Claude Opus 5 vs Claude Fable 5: Which Anthropic Model Should You Use?

What Is Claude Opus 5?

Claude Opus 5 is Anthropic's newest model in the Opus tier, released on July 24, 2026, and priced identically to its predecessor Opus 4.8.

Anthropic describes it as a thoughtful and proactive model built for daily use, and it's now the default model on Claude Max and the strongest model on Claude Pro. Its signature trait is agentic persistence: it verifies its own work and iterates until a task actually succeeds rather than stopping at a plausible-looking answer.

To see how it behaves in practice, check out our Claude Opus 5 API tutorial, where we build and benchmark a bug-fixing coding agent across all five reasoning-effort levels. You might also want our guide to Claude Code templates if you plan to run Opus 5 inside an agentic workflow.

What Is Claude Fable 5?

Claude Fable 5 is a Mythos-class model that Anthropic made safe for general use, released on June 9, 2026, at $10 per million input tokens and $50 per million output tokens. It was briefly pulled on June 12, 2026, to comply with a U.S. export-control order, and Anthropic restored access on July 1, 2026, once the order was lifted.

Anthropic positions it above the Opus tier in raw capability, and at launch, it was state-of-the-art on nearly all tested benchmarks, with its lead growing the longer and more complex the task. Its distinctive feature is a set of classifiers that route flagged cybersecurity, biology, and distillation requests to Opus 4.8 instead of answering directly.

Claude Opus 5 vs Claude Fable 5: Head-to-Head Comparison

Here's the summary before I dig into each dimension. The short version: Opus 5 and Fable 5 land within noise of each other on most benchmarks, but Opus 5 does it at half the price.

* current overall leader on this benchmark

Coding and agentic workflows

Coding is where most people will actually deploy these models, and it's also where the pricing gap looks least justified. On SWE-bench Verified, Opus 5 scores 96.0% against Fable 5's 95.0%, a lead of one point for the cheaper model. On Frontier-Bench v0.1, the terminal-coding eval, Opus 5 pulls further ahead at 43.3% versus 33.7%.

Fable 5 does claw back a narrow win on SWE-bench Pro, scoring 80.3% to Opus 5's 79.2%. That's a one-point gap in the other direction, which tells you these two models are effectively tied on repository-level engineering. Cognition's team reported that Opus 5 approaches Fable-level performance at half the cost on their FrontierCode 1.1 eval, which matches what the public numbers show.

The takeaway is that Opus 5 either matches or beats Fable 5 on the coding benchmarks, where both published numbers, with the single exception of SWE-bench Pro. Anthropic's own framing is that Opus 5 more than doubled Opus 4.8's Frontier-Bench score at a lower cost per task, and it beats Fable 5 there too. For coding, the pricier model gives you nothing extra.

Agentic and computer-use tasks

Both models are built for long-horizon agentic work, so this dimension matters if you're running autonomous pipelines rather than one-shot prompts. Here, Opus 5's proactive verification behavior is the differentiator that benchmarks capture partly.

On ARC-AGI 3, an eval where the model has to solve novel problems, Opus 5 jumps from Opus 4.8's 1.5% to 30.2%, which is almost four times as high as the next-best model (GPT-5.6 Sol at 7.8%). Sadly, Anthropic did not publish any score for Fable 5, so we cannot compare the two models directly.

On OSWorld 2.0, the computer-use benchmark, Opus 5 (70.6%) replaces Fable 5 (66.1%) as the overall leading model, and on Zapier's AutomationBench, it hit a 100% pass rate on a churn-prevention sequence that previous models couldn't complete at all.

Opus 5 wins agentic tasks on a per-dollar basis by a wide margin.

Reasoning and knowledge work

Knowledge work is Fable 5's marketing home turf, so I expected it to lead here, but it doesn't. On GDPval-AA v2, the knowledge-work Elo benchmark, Opus 5 scores 1,861 against Fable 5's 1,747. That's a 114-point Elo gap in favor of the cheaper model. 

Fable 5 still holds a genuine edge on the very hardest, longest scientific reasoning, and Anthropic is explicit that Mythos-class capability remains stronger for long-running autonomous biology research. But for the analytical and knowledge tasks most teams run, Opus 5's higher GDPval-AA score means you're paying double for Fable 5 to do worse.

Safeguards and refusal behavior

This is the dimension where the two models genuinely diverge, and it's the real reason Fable 5 exists as a separate product. It is also the dimension most likely to be read wrong, because two different things get called "safety" here, and they measure opposite ends of the problem.

The first is misuse resistance: an external classifier system that sits in front of the model and reroutes dangerous requests before they reach it. The second is alignment: how the model itself behaves, including whether it deceives, takes reckless, irreversible actions, or can be talked into misuse.

A model can score well on one and still need heavy support on the other, because the classifiers are not guarding against the model going rogue. They guard against a user pulling something harmful out of it.

