Claude Fable 5.2 vs Mythos 5: We Ran Both

Two Claude model IDs sit next to each other in the router, both with a mythic name, both from Anthropic, and the release notes make them sound like the same thing wearing different badges. So which one do you actually call? That's the real question behind Claude Fable 5.2 vs Mythos 5, and reading spec sheets won't settle it. Running them will.

So we did. Same three prompts, same settings, both models, through the each::labs LLM router. One coding task, one scheduling puzzle with a trap in it, one creative brief that ends in an image prompt. Below you'll find what Anthropic says about each model, what each one returned, and a plain verdict on when to pick which.

Two Names, One Family Tree

Start with the lineage, because it explains almost everything. Anthropic announced Claude Mythos 5 on June 9, 2026, together with Claude Fable 5. Anthropic describes Mythos 5 as the update to Mythos Preview, with gains on cybersecurity, biology and healthcare benchmarks, and it launched for a small group of vetted partners. Reporting from The Batch at the time described the two June models as identical except that Fable won't take on prompts in sensitive areas like security, biology and chemistry.

Then the line moved. Mythos 5.1 arrived on September 1. On November 9, Anthropic released Fable 5.2 and Mythos 5.2, and the official announcement says it outright: Claude Fable and Claude Mythos are the same underlying model. Fable carries strong cybersecurity and biology safeguards, which is why Anthropic can offer it to everyone. Mythos ships without some of those safeguards and stays limited to vetted organizations.

Here's the part that matters for this comparison. The Mythos you can call on each::labs is anthropic/claude-mythos-5, the June generation. The Fable you can call is anthropic/claude-fable-5.2, the November one. So this isn't a test of safeguards on the same brain. It's five months of model progress on one side of the table.

What Anthropic Changed in Fable 5.2

Anthropic positions Fable 5.2 as its strongest generally available model for hard technical work: long migrations, multi-step agents, research problems. According to the announcement, outputs come up to 30% faster than Fable 5.1, and the model triggers about 55% fewer safeguard blocks than its predecessor. On FrontierSWE v2 it scores 69.9%, ahead of Claude Opus 5.5 at 62.3%.

One number in the release is a nice sanity check for the "same model" claim. On Terminal-Bench-Science 0.1, Fable 5.2 scores 74.6% and Mythos 5.2 scores 75.3%. Less than a point apart. When the safeguards don't come into play, the two 5.2 versions behave like twins.

Anthropic doesn't publish Mythos 5 scores on that page, so we won't pretend to have them. What we can show is how the June Mythos and the November Fable handle the same real prompts today. We covered the full Fable benchmark table in our Claude Fable 5.2 launch post, if you want every row.

How We Ran Both

Nothing fancy. Each prompt went to both models through the router's chat completions endpoint with an output cap of 1,200 tokens, run in parallel, no system prompt, no retries. We timed each call end to end, so the seconds you'll see include the network round trip. It's one run per prompt, which makes the timings a hint, not a benchmark. The content of the answers is the part worth reading.

Round One: The Code Task

The first prompt was the kind of small utility you'd ask for twenty times a week:

Write a Python function dedupe_by_key(items, key) that removes duplicate dicts from a list based on a key, keeps the first occurrence, preserves order, and skips items that are missing the key. Include a short docstring and two asserts as tests. Code only.

Both models wrote essentially the same function: a seen set, a single pass, a continue for items without the key. Both added two asserts that cover duplicates and missing keys, and both respected "code only" without a word of commentary. Fable 5.2 came back in 4.6 seconds, Mythos 5 in 5.2.

Call it a draw, and a useful one. For well-specified, everyday code, you won't feel the generation gap. That's exactly why routine tasks are a poor way to choose between frontier models. Every serious model passes them.

Round Two: The Logic Trap

The second prompt hid a bottleneck. Three GPUs render 20 clips at 12 minutes each, and one reviewer checks each finished clip in 4 minutes, one at a time. What's the minimum time until everything is rendered and reviewed?

Both models landed on the right answer, 92 minutes. The way they got there is where they split. Fable 5.2 built a tight argument from the rendering side: at most 18 clips can be done before minute 84, so the last two clips can't be reviewed before 84 plus 8. Then it showed the schedule that hits that bound. Three short paragraphs, done in 9.0 seconds.

Mythos 5 reached 92 from the reviewer's side (the first clip lands at minute 12, plus 80 minutes of review work) and then walked through the schedule batch by batch. Solid reasoning, but longer, and it closed with a confused aside about a "last render" finishing at minute 88 that doesn't match its own timeline. It took 15.4 seconds. Fable's answer is the one you'd paste into a planning doc without editing.

