Claude Fable 5 vs Mythos 5: Pricing & Benchmarks

Claude Fable 5 and Mythos 5

Anthropic released Claude Fable 5 on June 9, 2026, putting a Mythos-class model in front of the general public for the first time. It is the most capable model the company has shipped openly, and at $10 per million input tokens and $50 per million output tokens, it is also the most expensive model it sells to everyone. That rate is double Claude Opus 4.8 and twice the input price of OpenAI's GPT-5.5. For any team already watching its AI bill, that premium demands a clear answer to one question: when does this model actually pay for itself?

In this guide we cover what Fable 5 and Mythos 5 are, why Anthropic chose to release a frontier-tier model publicly now, how the pricing and discount levers work, what a realistic workload actually costs, how the models compare on benchmarks against Opus 4.8 and GPT-5.5, and a practical framework for deciding which model belongs on which task. We close with cost-control tactics, migration guidance, and answers to the questions teams are asking this week.

Key takeaways
Fable 5 and Mythos 5 are the same model; Fable 5 is the safeguarded public version, Mythos 5 is restricted. Both cost $10 per million tokens for input and $50 for output, double Opus 4.8. They lead GPT-5.5 and Opus 4.8 by wide margins on agentic coding and reasoning, and on hard, token-heavy tasks their efficiency can make them cheaper per completed job despite the higher rate. Fable 5 is free on paid subscription plans through June 22, 2026.

Why Anthropic Released a Mythos-Class Model Now

Until this week, the Mythos tier was something the public could only read about. Anthropic first introduced Mythos as a restricted preview in April 2026, available only through a program called Project Glasswing, citing the model's ability to identify and exploit software vulnerabilities as the reason for holding it back. That preview, the Claude Mythos Preview, carried a steep price of roughly $25 per million input tokens and $125 per million output tokens, and access was limited to a small group of cybersecurity partners and critical-infrastructure providers.

Fable 5 changes that. It is the same frontier capability wrapped in a safeguard layer that makes it safe to offer broadly, and it lands at less than half the price of the preview. The timing is notable because it arrives just days after Anthropic publicly warned that AI capability is advancing faster than the safety tooling around it. Releasing Fable 5 with classifiers, while keeping the unrestricted Mythos 5 locked to vetted partners, is how the company is trying to square growing capability with that concern.

For the broader competitive backdrop, our Anthropic vs OpenAI 2026 comparison lays out how the two labs are diverging on safety posture as much as on raw capability.

What Are Claude Fable 5 and Mythos 5?

Fable 5 and Mythos 5 run on the same underlying model. The only meaningful difference is access control. Mythos 5 ships with its safeguards lifted in specific high-risk areas and stays restricted to approved Project Glasswing partners. Fable 5 is the version made safe for everyone, with classifiers layered on top that decline or redirect requests touching cybersecurity, biology and chemistry, and attempts to extract the model's capabilities for use in rival AI systems.

Capability-wise, both are built for autonomous knowledge work and coding rather than quick chat replies. They support a 1M-token context window and up to 128k output tokens per request, accept text, image, and file inputs, and ship with tool use, memory, context compaction, and adaptive reasoning. The design intent is end-to-end work that would otherwise take a person hours, days, or weeks: long-running, ambiguous, highly multi-step problems where the model executes a scoped task, self-corrects through verification loops, and returns a result a human can review. The model ID is claude-fable-5, and it is available today on the Claude API, Claude Code, Amazon Bedrock, and GitHub Copilot.

How the Safeguards Work in Practice

If you are building on Fable 5, the safety layer is worth understanding because it changes how your integration behaves. The classifiers intercept prompts that fall into restricted domains. Importantly, the safeguards rarely interrupt ordinary work: by Anthropic's early data, more than 95% of Fable 5 sessions are handled entirely by Fable 5 itself.

• Refusals come back cleanly. When a request is declined, the Messages API returns a successful HTTP 200 response with stop_reason set to refusal, not an error. The response also reports which classifier triggered, so you can handle it gracefully instead of treating it as a failure.
• Some requests are handed off, not blocked. A small share of restricted prompts are quietly routed to Claude Opus 4.8, a lower-capability model, rather than answered by Fable 5 directly. For most teams this handoff is uncommon.
• Mythos 5 removes these classifiers. That is the entire reason it stays restricted to vetted partners. If you are on the public path, you are on Fable 5.