Fable 5 is a Mythos-class model with the heavier classifier system of the two. Its classifiers cover three areas, and when they fire, the request falls back to Opus 4.8:

• Cybersecurity: both exploitation and broader offensive cyber tasks are blocked, flatlining Fable 5's performance to 0% on cyber evals when classifiers are active.
• Biology and chemistry: most biology and chemistry requests fall back to Opus 4.8 by default.
• Distillation: requests flagged as attempts to extract Fable 5's capabilities route away.

Anthropic tuned these conservatively: early data shows more than 95% of Fable 5 sessions trigger no fallback at all.

Opus 5 carries a lighter version of the same system. Its cyber classifiers allow source-code vulnerability finding but block binary-based scanning, penetration testing, and exploit generation, and Anthropic expects them to fire around 85% less often than Fable 5's.

On its automated behavioral audit, Opus 5 scored 2.3 on misaligned behavior, which Anthropic calls its most aligned model to date, ahead of Opus 4.8, Sonnet 5, and Fable 5.

So this is a fork, not a scoreboard. For general coding, knowledge work, and computer use, Opus 5 is a well-aligned model that gets in the way less. For the frontier of a guarded domain, such as offensive security research or long-horizon biology, Fable 5 (or Mythos 5 through the Cyber Verification Program) is the more capable tool, and the heavier refusal behavior is the price of reaching it.

Pricing: what you actually pay

Fable 5 costs exactly twice as much as Opus 5 on both input and output. Opus 5's Fast mode, which runs at roughly 2.5x default speed, costs $10 input and $50 output, which lands it at exactly Fable 5's standard rate.

What a real workload costs

Both models share Anthropic's newer tokenizer that arrived with the Opus 4.7 generation, so identical text costs the same number of input tokens on each, and neither vendor has published a controlled token-count comparison between the two for the same task.

At the heavy level, that's a $1,125 monthly difference for a production feature or a continuously running agent, and the benchmarks say you're giving up nothing on coding or knowledge work to capture it.

That said, treat these totals as a rate comparison rather than a precise forecast. Opus 5 can cost more than Opus 4.8 in practice because it tends to emit more thinking tokens, but it doesn't change the general picture that Opus 5 will be the significantly cheaper model of the two in most cases.

When to Choose Claude Opus 5 vs Claude Fable 5

For most readers this is a short decision: Opus 5 is the default, and Fable 5 is the exception. The table below covers the scenarios where that flips.

• You're building coding or agentic tools and want the best per-dollar performance, since it beats Fable 5 on Frontier-Bench and ties it on SWE-bench.
• You run high-volume workloads where a 50% rate cut translates into real budget, and the benchmarks confirm you lose nothing on quality.
• You want the most aligned model available, given its 2.3 misaligned-behavior audit score, Anthropic's best to date.
• You need lighter refusal behavior for legitimate security research, since its classifiers fire roughly 85% less often than Fable 5's.

Choose Fable 5 if...

• Your work is long-horizon autonomous biology research, the one area where Anthropic says Mythos-class capability still leads.
• You specifically need the broader classifier coverage across cyber, biology, and distillation that Fable 5's safeguard system provides.
  

How to Get Started With Claude Opus 5 and Claude Fable 5

Both models are broadly available across the Claude apps, the API, the major cloud platforms, and Anthropic's coding agents. The practical differences are narrow and mostly show up at the consumer plan level.

On the API, both are a single string swap, using claude-opus-5 and claude-fable-5, respectively.

The clearest gap: on Pro and Team Standard, Fable 5 isn't included at all. It runs on pay-as-you-go usage credits at API rates, while Opus 5 is included on those plans.

Using Opus 5 and Fable 5 in a coding agent

Both models are selectable in Claude Code, and switching between them is a matter of changing the model string in your request or agent config. Because they share the same SDK, moving a working agent from one to the other is a one-line change.

from anthropic import Anthropic

client = Anthropic()
response = client.messages.create(
    model="claude-opus-5",  # swap for "claude-fable-5"
    max_tokens=1024,
    messages=[{"role": "user", "content": "Refactor this function..."}],
)
print(response.content.text)

One implementation detail from our hands-on testing: Opus 5 returns thinking blocks by default, so accessing response.content.text directly can raise an AttributeError when the first block is a thinking block. We cover the full setup, effort-level tuning, and prompt-caching behavior in our Claude Opus 5 API tutorial.

Final Thoughts

I recommend using Opus 5 unless you have a specific reason not to. It matches or beats Fable 5 on coding, wins the knowledge-work Elo benchmark, triples the field on novel reasoning, and does all of it for half the price while posting Anthropic's best alignment score.

Fable 5 clearly is not a bad model. It's just not made for most of what teams actually do. It carries heavier safeguards and holds a real edge on the frontier of long-running autonomous work like biology research, but outside that narrow slice, paying double for Fable 5 means paying double to score slightly lower on the benchmarks that matter for coding and analysis.

What I find most interesting is that this comparison inverts the usual model-tier logic. Normally, the newer, cheaper model trades some capability for accessibility. Here, Anthropic shipped a cheaper model that is also the more capable choice for nearly every mainstream task, which suggests Fable 5's price is buying you its safeguard posture and Mythos lineage more than raw performance.