Round Three: The Creative Brief

The last prompt is the one closest to how teams use an LLM next to image and video models:

Turn this brief into a 4-shot list for a 15-second product video, one line per shot with camera move and duration: 'A matte black travel mug for commuters. Mood: calm early morning. Must show the leak-proof lid and that it fits a car cup holder.' Then write one image-generation prompt for shot 1.

Here the difference was clear, and it wasn't about speed (Mythos 5 actually answered faster, 8.2 seconds against 12.4). Fable 5.2 treated "must show the leak-proof lid" as a requirement to prove, not just mention. It gave the lid click its own macro shot, then added a third shot where the mug slips sideways into a tote bag next to a laptop, holding on the lid. That's a visual argument a viewer understands in a second.

Mythos 5 covered both must-haves in fewer shots, with a sideways tilt during the lid close-up. Fine. Its first shot, though, has steam curling from the closed lid, which quietly undercuts the leak-proof story you're trying to tell. The image prompts split the same way. Mythos wrote a tidy comma-separated list of tags. Fable wrote a full scene: pale stone counter, cool pre-dawn light with a warm edge on the horizon, a 50mm lens at counter height, soft highlights along the matte finish, no logos or text. Hand that to Seedream 5.0 Pro and you're iterating on taste, not fixing gaps.

Claude Fable 5.2 vs Mythos 5: The Verdict

Default to Fable 5.2. On the easy task the two tied. On the reasoning task, Fable gave the cleaner proof with no stray mistakes. On the creative brief, Fable read the requirements more carefully and wrote a far more usable prompt. Three prompts don't make a leaderboard, but every gap we saw pointed the same direction, and it's the direction you'd expect from a model five months newer that Anthropic calls its strongest for hard work.

So where does Mythos 5 fit? If you already have prompts, evals or agents tuned on Mythos 5, keep it as your baseline and run the same suite on Fable 5.2 before you switch production traffic. A second opinion takes one extra line of code when both models share one endpoint, and a disagreement between two strong models is often the most useful signal in a review step. Mythos is also the line Anthropic builds without some cyber and biology safeguards, which is why access to it is gated in the first place. Our three tests didn't touch those areas, so we're not drawing conclusions there.

One more thing the test made obvious: the gap shows up in judgment, not in boilerplate. If your workload is mostly short, well-specified requests, either model will do. If it involves briefs, plans and anything ambiguous, Fable 5.2 earns its place.

Call Both With One Key

The router speaks the OpenAI chat completions format, so the official OpenAI SDK works as is. Point it at the each::labs base URL, then loop over both model IDs to run your own side-by-side:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.eachlabs.ai/v1",
    api_key="YOUR_EACHLABS_API_KEY",
)

prompt = "Turn this brief into a 4-shot list for a 15-second product video..."

for model in ["anthropic/claude-fable-5.2", "anthropic/claude-mythos-5"]:
    response = client.chat.completions.create(
        model=model,
        messages=[{"role": "user", "content": prompt}],
    )
    print(f"--- {model}")
    print(response.choices[0].message.content)

Swap in twenty of your own real prompts and you'll have a better answer than any chart can give you. The full list of model IDs is in the each::labs LLM router docs.

From Shot List to Finished Video

The creative round hints at the workflow that makes a strong LLM worth having. Let Fable 5.2 turn a brief into a shot list and a consistent set of prompts, then send those prompts to image models for stills and to Kling 4.0 or other text-to-video models for motion. Same key, same platform, and you can chain the whole thing in each::labs flows. If you like head-to-head tests like this one, we ran the same treatment on video in Seedance 2.5 vs Kling 4.0 and on images in Seedream 5.5 vs GPT Image 2.6.

Frequently Asked Questions

Is Claude Fable 5.2 better than Mythos 5?

In our three-prompt test, yes, wherever there was a difference. The code task was a tie. Fable 5.2 gave the cleaner reasoning on the logic puzzle and the more careful, more usable answer on the creative brief. It's also five months newer, released by Anthropic on November 9, 2026.

How is Mythos 5 different from Mythos 5.2?

Generation and access. Anthropic says Fable 5.2 and Mythos 5.2 are the same underlying model with different safeguards, and Mythos 5.2 is limited to vetted organizations. Mythos 5 is the earlier Mythos release from June 2026, two versions back. On each::labs the Mythos model ID is anthropic/claude-mythos-5.

Can I call both models on each::labs?

You can. Both sit behind the same LLM router endpoint: anthropic/claude-fable-5.2 and anthropic/claude-mythos-5. Switching between them is a one-string change in your code, which makes side-by-side evals easy.

Which one should I use for agents and long tasks?

Start with Fable 5.2. Anthropic built it for long migrations, multi-step agents and research work, and on each::labs it comes with a one million token context window. Keep Mythos 5 in the loop as a comparison baseline if your existing pipeline was tuned on it.