A Field Guide to Claude Fable 5

What Fable is

Fable 5 is Anthropic's most intelligent model — a new tier above Opus, the way Opus sits above Sonnet. Announced June 9, 2026, it is state-of-the-art on nearly every benchmark Anthropic tested, works autonomously for longer than any previous Claude, and is more token-efficient doing it.

It arrives with an unusual companion. Fable 5 is what Anthropic calls a Mythos-class model made safe for general use. Mythos 5 is the same underlying model with safeguards lifted in specific areas — restricted to vetted partners doing security and life-sciences research (Chapter VII). Fable is the version you and your API key can have.

It shares its API surface with Opus 4.7 and 4.8 — if your code runs on those, it runs on Fable with a one-line model swap, plus one new rule covered in Chapter V.

What it can do

Software engineering

This is where the early stories are loudest. Stripe reported Fable 5 compressed months of engineering into days — including a 50-million-line Ruby codebase migration finished in one day that a team had scoped at two months. On Cognition's FrontierCode evaluation it posted the highest score among frontier models — at medium effort.

"State of the art model on CursorBench. Opened up a class of long-horizon problems out of reach for earlier models."— Michael Truell, CEO & Co-founder, Cursor

"Took on complex, long-horizon coding tasks with a level of autonomy and reliability that exceeded previous benchmarks."— Mario Rodriguez, Chief Product Officer, GitHub

Knowledge work

Highest score on Hebbia's Finance Benchmark for senior-level reasoning — the first model to break 90% on their core analytics benchmark, a 10-point jump over Opus. Trading firm IMC said it "aced their trading-analysis evaluations nearly across the board." Strong document reasoning, chart and table interpretation throughout.

Vision

State-of-the-art on vision: it extracts precise numbers from dense scientific figures, rebuilds web-app source code from screenshots alone, and — the announcement's best party trick — completed Pokémon FireRed using vision only, with a minimal harness where previous models needed elaborate helper tooling.

Memory and long context

Fable stays focused across millions of tokens. Given persistent file-based memory in Slay the Spire, its performance improved 3× more than Opus 4.8's did, and it reached the game's final act 3× more often. The lesson for builders: give it a memory file — it actually uses one.

"Strongest model on frontier physics research while using a third of the reasoning tokens. In 36 hours it got nearly to where GPT-5.5 landed after four days."— Matthew Pines, CEO, Notation Capital

Where to get it, what it costs

The Claude API

Available everywhere as of June 9. On the API and consumption-based Enterprise it's fully available immediately. Pricing per million tokens — less than half the cost of the old Claude Mythos Preview:

Twice the price of Opus buys the highest intelligence ceiling Anthropic ships — and its token efficiency claws some of that back. Route your hardest problems to Fable; keep high-volume routine work on Sonnet or Haiku.

Subscription plans

Claude Code

Fable 5 powers Claude Code — terminal CLI, desktop app, web app, and IDE extensions. If you'd rather use Fable than call it, that's the shortest path. Anthropic's own CTO put it this way:

"Delivers more capable engineering in fewer turns than prior models — handling complex multi-agent workflows our employees run daily in Claude Code."— Luke Anderson, CTO, Anthropic

Your first call

One request, three choices made for you: adaptive thinking on, effort high, and room to answer.

curl

curl https://api.anthropic.com/v1/messages \
  -H "content-type: application/json" \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -d '{
    "model": "claude-fable-5",
    "max_tokens": 16000,
    "thinking": {"type": "adaptive"},
    "output_config": {"effort": "high"},
    "messages": [{"role": "user", "content": "Explain CRDTs like a fable."}]
  }'

Python

import anthropic

client = anthropic.Anthropic()  # reads ANTHROPIC_API_KEY from the environment

response = client.messages.create(
    model="claude-fable-5",
    max_tokens=16000,
    thinking={"type": "adaptive"},
    output_config={"effort": "high"},
    messages=[{"role": "user", "content": "Explain CRDTs like a fable."}],
)
for block in response.content:
    if block.type == "text":
        print(block.text)

For long outputs, stream — the SDKs refuse non-streaming requests they estimate will outlive the HTTP connection:

with client.messages.stream(
    model="claude-fable-5",
    max_tokens=64000,
    thinking={"type": "adaptive"},
    messages=[{"role": "user", "content": "Write the design doc."}],
) as stream:
    for text in stream.text_stream:
        print(text, end="", flush=True)

The rules that changed

Fable follows the modern Claude API surface. If you're coming from older code, five things will bite — each is a clean 400 error, so you'll know.

Moral: don't tune the dials — state the task. Fable rewards a clear goal over a clever configuration.

