus 5.5 at $4/$20 or Sonnet 5.5 at $2/$10 finish most day-to-day tasks for a fraction of the price), when your work regularly touches domains the classifiers reroute to Opus anyway, when your organization cannot accept 30-day data retention, or when every Fable token on your plan tier bills to usage credits nobody budgeted. And as with every launch: a model earns a default on your repository, your tests, and your incumbent, not on a leaderboard.

Where AQ fits

AQ is the multiplayer coding harness where engineering teams run AI coding agents like Claude Code and Codex together: shared live terminals, a code editor, and app previews, in your own cloud. A model built for 38-hour unattended runs needs somewhere to run for 38 hours, and a laptop is not it: on AQ, Claude Code runs in a persistent tmux session on your team's VM, so a Fable run survives a closed laptop and resumes from any device, with teammates watching the same live session. Evaluating Fable is a same-day experiment: open two workspaces on the same repository (each an isolated git worktree on its own branch), run /model fable in one and your incumbent in the other, give both the identical task, and compare the sessions, not just the diffs (our guide to reviewing an AI coding session covers what to look for). Everyone signs into Claude Code with their own Claude account, so Fable's plan rules and consent prompts behave exactly as they do locally, and AQ never marks up usage on your own subscriptions.