Claude Fable 5: Pricing, the API, and When to Use It vs Opus 4.8 (2026)

Claude Fable 5 is the public, safeguarded version of Anthropic's new Mythos-class model, released June 9, 2026. It calls as model ID claude-fable-5 and costs $10 per million input tokens and $50 per million output — double Claude Opus 4.8's rate. In high-risk areas (cybersecurity, biology, chemistry, distillation) it silently falls back to Opus 4.8, a switch that trips in under 5% of sessions. Reach for it on hard, long-running agentic and analysis work; wait if you need cheap tokens, a single guaranteed model, or you are bound by the new 30-day traffic-retention rule.

By Rohit Raj — Founding Engineer · 10+ yrs MVP shipping · LinkedIn

On June 9, 2026, Anthropic released Claude Fable 5 — the public face of a new model family it calls Mythos-class, and, by its own description, more capable than anything Anthropic has previously made generally available. The same announcement introduced Claude Mythos 5, the unsafeguarded sibling, which stays locked behind a restricted-access program. Fable is the one you and I can actually call today.

The capability that changed is not a benchmark bump — it is durability. TechCrunch reported that analytics firm Hex measured Fable as the first model to clear 90% on its benchmark of complex, long-running analytical tasks. That is the headline for builders: this model holds a thread across long agentic runs better than the frontier did a month ago, which is exactly where agents broke in practice.

The catch is that "more capable" came with a price hike and a leash. Fable lists at double Opus 4.8's token rate, and on a short list of sensitive topics it quietly hands the conversation to Opus 4.8 instead. This is the developer-only read: what shipped, the claude-fable-5 API and its real cost, where it earns the premium, the fallback behavior you have to design around, and when you should just keep using Opus 4.8. It launched the same week Anthropic warned regulators that frontier AI is getting dangerous enough to need a "coordinated brake pedal," so the safety scaffolding around Fable is not incidental — it is the product.

The structural news is the Fable/Mythos split. Anthropic trained one Mythos-class model and shipped it in two trims: Fable 5, with safety classifiers turned on for general release, and Mythos 5, the same underlying weights with those classifiers relaxed for vetted users. The official post frames Fable as "made safe for general use" while Mythos goes only to a small set of cyberdefenders and infrastructure providers through a US-government partnership it calls Project Glasswing, plus select biomedical researchers under a trusted-access program.

On raw capability, Anthropic claims gains across software engineering, knowledge work, vision, and scientific research, citing named evaluations — FrontierCode, the Hebbia Finance Benchmark, Cognition's internal evals, and, tellingly, long-horizon game-playing runs on Pokémon FireRed and Slay the Spire. The single number that traveled is Hex's: 90% on its long-running analytics benchmark, a first.

Two facts matter more than any leaderboard for production use. First, Anthropic ran an external bug bounty that surfaced no universal jailbreaks across 1,000+ hours of testing before launch. Second, it now retains all traffic for 30 days as a defense against novel attacks — a real change to the data-handling contract that I will come back to, because it can be a dealbreaker for regulated teams.

Fable 5 lists at $10 per million input tokens and $50 per million output tokens (Anthropic). TechCrunch notes that is double Opus 4.8's rate, which puts Opus 4.8 at roughly $5 / $25 per million — the comparison that should drive your routing decision.

Run the math on a realistic agentic task — say 40K input tokens of context and 8K tokens of generated code and reasoning per turn:

• Opus 4.8: 40K × $5/M + 8K × $25/M = $0.40 per turn
• Fable 5: 40K × $10/M + 8K × $50/M = $0.80 per turn

So a multi-step agent loop that does 50 turns swings from about $20 to $40 on the model line alone. That is not nothing at scale, and it is the whole reason "always use the best model" is bad engineering. Fable earns the 2× when the task is hard enough that Opus 4.8 fails or loops; it is pure waste on a task Opus already nails.

