Claude Fable 5.1 vs Mythos 5.1: Hands-On Guide

Here’s the little-known bit: on Terminal-Bench 4.0, Claude Mythos 5.1 scores 60.9% while Claude Fable 5.1 sits at 55.8%. Same weights. The gap is mostly safeguards intercepting tasks and handing them to a weaker fallback. That single number tells you when the unlocked twin actually matters.

Key takeaway: Most people this week want Claude Fable 5.1, not Mythos. Treat Mythos as a specialist enable. Set effort on purpose, keep history append-only, and you get the long-run agent gains without torching Max quota or API spend.

Quick background (what just shipped)

Anthropic shipped both on September 1, 2026. Shared kit: 1M-token context, up to 128k output, always-on adaptive thinking. Depth runs through effort – low, medium, high, xhigh, max. List price matches Fable 5 at $10/MTok input and $50/MTok output. Cache reads fell to $0.25/MTok from $1. Anthropic’s own estimate: ~25% cheaper on typical billed work, up to ~45% on heavy agent loops.

Fable 5.1 (claude-fable-5-1) is the general path – API, claude.ai, Claude Code. Mythos 5.1 (claude-mythos-5-1) is the same capability with looser cyber and life-sciences classifiers, limited to vetted Project Glasswing and Cyber / Life Sciences Verification Program orgs (US-first at launch, per the announcement).

Low or medium effort on 5.1 often matches or beats full Fable 5 – cheaper. Most people skip this on day one. Details sit in the official announcement.

Method A vs Method B: Fable 5.1 or Mythos 5.1?

Fable still refuses some cyber-adjacent prompts. Mythos exists so defenders and life scientists don’t hit that wall. No program access? Mythos isn’t on the menu. Fable’s classifiers already cut cyber false positives about 60% versus earlier versions (Anthropic’s figure).

Worth a pause: if your work never touches those blocked domains, the 5-point Terminal-Bench gap is academic. You’re paying for rails you won’t feel – and Fable is the model you can actually call.

Winner for this walkthrough: Claude Fable 5.1. Below is how to drive it.

Hands-on: Claude Fable 5.1 without wasting tokens

Two doors: Claude Code and the Messages API.

Claude Code path

1. Update the CLI: npm install -g @anthropic-ai/claude-code@latest.
2. Start on the model: claude --model claude-fable-5-1, or /model inside a session.
3. Set effort: /effort high (Code default). Drop to medium or low for mechanical refactors once quality holds on your repo.
4. Give outcomes, not step lists. Example: “You’re responsible for making the payment service handle idempotent retries across the three call sites. Keep the public API stable. Verify with the existing integration tests.”

On Max plans it pulls the regular weekly limit and burns faster than older coding defaults – watch the meter. Community threads from launch week already flagged multi-million-token surprises on small jobs.

API path (minimal call)

from anthropic import Anthropic
client = Anthropic()

response = client.messages.create(
 model="claude-fable-5-1",
 max_tokens=8192,
 output_config={"effort": "high"},
 messages=[{
 "role": "user",
 "content": "I'm shipping a multi-service migration for the payments team. They need a safe cutover plan they can execute in one afternoon. Write the plan and the exact verification queries."
 }]
)
print(response.content[0].text)

Thinking stays on. No off switch, no fixed budget_tokens. Effort is the depth knob – default high in the API and Claude Code, medium on claude.ai.

Pro tip: Medium effort often matches old Fable 5 quality with fewer thinking tokens. Run your own evals before parking everything on xhigh or max. Those levels chew output budget fast.

Quiet progress is the gotcha in long tool chains. Default narration is softer than Fable 5, so the agent can look frozen for minutes. Ask for brief status between tool calls, and nudge it to batch independent tools in one turn instead of serializing everything.

Patterns that work are in the Fable 5.1 prompting guide. Breaking changes and IDs live in the what’s-new docs.

Edge cases that burn money or break sessions

• Forced tools die.tool_choice: {"type": "any"} or a forced named tool returns 400 – thinking is always on, so forced tools skip reasoning and trash argument quality. Use auto + strict schemas, or name the tool in plain language.
• History must stay append-only. Edit system, tools, or an earlier message after a thinking block and later blocks invalidate (400, or a silent drop with some betas). Accounts after Aug 31, 2026 bind thinking to the conversation prefix. Inject reminders as turn-scoped system messages. Don’t rewrite the past.
• Full-file rewrites and agent explosions. Docs and launch-week Reddit threads agree: Fable 5.1 rewrites whole files more than Fable 5, and high/xhigh effort on vague tasks can spawn huge sub-agent trees. Users have seen 100+ agents and multi-million-token burns on five-file audits. Cap agents in CLAUDE.md or the prompt. Say “smallest targeted patch – do not rewrite the file.”
• Cache math flipped a bit. Reads at $0.25 make aggressive early compaction less urgent for cost. Compacting later can keep more context intelligence. Still warm the cache on long agent prefixes. Exact Max weekly burn multiplier for 5.1 vs 5 still isn’t published.

Safeguards still fire on some cyber/bio-adjacent work after the false-positive cuts. Keep an Opus-path fallback if your agent loop hits refusals and retries forever.

Is the extra capability worth the burn rate on every task? Only your evals answer that.

FAQ

Can I use Claude Mythos 5.1 today?

No – not unless you’re already in Project Glasswing or the trusted verification programs. No public signup. Use Fable 5.1; same core model.

Does Fable 5.1 work on the free Claude tier?

As of the September 2026 launch material, Fable 5.1 is aimed at paid API and paid Claude surfaces (claude.ai plans, Claude Code), not a free-tier default. Plan for token burn like premium coding compute; community metering complaints showed Max weekly limits disappearing fast on high-effort agent runs. Check your plan’s live meter – published multipliers vs Fable 5 still lag.

I switched mid-chat from Opus and lost reasoning continuity. Why?

Thinking blocks are model-bound. Earlier models cannot read Fable 5.1 thinking blocks. Leave a long-horizon run on Fable, or accept a fresh think after you switch. And don’t edit the prefix once blocks exist – that binding error is a separate foot-gun from the model swap.

Open Claude Code or the API, set claude-fable-5-1 at medium or high effort, and throw it one real multi-file job you kept postponing. Measure tokens and quality against your old default. That’s the test that matters this week.