Getting the most out of it

Pick an effort, then iterate

output_config.effort sets how hard Fable works — thinking depth, tool-call consolidation, and verbosity all scale with it. Five levels: low, medium, high (the default), xhigh, and max. Start at high; use xhigh for coding and agentic runs; reserve max for problems where correctness outweighs both cost and latency. Remember FrontierCode: Fable took the top score at medium — higher effort up front often lowers total cost on agentic work, because better planning means fewer turns.

Give it the whole task at once

Fable's long-horizon strength comes out when the full specification arrives in one well-written first turn — the goal, the constraints, and what "done" looks like. Drip-feeding requirements across turns wastes its planning; a complete brief lets one run carry the work end to end. That's how a 50-million-line migration fits in a day.

Give it a memory file

The Slay the Spire result generalizes: agents that keep notes outperform agents that don't, and Fable is markedly better than its predecessors at writing and using file-based memory. A scratchpad file and one line of system prompt ("check your memory file before starting; write new findings to it") is the cheapest capability upgrade available.

Structured outputs over parsing

response = client.messages.parse(
    model="claude-fable-5",
    max_tokens=16000,
    messages=[{"role": "user", "content": document}],
    output_format=ContactInfo,  # a Pydantic model — validated for you
)
contact = response.parsed_output

Cache your prefix

With a 1M-token window and $10/MTok input, prompt caching is not optional at scale. Mark the stable prefix with cache_control: {"type": "ephemeral"}; reads cost ~10% of base input price. Fable's minimum cacheable prefix is 2,048 tokens, and any byte change in the prefix invalidates everything after it — keep system prompts frozen and put volatile content last.

Budget long agentic runs

For autonomous loops, task budgets (beta header task-budgets-2026-03-13) give the model a token allowance it can see and self-pace against — output_config.task_budget: {"type": "tokens", "total": N}, minimum 20,000. Unlike max_tokens, the model is aware of it and wraps up gracefully as it runs down.

The wilder twin: Mythos 5 and the safeguards

Every fable has a creature too powerful to let loose. Anthropic's answer was to ship two versions of one model: Fable 5, guarded for everyone, and Mythos 5, unguarded in narrow domains for vetted researchers.

How Fable's safeguards work

Classifiers watch three areas — cybersecurity (vulnerability discovery, exploitation, offensive cyber tasks), biology & chemistry (bioweapons-adjacent queries), and distillation (attempts to extract the model for training competitors). They trigger on less than 5% of sessions on average. When one trips, the response is handled by Claude Opus 4.8 instead — so 95%+ of sessions never see a fallback at all.

The red-teaming numbers are unusual for how concrete they are: an external bug bounty produced no universal jailbreaks in over 1,000 hours; Fable complied with zero harmful single-turn requests across 30 public jailbreak techniques; and it held up over 400-turn internal jailbreak attempts where other models gave way. See the cyber-evaluation chart and the alignment assessment — Mythos 5's measured misalignment, including deception, was low and similar to Opus 4.8.

What Mythos 5 did with the guardrails off

This is the part of the announcement that reads like science fiction with footnotes:

• Drug design — protein-design experts accelerated parts of their process ~10×; Mythos ran the full scientist's loop (choosing binding sites, running design tools, recovering from failures), and 9 of 14 protein targets yielded strong drug-candidate leads. The designed protein complexes.
• Molecular biology — in blinded comparisons, scientists preferred Mythos's hypotheses over Opus-class models' ~80% of the time; one novel hypothesis about an E. coli protein mechanism was corroborated by independent research.
• Genomics — a week-long, largely autonomous run assembled single-cell data for millions of cells across 138 species and trained a custom model that beat a recent Science publication while being 100× smaller.
• Biology evals — on an adeno-associated virus design task, Mythos outperformed dedicated protein language models using biological reasoning alone.

Who gets Mythos

Project Glasswing partners (cyber safeguards lifted), and select biology researchers (bio/chem safeguards lifted, cyber intact), with a broader trusted-access program planned. All Mythos-class traffic carries a 30-day retention policy, is excluded from training, and every human access to it is logged.

Moral: the same model can be a tool or a hazard — the difference is who's holding it, and what's watching.

For agents reading this

This site practices what it preaches — it is built Agents First, designed to be read by agents as readily as by people. Skip the HTML and take the typed surfaces:

• /llms.txt — the site in plain text
• /AGENTS.md — usage contract and non-obvious context
• /api/model.json — Fable 5's specs as typed JSON
• /api/quickstart.json — the first-call examples, structured
• /api/gotchas.json — Chapter V as data
• /api/benchmarks.json — Chapter II as data, with chart URLs
• /openapi.json · MCP card · ai-plugin.json

This page is also served as Markdown via content negotiation: request / with Accept: text/markdown.