Claude Fable 5.2 Tested: Effort, Cost and How Mythos 5.2 Differs

Anthropic has released Claude Fable 5.2 and Claude Mythos 5.2, one underlying model offered two ways. Fable 5.2 is available to every customer, with cyber and biology safeguards, and Mythos 5.2 is for organisations Anthropic approves for security or life-sciences work.

Launch benchmarks show which model scores highest, but do not make it clear which Claude model is worth paying for on your workload.

We gave Fable 5.2 small tasks with built-in traps (a GPU job schedule, a SQL query over a messy database and a dense line chart), ran the schedule five times at each effort level and measured accuracy, cost and speed. We tested Fable 5.2 on the Claude API against Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5.

Tests ran in November 2026. We have included Anthropic’s own scores, how to choose an effort level and where cheaper Claude models are enough. We also cover speed, fast mode and the code changes that break older integrations.

One Model, Two Ways In

Prices fall from $10 and $50 on Fable 5.1, and Anthropic reports output up to 30% faster; its developer notes list the API changes.

Claude Fable 5.2 Benchmarks

In Anthropic’s scores, the largest gains over Fable 5.1 come in robotics, chart reading and knowledge work. The biggest jump is in simulated robotics, where RoboDojo Sim rises from 27.6% to 64.9%. Chart reading (Chartography) goes from 44.8% to 73.4%, and the Remote Labor Index from 17.9% to 41.3%. In the chart’s Coding and agents view, the biggest rise is Terminal-Bench-Science, from 52.6% to 74.6%. Fable 5.2 is also ahead of Opus 5.5 on all ten benchmarks shown:

What Fable 5.2 and Mythos 5.2 Built

Anthropic’s launch videos show the range:

Operating system

Electronics

Robotics

Mythos capsid

Videos: Anthropic.

Choosing an Effort Level

Thinking is always on, so effort (an API setting from low to max that controls how much the model reasons) sets depth, time and cost. To test it, we set a trap: 12 jobs to schedule on three GPU nodes, where job C needs all three nodes at once. Starting whatever is ready whenever nodes are free keeps job C waiting until hour 12, and the plan runs to hour 30. The best plan finishes at hour 21 by deliberately leaving nodes idle so job C can start at hour 2.

with client.messages.stream(
    model=model,
    max_tokens=max_tokens,
    thinking={"type": "adaptive", "display": "summarized"},
    output_config={"effort": effort, "format": {"type": "json_schema", "schema": puzzle.SCHEMA}},
    messages=[{"role": "user", "content": puzzle.prompt_text()}],
) as stream:

low     correct=True   finish=21h  output= 1,445 tokens    22.4s  $0.065
medium  correct=False  finish=22h  output= 1,470 tokens    21.9s  $0.066
high    correct=True   finish=21h  output= 1,724 tokens    23.8s  $0.076
xhigh   correct=True   finish=21h  output= 2,400 tokens    32.6s  $0.103
max     correct=True   finish=21h  output= 3,770 tokens    48.2s  $0.158

Over five runs per effort level, xhigh and max found the 21-hour schedule every time, and every miss was 22 hours. Low and high found it in four runs out of five, but medium found it only once, so more effort did not always mean a better answer. With five runs per level, treat that dip as a warning rather than a rule.

Cost rises with every level, from $0.068 a run at low to $0.176 at max, 81% more than xhigh for the same result. Xhigh is the cheapest level that never missed; low suits jobs where you can check the answer and retry.

Claude Fable 5.2 vs Cheaper Claude Models

On short, well-defined work, cheaper models kept pace: Opus 5.5, Sonnet 5.5 and Haiku 5.5 solved the trap in all three runs at high effort, and all five models found the exact revenue in a trap-laden SQL database. Haiku 5.5 did both at about 1/26 to 1/80 of Fable 5.2’s cost.

We also asked Fable 5.2, Fable 5.1 and Opus 5.5 to read every value from a dense line chart with 48 closely spaced points, in one run each. Fable 5.2 read all 48 exactly, against 47 for Opus 5.5 and 5 for Fable 5.1, though none of Fable 5.1’s readings was more than 3 GPU hours out.

