Claude Fable 5 vs Opus 4.8: Should You Actually Switch? (Mythos-Class, Benchmarked)

Claude Fable 5 vs Opus 4.8: Should You Actually Switch? (Mythos-Class, Benchmarked)

Anthropic shipped Claude Fable 5 today, June 9, 2026. It scores 80.3% on SWE-Bench Pro, the agentic-coding benchmark where Opus 4.8 manages 69.2% and GPT-5.5 trails at 58.6%. That's not a rounding-error gap. And there's a twist: Fable 5 is the public, safeguarded half of a new "Mythos-class" model, with a gated sibling called Claude Mythos 5 kept behind a restricted access program. Here's what changed, what a real task costs at $10/$50 per million tokens, and whether you should actually move your stack off Opus 4.8.

Quick Answer

• Fable 5 scores 80.3% on SWE-Bench Pro vs Opus 4.8's 69.2% and GPT-5.5's 58.6%.
• Pricing is $10 input / $50 output per million tokens, less than half of Mythos Preview.
• Fable 5 auto-falls back to Opus 4.8 for high-risk cyber/bio/chem queries (under 5% of sessions).
• Free on Pro/Max/Team/Enterprise plans until June 22, 2026; usage credits apply from June 23.

What Actually Launched: Fable 5, Mythos 5, and the "Mythos-Class" Line

Mythos-class is Anthropic's new top capability tier, and it ships as two models. Claude Fable 5 is the public, safeguarded version you can call today. Claude Mythos 5 is the same frontier model with the safeguards removed, kept behind gated access. Same brains, two release postures.

The split is the whole story here. According to Anthropic's newsroom announcement, Mythos 5 retains full frontier capability, including offensive-cyber skills that could cause real damage, so Anthropic doesn't hand it to everyone. Fable 5 takes that same model and wraps it in classifiers that detect high-risk requests and quietly route them to Opus 4.8 instead.

So when you call Fable 5, you get Mythos-class reasoning on everything that isn't dangerous, and a safer fallback model on the small slice that is. CNBC framed it as Anthropic releasing a "Mythos-like AI to the public," which is accurate, but the precise mechanic matters more than the headline.

Fable 5 and Mythos 5 are the same frontier model split in two: one you can use today, one Anthropic keeps gated. If you only remember one thing about the naming, remember that. Mythos-class is the tier; Fable and Mythos are the two doors into it.

What Changed vs the Opus 4.x Line

On every coding and reasoning eval Anthropic published, Fable 5 clears Opus 4.8, not by a hair but by double-digit margins. SWE-Bench Pro jumps 69.2% to 80.3%. FrontierCode Diamond goes from 13.4 to 29.3. The pricing dropped too, and the model scales with a new reasoning-effort dial. This is a genuine frontier step over the Opus 4.x line, not a point release.

The 13-benchmark sweep above is the at-a-glance version. Here's the headline subset against the three models most teams actually compare:

The comparison anchor here is Claude Opus 4.8, the model Fable 5 replaces at the frontier, and the same model Fable 5 falls back to on high-risk queries. If you've tracked the Opus 4.x line across the last two releases, the pattern was incremental gains. Fable 5 breaks that pattern.

Two things to flag before you read the table as gospel. First, these are Anthropic's own published evals, not an independent test. Second, the reasoning-effort dial means a single benchmark number hides a cost curve. More on that in the pricing section, because it changes the math.

The Benchmarks That Matter for Builders

Four numbers carry the practical weight: SWE-Bench Pro 80.3%, FrontierCode Diamond 29.3, Terminal-Bench 2.1 88.0%, and GPQA-AA 1932 Elo. Together they cover real repo work, the hardest synthetic coding problems, agentic shell tasks, and graduate-level reasoning. If your workload touches any of those, this is the section that tells you whether the number maps to your job.

The hero chart at the top of this post is the proof: SWE-Bench Pro 80.3 vs 69.2 vs 58.6, and FrontierCode 29.3 vs 13.4 vs 5.7. Now what do those actually measure?