Claude Fable 5 and Mythos 5 Pricing

Both models carry identical list pricing. The headline rate is high, but the standard discount levers still apply, and for context-heavy agentic work the effective cost can land well below the sticker number.

The free evaluation window most teams will miss:
Fable 5 is included at no extra cost on Pro, Max, Team, and seat-based Enterprise plans from launch through June 22, 2026. From June 23 it moves to usage credits billed at API rates. If your team relies on subscription quota rather than pay-as-you-go billing, run your evaluations before that cutover so you are not surprised by the switch.

Pricing Modifiers Explained

The three levers that move your effective Fable 5 bill are worth understanding individually, because they stack and they behave differently from one another.

• Prompt caching. Reused context, such as a long system prompt, a style guide, or an entire codebase you reference across calls, can be cached so repeat reads cost about $1 per million tokens instead of $10. There is a separate, modest cost to write to the cache, but for any workload that reuses the same context across many requests, the savings dwarf the write cost. Agentic coding and RAG pipelines benefit the most.
• Batch processing. The Batch API returns results within a 24-hour window in exchange for a flat 50% discount on every token, taking Fable 5 to $5 input and $25 output per million. It is ideal for evaluations, content pipelines, data enrichment, and any job that does not need a real-time answer. There is no quality penalty, only a latency tradeoff.
• Inference geography. Global routing is the default and uses standard pricing. If you must guarantee US data routing, setting inference_geo to us applies a 1.1x multiplier across all token categories. Only opt in when a compliance requirement genuinely calls for it, since it raises every line of your bill by ten percent.

Stacked sensibly, these levers mean the headline $10/$50 is closer to a ceiling than a typical cost. A cached, batched agentic workload can run at a fraction of the sticker number, which is the single most important fact to internalize before judging Fable 5 as too expensive.

What a Real Workload Actually Costs

List prices are abstract until you map them to a task. Consider a single agentic coding job that reads 500,000 input tokens of context and produces 100,000 output tokens. Here is what that one task costs across the relevant models at standard rates, and what it drops to once you apply Fable 5's discount levers.

API Cost Comparison: Fable 5 vs Opus 4.8 vs GPT-5.5

On sticker price, Fable 5 sits clearly at the top of the publicly available frontier tier. The table below lines up the current public rates so you can see exactly where each model lands.

Benchmark Comparison

Pricing only matters relative to capability. On Anthropic's reported figures, Fable 5 opens a wide gap in agentic coding and complex reasoning, which is precisely the class of work where a premium model is supposed to justify itself.

What these benchmarks actually measure

SWE-bench Pro evaluates whether a model can resolve real software engineering issues end to end, editing across files, running tests, and iterating until the task passes, rather than just generating a plausible snippet. FrontierCode Diamond, from Cognition, is a harder, more agentic coding evaluation designed to separate frontier models from the pack on long, complex tasks. Both reward sustained, self-correcting work, which is why a model built for autonomous knowledge work pulls ahead. The gap matters most if your use case is genuinely agentic; for short completions or single-file edits, the difference narrows and cheaper models remain competitive.

Token Efficiency: Why the Sticker Price Misleads

The most important cost insight is that token volume is not constant across models. A more capable model that solves a problem in fewer steps consumes fewer tokens, and on hard tasks that efficiency can outweigh a higher per-token rate. One early customer completed a frontier physics research task in 36 hours using roughly one-third of the reasoning tokens GPT-5.5 needed across four days.

Put numbers on it. Suppose a complex task takes GPT-5.5 three million input tokens and 600,000 output tokens, while Fable 5 finishes the same work in one-third the tokens.

Real-World Results So Far

Early adopters point to the same theme: long, autonomous work that was previously out of reach. Stripe reported migrating a 50 million line Ruby codebase in a single day. Cursor and GitHub have cited tasks that earlier models simply could not complete. Rakuten highlighted that at its highest effort setting, the model reflects on and validates its own output, the kind of self-checking that makes unattended runs trustworthy. None of these are everyday chat prompts; they are the multi-day engineering and research jobs where the cost of human time dwarfs the model bill.