Claude Fable 5.2 Speed and Fast Mode

On one prompt at low effort, Fable 5.2 averaged 78 output tokens per second against 57 for Fable 5.1, 36% faster, above Anthropic’s figure of up to 30%. Opus 5.5 generated tokens faster still, at 86 per second, and cost about half as much per answer: $0.026 against $0.052. Fast mode takes one parameter and reaches 220 tokens per second, finishing in about a third of the time (5.4 seconds against 16.5 in standard mode) for 1.8 times the cost per answer:

extra = {"speed": "fast"} if speed == "fast" else {}  # fast mode needs no beta header on Claude Fable 5.2
# ...
with client.messages.stream(
    model=model,
    max_tokens=4000,
    output_config={"effort": "low"},
    messages=[{"role": "user", "content": PROMPT}],
    **extra,
) as stream:

claude-fable-5-2  standard  1119 tokens  first text  1.72s  total  14.74s    75.9 tokens/s  $0.0452
claude-fable-5-2  fast      1191 tokens  first text  0.84s  total   5.39s   220.9 tokens/s  $0.0962
claude-fable-5-1  standard  1127 tokens  first text  0.62s  total  19.41s    58.1 tokens/s  $0.0569
claude-opus-5-5   standard  1301 tokens  first text  1.56s  total  15.11s    86.1 tokens/s  $0.0262

Claude Fable 5.2 API: Getting Started

The Models API shows which models a key can call:

import anthropic

client = anthropic.Anthropic()  # reads ANTHROPIC_API_KEY

available = {m.id: m for m in client.models.list()}
for model_id in ["claude-fable-5-2", "claude-mythos-5-2", "claude-fable-5-1", "claude-opus-5-5"]:
    print(f"{model_id:20} {'available' if model_id in available else 'not available to this key'}")

fable = client.models.retrieve("claude-fable-5-2")
caps = fable.capabilities
print("\n" + fable.display_name)
print("context window   :", f"{fable.max_input_tokens:,} tokens")
print("max output       :", f"{fable.max_tokens:,} tokens")
# ...
print("thinking types   :", [t for t, v in caps.thinking.types.to_dict().items() if v["supported"]])
print("effort levels    :", [e for e in ("low", "medium", "high", "xhigh", "max") if getattr(caps.effort, e).supported])

claude-fable-5-2     available
claude-mythos-5-2    not available to this key
claude-fable-5-1     available
claude-opus-5-5      available

Claude Fable 5.2
context window   : 1,000,000 tokens
max output       : 128,000 tokens
thinking types   : ['adaptive']
effort levels    : ['low', 'medium', 'high', 'xhigh', 'max']

Mythos 5.2 accepts the same requests with the model ID claude-mythos-5-2; access runs through the Cyber Verification Program and the Life Sciences Verification Program.

What Changes in Your Code

Code written for earlier Claude models can break on four rules from Anthropic’s migration and prompting guides:

response = client.beta.messages.create(
    model="claude-fable-5-2",
    max_tokens=8000,
    output_config={"effort": effort},
    # ...
    fallbacks="default",  # retry a declined request on the model Anthropic recommends for its category
    betas=["server-side-fallback-2026-07-01"],
    messages=[{"role": "user", "content": prompt}],
)
# ...
if response.stop_reason == "refusal":  # check this before reading content
    print(f"declined: category={response.stop_details.category} | {response.stop_details.explanation}")

Running the Capsid Folding Models on Hyperstack

Anthropic’s gene-therapy capsid example (the Mythos capsid clip) shows how the work splits. Mythos 5.2 wrote the method, and for the AAV2 capsid a single NVIDIA H200 GPU ran the open protein-folding models OpenFold3-preview2 and ESMFold2:

Direct access to Mythos 5.2 is limited to approved organisations, but those folding models are open, so anyone can run them. NVIDIA H200 SXM is available on demand on Hyperstack in eight-GPU virtual machines, billed per minute. AI Studio also serves open models.

Film: Anthropic.

Hyperstack GPU Cloud

Run the GPU Side on Hyperstack

Run folding pipelines and open models on NVIDIA H200 and NVIDIA H100 GPUs.

NVIDIA H200 SXMAI StudioPer-minute billing

Get started with Hyperstack in minutes.

FAQs

How do Claude Fable 5.2 and Claude Mythos 5.2 differ?

One model, one price: Fable 5.2 is open to everyone, with safeguards, and direct access to Mythos 5.2 goes to approved organisations.

How much does Claude Fable 5.2 cost?

$8 per million input tokens and $40 per million output tokens; batch requests cost half, and fast mode $16 and $80.

How do I access Claude Mythos 5.2?

Apply to Anthropic’s Cyber or Life Sciences Verification Program; approved organisations use the model ID claude-mythos-5-2. Claude Enterprise customers can also use it inside Claude Security.

Is Claude Fable 5.2 better than Claude Opus 5.5?

Anthropic’s scores put it ahead on all ten charted benchmarks; in our tests, Opus 5.5 matched it on scheduling and SQL for less and generated output faster.