On access, there is a free window: Fable is included in Pro, Max, Team, and Enterprise subscriptions at no extra cost from June 9 through June 22, 2026, after which it draws on usage credits (Anthropic). On the API and consumption-based Enterprise it is metered from day one. If you only want to kick the tires, do it before June 23. For sizing token spend across vendors more broadly, I keep a running Claude vs OpenAI vs Gemini API cost breakdown.

Nothing about the call shape changed — it is the standard Messages API with a new model string, claude-fable-5. The one thing you must add is awareness that the model that answers may not be the model you asked for.

from anthropic import Anthropic

client = Anthropic()  # reads ANTHROPIC_API_KEY

resp = client.messages.create(
    model="claude-fable-5",
    max_tokens=4096,
    messages=[
        {"role": "user", "content": "Reproduce the failing test, fix it, and explain the change."}
    ],
)

print(resp.content[0].text)

# A safeguard classifier can route the request to Opus 4.8 instead.
# Don't assume Fable answered — read the model the API actually served:
if resp.model != "claude-fable-5":
    print(f"[fallback] served by {resp.model} — expect a capability + style shift")

That resp.model check is the single most important line for anyone shipping Fable. The Messages response echoes the model that produced the reply, so a fallback to claude-opus-4-8 is observable — but only if you look. Wire it into your logs from day one, because silent model swaps are the kind of thing that corrupts an eval set or an A/B test without anyone noticing for a week.

The premium is justified in exactly the places the previous frontier got flaky. Three concrete ones:

1. Long-running coding agents. The whole pitch of agent-first development — an agent that reproduces a bug, fixes it, opens the PR, and answers review — falls apart when the model loses coherence at hour two. Fable's reported strength on long-horizon tasks (the 90% Hex result, the game-playing runs) is aimed straight at this. If you are running the kind of harness-engineering loop OpenAI documented where agents open multiple PRs a day, model durability is the bottleneck, and Fable is a credible upgrade for the hardest tickets.

2. Complex, multi-source analysis. The Hebbia finance benchmark and the Hex analytics result both point at messy, long-context reasoning over many documents — the workload where models quietly drop a constraint halfway through. For a research or data-analysis agent, Fable is worth A/B-testing against Opus 4.8 on your own task before you commit.

3. Vision and scientific reasoning. Anthropic explicitly calls out vision and scientific research as areas of improvement. If your product reads charts, diagrams, or scientific PDFs, this is the axis to re-test.

The honest framing: none of these are "Fable is better at everything." They are "Fable fails less on the hard 20%." For the easy 80% of calls, Opus 4.8 at half the price is still the right default — route, do not replace.

Here is the decision table. The key insight is that Fable and Mythos are the *same model* — the difference is which safety classifiers are active and who is allowed to call it.

Sources: Anthropic launch post and TechCrunch. Opus 4.8 pricing is inferred from TechCrunch's "double" framing; if you are deciding between Opus generations, I broke down Opus 4.8 vs 4.7 for developers separately.

Three reasons to stay on Opus 4.8 — or wait — that the launch coverage glosses over.

The silent fallback breaks determinism. When a classifier fires, your request is answered by Opus 4.8, not Fable, in under 5% of sessions on average. That means at least 1 in 20 of your sensitive-topic calls returns a different model's output — different style, different capability, sometimes a refusal. For a security-tooling product like an AI vulnerability scanner, where "cybersecurity" is the entire job, that fallback could fire on a large fraction of calls and quietly gut your results. Test your real prompts before you trust the model string.

The 30-day retention rule is a compliance event. Anthropic now retains all traffic for 30 days as an attack defense. If you are under a data-processing agreement that forbids third-party retention of user content — common in health, legal, and fintech — that single line can disqualify Fable until your DPA is renegotiated. This is the failure mode I would actually worry about: not a bad answer, but a contract violation you discover in an audit.