• SWE-Bench Pro tests whether a model can ship a real pull request to a real repo: clone, understand, patch, pass tests. SWE-Bench Pro measures whether a model can ship a real pull request to a real repo, and Fable 5 lands 8 in 10. This is the closest proxy to "can it do my actual job."
• Terminal-Bench 2.1 scores agentic shell tasks: running commands, reading output, recovering from errors. 88.0% matters if you're building AI coding agents that live in a terminal.
• FrontierCode Diamond is the hardest tier of coding problems, the ones where most models score single digits. 29.3 is low in absolute terms but more than double Opus 4.8's 13.4.
• GPQA-AA is graduate-level science reasoning, scored as an Elo rating. 1932 signals strong multi-step reasoning beyond code.

In our Claude Code workflows across the Opus 4.x line, the recurring ceiling was long-horizon agentic tasks: the model would lose the thread halfway through a multi-file change. Fable 5's Terminal-Bench and SWE-Bench Pro numbers suggest that ceiling moved. Whether it moved for your repo is the only test that counts, but the published numbers are the right kind of number to care about.

Real Test Cases: What Fable 5 Does That Prior Models Couldn't

Anthropic published three concrete examples that show the class of problem Fable 5 can now handle: a Stripe codebase migration compressed from months to a day, a video game completed on vision alone, and a 3× performance jump from self-validation. Each one points at a capability prior models couldn't reach. All three are Anthropic-sourced; we're citing their announcement, not a test of our own.

Here's what each one proves:

1. Stripe Ruby migration. Anthropic reports Fable 5 compressed a Ruby codebase migration from roughly two months to about one day. Stripe compressed a Ruby migration from two months to a single day, and that's the size of problem this model changes. This is the class of long-horizon refactor that used to need a human team holding the whole codebase in their heads.

2. Pokémon FireRed on vision alone. Per Anthropic, Fable 5 completed Pokémon FireRed using vision only, where prior models needed complex custom harnesses to feed game state in as text. That's a spatial and visual-reasoning leap: the model reads the screen and acts, no scaffolding.

3. Slay the Spire, 3×. At max reasoning effort, Anthropic says Fable 5 self-validates its own moves and reached 3× better performance via persistent memory. The takeaway isn't the game; it's that the model can check its own work over a long planning horizon and improve from it.

None of these are things you should take on faith for your workload. But they map cleanly to the benchmark jumps in the table above: long-horizon coding (SWE-Bench Pro), spatial reasoning (vision), and self-validation (the reasoning-effort dial). The qualitative wins and the quantitative ones tell the same story.

Pricing and the Cost-Per-Task Reality

Fable 5 costs $10 per million input tokens and $50 per million output tokens. Anthropic frames that as "less than half the price of Claude Mythos Preview." It's also roughly 2× Opus 4.8 per token. Both are true. The catch is that token price is the wrong unit. What you actually pay for is a finished task, and that's where Fable 5's math gets interesting.

Reasoning effort is a dial you set per request. Low effort means fewer internal reasoning tokens and a cheaper, faster answer; max effort means the model thinks longer, spends more output tokens, and scores higher. The chart above shows Fable 5 climbing from ~11% to ~31% on FrontierCode as effort scales low to max, while Opus 4.8 peaks around 13% and GPT-5.5 sits flat near 5-6%. So cost per task isn't one number; it's a curve you choose a point on.

Let's work a concrete example from the published $10/$50 pricing. This is arithmetic from Anthropic's posted rates, not a benchmarked result.

Take a single agentic call with 50,000 input tokens and 15,000 output tokens at high effort:

Input:  50,000 / 1,000,000 × $10 = $0.50
Output: 15,000 / 1,000,000 × $50 = $0.75
Per call:                          $1.25

A real agentic task chains many such calls: read repo, plan, edit, run tests, fix, repeat. Say the loop runs 12 calls at that shape: roughly $15 for the finished task. At max reasoning effort with heavier context, Anthropic's own cost curve puts hard FrontierCode tasks closer to the $20-per-task range. Run the same task on GPT-5.5 and you'll pay less per token. But if it can't finish the task at any effort level, your cost-per-finished-task is infinite. Fable 5 costs more per token than GPT-5.5, and still wins on cost-per-finished-task, because it completes work GPT-5.5 can't.