Is the Premium Worth It? A Selection Framework

The decision is rarely all-or-nothing. Most teams get the best economics by routing each job to the cheapest model that clears the quality bar, and reserving Fable 5 for the work that genuinely needs it.

• Reach for Fable 5 or Mythos 5 when the task is long-running, ambiguous, and highly multi-step: large codebase migrations, end-to-end agentic workflows, or deep research that would otherwise take a person days. This is where token efficiency and higher success rates flip the cost math in its favor.
• Stay on Opus 4.8 for strong reasoning at half the price when tasks are demanding but bounded and you do not need the extra agentic headroom. It remains an excellent default for hard work that is not open-ended.
• Use Sonnet or Haiku-class models, or GPT-5.5, for everyday generation, extraction, routing, and high-volume jobs. Our LLM cost optimization guide walks through how to set up this kind of tiered routing.

Should You Switch from GPT-5.5 or Opus 4.8?

If you are on GPT-5.5 today, the case for adding Fable 5 is strongest for agentic coding and autonomous research, where the benchmark and token-efficiency gaps are largest. It is weakest for high-volume, latency-sensitive, or short-context work, where GPT-5.5's lower rate wins. If you are on Opus 4.8, the question is whether your hardest tasks are hitting a ceiling. If Opus completes them well, there is little reason to pay double. If it stalls on long, multi-step jobs, Fable 5 is the obvious next step.

The clean way to decide is to run a head-to-head evaluation on ten to twenty representative tasks during the free subscription window, scoring each on success rate, total tokens, and wall-clock time rather than per-token price. Let those measured results, not the sticker rate, drive the choice.

Availability, Access, and the Usage-Credit Switch

Fable 5 is live immediately across several surfaces, which makes it easy to run a real evaluation today rather than waiting for a phased rollout. You can reach it through the Claude API directly, inside Claude Code, on consumption-based Enterprise plans, through Amazon Bedrock, and via GitHub Copilot. The model ID is claude-fable-5.

The detail that catches teams off guard is the billing transition. From launch on June 9 through June 22, 2026, Fable 5 is included on Pro, Max, Team, and seat-based Enterprise plans at no extra cost. On June 23 it leaves those plan limits, and continued use is billed against usage credits at API rates. If your testing depends on subscription quota rather than pay-as-you-go billing, schedule it inside that window. Anthropic has signaled it expects demand to be high and hard to predict, so capacity and limits may shift, which is another reason to verify live availability before you build production dependencies on it.

How to Control Fable 5 Costs

Because the output rate is $50 per million tokens, long generations are where bills grow fastest. Three levers keep spend predictable: enable prompt caching so reused context like a system prompt or codebase reads at about $1 per million, batch any non-urgent job for a flat 50% cut, and route aggressively so only the hardest work touches Fable 5. Set US-only inference only when you genuinely need guaranteed US data routing, since it adds a 1.1x multiplier. These are the same disciplines we cover in our token budgeting framework.

Frontier models also tend to split a single request into many sub-tasks, which is exactly how costs creep up quietly. Treating tokens as a metered utility, with per-team and per-feature budgets and anomaly alerts, is what keeps a powerful model from becoming an unpredictable one. Our guide to FinOps for AI token and GPU costs lays out how to instrument and allocate that spend so finance and engineering stay aligned.

Conclusion

Claude Fable 5 and Mythos 5 are the most capable models Anthropic has made available, and the $10/$50 pricing reflects that. They are double Opus 4.8 and a clear premium over GPT-5.5, but on agentic coding and complex reasoning the capability gap is large, and strong token efficiency can flip the cost math in Fable 5's favor on the hardest tasks. For most teams the smart move is selective: route the bulk of work to cheaper models and unleash Fable 5 where its autonomy saves days of human effort. Run your evaluation before the June 22 free window closes, instrument your token spend, and let measured results decide. If you want help turning these rates into a forecast and keeping AI spend allocated to the right teams and products, that is exactly the discipline FinOps brings.