The price only pays off on hard tasks. At 2× the cost, Fable on easy calls is a budget leak. If your eval shows Opus 4.8 already passing, switching to Fable buys you nothing but a bigger bill. And if you need cheap, high-volume classification or extraction, neither is the right tool — a smaller model is.

If a client asked me to put Fable into a live product this week, here is the wiring I would insist on — the parts that are not in the announcement.

Route, do not replace. I would keep Opus 4.8 as the default and send only the hard class of requests to Fable, gated by a difficulty signal (retry-after-failure, long-context, or a task-type flag). The cheapest reliable router is "try Opus 4.8, escalate to Fable on a failed self-check." This keeps the 2× premium on the 20% of calls that need it. I keep provider-routing options like OpenRouter, LiteLLM, and Portkey compared for exactly this.

Instrument the fallback. Log resp.model on every call and alert if the Opus-4.8 fallback rate drifts above your baseline — a spike means your traffic is hitting a safeguarded topic and your results are silently degrading. Without this metric, you are flying blind on whether you are even getting the model you pay for.

Treat retention as a feature flag. Make "send to Fable" a per-tenant toggle that is off by default for any customer whose contract forbids third-party retention. Do not bake the model choice into the prompt layer where compliance cannot see it.

Budget-cap before you scale. Set a hard monthly token ceiling per environment. A long-running agent on a 2×-priced model is precisely how a $200 test bill becomes a $4,000 surprise. The failure mode the README never warns about is not a wrong answer — it is an agent that loops on Fable's output tokens at $50/M while you sleep.

This is the kind of routing-and-guardrail layer I wire in from commit one when I build an MVP in 6 weeks — model choice behind a clean interface, cost caps, and observability, so swapping or sandboxing a model is a config change, not a rewrite.

What is Claude Fable 5? It is the publicly available, safeguarded version of Anthropic's Mythos-class model, released June 9, 2026. It is, by Anthropic's account, more capable than any model it had previously made generally available, with gains in software engineering, knowledge work, vision, and scientific research. Its restricted twin is Claude Mythos 5.

How much does Claude Fable 5 cost? $10 per million input tokens and $50 per million output tokens — about double Claude Opus 4.8's rate. It is included free in Pro, Max, Team, and Enterprise subscriptions from June 9 to June 22, 2026, then draws on usage credits.

What is the Claude Fable 5 model ID? claude-fable-5 on the standard Messages API. The call shape is unchanged from other Claude models.

Is Claude Fable 5 safe to use? Anthropic ships it with classifiers that fall back to Claude Opus 4.8 on cybersecurity, biology, chemistry, and distillation requests, triggering in under 5% of sessions. An external bug bounty found no universal jailbreaks in 1,000+ hours of testing. Note that all traffic is retained for 30 days.

What is the difference between Claude Fable 5 and Claude Mythos 5? They are the same underlying model. Fable has safety classifiers active and is public; Mythos has them relaxed and is restricted to Project Glasswing partners — cyberdefenders, infrastructure providers, and select biomedical researchers.

Should I switch from Opus 4.8 to Fable 5? Only for the hard tasks where Opus 4.8 fails — long-running agents and deep multi-source analysis. For the routine 80% of calls, Opus 4.8 at half the price is still the right default. Route by difficulty rather than replacing wholesale.

Claude Fable 5 is a real capability jump where it counts — long agentic runs and hard analysis — wrapped in the most conservative safety posture Anthropic has shipped on a public model. The two things to internalize are not on the marketing page: the Opus-4.8 fallback means the model that answers is not guaranteed to be the one you requested, and the 30-day retention rule is a compliance change, not a footnote. Build for both — log resp.model, gate retention per tenant, and route by difficulty so the 2× price lands only on the 20% of work that needs it.

If you want an AI feature built so the model layer is swappable, observable, and cost-capped from the first commit — instead of hard-wiring claude-fable-5 everywhere and discovering the fallback and the retention bill in production — that is the work I do. I ship production MVPs in 6 weeks and take founding-engineer engagements for India-based teams building on the current model stack.