Here's a minimal call. Anthropic's announcement gives claude-fable-5 as the model id; verify the exact versioned string against the Anthropic API docs before you ship, since dated aliases sometimes differ:

import anthropic

client = anthropic.Anthropic()

message = client.messages.create(
    model="claude-fable-5",          # verify exact dated id in API docs
    max_tokens=4096,
    thinking={"type": "enabled", "budget_tokens": 8000},  # reasoning effort
    messages=[
        {"role": "user", "content": "Migrate this module from Ruby 2.7 to 3.3 and run the tests."}
    ],
)

print(message.content)

The economics thesis in one line: you pay more per token than GPT-5.5, but you finish tasks GPT-5.5 can't finish at any price. For low-stakes high-volume work, that math favors a cheaper model. For hard agentic coding, it favors Fable 5.

The Safeguards Split: Why Fable Falls Back to Opus 4.8

Fable 5 runs classifiers on every request. When it detects a high-risk query (offensive cyber, bio, chem, or model-distillation attempts), it falls back to Opus 4.8 instead of answering with full Mythos-class capability. Anthropic says this fallback fires on under 5% of sessions, so most users never trigger it. The point is to keep the dangerous capability behind the gate while leaving everyday work untouched.

The harder number is attack-success rate: how often an automated attacker can jailbreak the model into doing something it shouldn't. Lower is better. Anthropic's red-team results show a steep drop across releases, from Opus 4.6 at 83.2%, to Opus 4.7 at 72.7%, to Opus 4.8 at 56.6%, down to Fable 5 at 5.4%. Under automated red-teaming, attacks succeeded against Fable 5 just 5.4% of the time, down from 56.6% on Opus 4.8.

Why should you care? If you deploy an agent with tool access (shell, file system, network), a jailbreak isn't a bad chat reply, it's a model executing real commands an attacker steered it into. A 5.4% attack-success rate means it's roughly ten times harder to weaponize than Opus 4.8. IT Pro's coverage led with the fallback mechanic for exactly this reason: it's the feature that makes a frontier model safe enough to hand the public.

The trade-off is latency. If your workflow legitimately touches security tooling, the fallback can fire mid-task and swap models on you, something to plan around.

Claude Mythos 5 and the Dual-Use Question

Claude Mythos 5 is the gated sibling, the same model without the safeguard fallback. It keeps offensive-cyber capability that Fable 5 blocks, which is exactly why Anthropic doesn't release it publicly. Access runs through Project Glasswing, a vetted program for defenders and researchers who need full capability. Same model, two release postures: one open, one gated.

The chart makes the gating visible. On offensive-cyber evals (Firefox, OSS-Fuzz, CyberGym, CyScenarioBench) Mythos 5 scores 88.4, 24.0, 83.8, and 38.7. Fable 5 returns a flat 0.0 across all four. On offensive-cyber evals, Mythos 5 scores up to 88.4 where Fable 5 returns a flat 0.0: the safeguard split, made visible.

That zero isn't a capability gap. It's the fallback firing every time, blocking the request before the Mythos-class model answers. TechCrunch noted the timing: Anthropic shipped this days after warning that AI is getting too dangerous to release openly. The Fable/Mythos split is the policy answer to that warning, publish the capability, gate the danger.

For most developers, Mythos 5 is irrelevant. You'll never qualify for Project Glasswing and you don't need to. What matters is knowing that the model you're calling is deliberately capped on a narrow band of tasks, by design.

What This Means: Should You Switch?

Switch to Fable 5 for hard agentic coding, long-horizon multi-step tasks, and vision or spatial reasoning, anywhere finishing the task beats saving pennies per token. Stay on Opus 4.8 for cost-sensitive high-volume work, or for anything that triggers the fallback anyway. That's the whole framework. The rest is matching your workload to the right side of it.

When Fable 5 wins:

• Hard agentic coding where Opus 4.8 stalls; the SWE-Bench Pro gap is real
• Long-horizon tasks (multi-file refactors, migrations) where self-validation pays off
• Vision and spatial reasoning workloads
• Any job where one finished task is worth more than the token premium

When Opus 4.8 still wins:

• High-volume, cost-sensitive work where per-token price dominates
• Cyber/bio/chem workloads, where Fable 5 falls back to Opus 4.8 anyway, so call it directly
• Latency-sensitive flows that can't tolerate a mid-task model swap

On availability: Fable 5 is free on Pro/Max/Team/Enterprise plans until June 22, 2026, with usage credits applying from June 23. It's generally available on the API, AWS Bedrock, GitHub Copilot per the GitHub changelog, and Harvey. So you can test the switch in your editor today without changing your billing.

Frequently Asked Questions

What is Claude Fable 5?

Claude Fable 5 is Anthropic's first publicly available Mythos-class model, launched June 9, 2026. It scores 80.3% on SWE-Bench Pro and runs safeguard classifiers that fall back to Opus 4.8 on high-risk queries. It costs $10 input / $50 output per million tokens.

What's the difference between Fable 5 and Mythos 5?

They're the same frontier model with different release postures. Fable 5 is public and safeguarded; it falls back to Opus 4.8 on dangerous queries. Mythos 5 is the gated full-capability version, accessible only through Anthropic's Project Glasswing program for vetted defenders and researchers.

Is Claude Fable 5 actually better than Opus 4.8?

Yes on hard agentic coding, vision, and spatial reasoning: Fable 5 hits 80.3% on SWE-Bench Pro vs Opus 4.8's 69.2%. But Opus 4.8 still wins on cost-sensitive high-volume work, and Fable 5 falls back to Opus 4.8 for high-risk cyber/bio/chem topics anyway. "Better" depends on your workload.

How much does Claude Fable 5 cost?

Claude Fable 5 costs $10 per million input tokens and $50 per million output tokens. Anthropic's framing is "less than half the price of Mythos Preview." That's roughly 2× Opus 4.8 per token, but the value case is per finished task, since Fable 5 completes work cheaper models can't.

What is a "Mythos-class" model?

Mythos-class is Anthropic's new top capability tier, launched with Fable 5 and Mythos 5. It represents a frontier jump over the Opus 4.x line on coding and reasoning benchmarks. Fable 5 and Mythos 5 are two release versions of the same Mythos-class model: one public and safeguarded, one gated.

Why does Fable 5 fall back to Opus 4.8?

Fable 5 runs classifiers that detect high-risk requests (offensive cyber, bio, chem, or model distillation) and routes them to Opus 4.8 instead of answering with full Mythos-class capability. This fires on under 5% of sessions. It cut adversarial attack success from 56.6% (Opus 4.8) to 5.4%.

Can I access Claude Mythos 5?

Probably not. Mythos 5 is gated behind Project Glasswing, Anthropic's vetted access program for defenders and security researchers who need full offensive-cyber capability that Fable 5 blocks. Most developers won't qualify and don't need to, since Fable 5 covers everyday agentic and coding work.

Where can I use Claude Fable 5?

Claude Fable 5 is generally available on the Anthropic API, AWS Bedrock, GitHub Copilot, and Harvey. It's free on Pro, Max, Team, and seat-based Enterprise plans until June 22, 2026; usage credits apply from June 23. You can test it in your editor or via API today.

The Bottom Line

Claude Fable 5 is a real frontier step over Opus 4.8: 80.3% vs 69.2% on SWE-Bench Pro, double the FrontierCode score, and an attack-success rate of 5.4% that makes it far safer to deploy with tools. Switch for hard agentic coding, long-horizon tasks, and vision work. Stay on Opus 4.8 for cost-sensitive volume or anything that triggers the fallback anyway. It's free until June 22, so you can test the switch before it costs you anything. The Techsy team builds production agents on exactly this stack, so reach out if you want a hand.