Claude Mythos 5 Pricing and Production Tradeoffs

A pricing post that turned into a procurement post halfway through. Here’s how I’d budget for it anyway.

I went to route a long-running agent task to Fable 5 last week and got an error. Not a rate limit. Not a billing wall. The model was just… gone. Both Fable 5 and Mythos 5 had been pulled for every customer two days earlier, on June 12, over a US government export directive. So before I tell you anything about Claude Mythos 5 pricing, I have to tell you the thing the sticker price won’t: as of mid-June 2026, you can’t actually buy access to run it. That changes how you should think about every number below.

This is a cost post, but it’s also a reality check. I’ll give you the per-token rates, the Fable-vs-Mythos distinction, how it compares to Opus 4.8, and the production controls that keep frontier-tier spend from eating your month. I’ll also tell you where the math stops applying. Because right now, a chunk of it does.

The 30-second version

If you skip everything else, here’s the decision input:

List price for both Fable 5 and Mythos 5: $10 per million input tokens, $50 per million output tokens. Same rate. The name difference is about safeguards, not cost.
That’s exactly 2x Claude Opus 4.8 ($5 / $25). Output is the expensive half — 5x the input rate — so response length drives your bill more than prompt size.
Mythos 5 was never something you could just buy. It’s restricted to Project Glasswing partners. Fable 5 was the public version.
And as of June 12, 2026, both are suspended for all customers pending a government directive. Opus 4.8 and everything below it are unaffected.

So the honest answer to “how much does Claude Mythos 5 cost” is two-part: the rate is $10/$50, and the access is currently zero. Budget for the model only if you’re prepared for the restore date to slip.

What Claude Mythos 5 pricing actually means

Official pricing for Fable 5 and Mythos 5

Anthropic priced both Mythos-class models identically. From the official launch announcement, pricing for both models is $10 per million input tokens and $50 per million output tokens, and developers can use claude-fable-5 via the Claude API. The model IDs are claude-fable-5 and claude-mythos-5.

One detail that matters for budgeting: the input side carries the existing prompt-caching discount. Per the Fable product page, Claude Fable 5 is priced at $10 per million input tokens and $50 per million output tokens, with the existing 90% input token discount for prompt caching, and US-only inference is available at 1.1x pricing for input and output tokens. If you replay a lot of static context — system prompts, retrieved docs, tool schemas — caching does real work here. More on that below.

How current pricing compares with Mythos Preview

The $10/$50 rate is actually a markdown. The original restricted frontier tier, Claude Mythos Preview, ran far higher. Anthropic frames the new rate as less than half the price of Claude Mythos Preview, and reporting put the Preview at roughly $25/$125 per million tokens. So the trend line is the interesting part — frontier-tier inference is getting cheaper as capacity scales, not more expensive. That’s the right thing to extrapolate from. The specific number you should treat as provisional, because it already moved once.

Why access status changes the pricing question

Here’s where I have to break the usual “here’s the rate, go build” rhythm. Mythos 5 was always a restricted model, deployed through Project Glasswing. Claude Mythos 5 is restricted to Glasswing partners, with cyber safeguards lifted, and soon to select biology researchers, until the broader trusted access program is available. If you’re not a US-government cyber defender or an approved biomedical research org, Mythos 5 was never going to show up in your console. The version you’d actually run was Fable 5.

And then both got pulled. On June 12, Anthropic stated that the US government issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, and the net effect is that access must be disabled for all customers, while access to all other Anthropic models is not affected. Anthropic says it disagrees with the directive and is working to restore access. I have no idea when that lands, and neither does anyone outside the room. So: the price is real, the access is not — today.

Pricing vs actual workload cost

Sticker rate is the least useful number for planning. What you actually pay is rate × tokens, and on Mythos-class models the token count behaves differently than on smaller models. Fable 5 is built for long-horizon work — the kind of task that runs for hours and generates a lot of output. That’s exactly the profile where a 5x output multiplier compounds. A model that’s “only” 2x Opus on paper can be 3–4x on the invoice if your tasks are output-heavy and you let them run unconstrained.

Flip side: Anthropic and early customers report Fable 5 is more token-efficient on hard reasoning. One partner said Fable 5 completed a frontier physics research task in 36 hours using a third of the reasoning tokens it took GPT-5.5 four days to match. If you genuinely use fewer tokens to hit the same result, the effective cost gap narrows or flips. That efficiency shows up on complex multi-step reasoning, not on short queries — so the ROI is task-shaped, not flat.

Fable 5 vs Mythos 5: the cost reality

Same listed price, different access model

I’ll say it plainly because the naming is confusing: Fable 5 and Mythos 5 are the same underlying model at the same price. Anthropic’s own framing is that Mythos 5 is the same model as Fable 5 but with the safeguards lifted in some areas. The split is safety classifiers, not capability tiers and not a price tier. So “which is cheaper” is the wrong question. There is no cheaper one.

Why most teams would evaluate Fable 5 first

For anyone who isn’t a Glasswing partner, the choice was never really a choice — Fable 5 was the only door. The thing to understand about Fable’s cost behavior is the fallback mechanism, which the Fable 5 API documentation lays out for integrations. When its classifiers flag a request, the response is automatically handled by Claude Opus 4.8 instead, and more than 95% of Fable sessions involve no fallback at all. The classifiers trigger, on average, in less than 5% of sessions.

What does that mean for your bill? On a fallback, you’re paying for an Opus 4.8 response, not a Fable one — and you’re informed when it happens. For most workloads the blended effect is small. But if your domain brushes against cybersecurity, biology, or chemistry, you’ll hit fallbacks more often, and your effective model — and effective price profile — shifts toward Opus without you choosing it. That’s a planning input, not a gotcha.

Restricted access and procurement friction

There’s also a cost that doesn’t appear on any rate card: the 30-day data retention requirement. Anthropic now requires 30-day retention for all traffic on Mythos-class models, on both first- and third-party surfaces, and won’t use the data to train models or for any non-safety purpose. For some regulated buyers that’s a procurement conversation in itself. Not a dollar amount, but a real line item in the “can we even use this” column. Combined with the current suspension, the procurement story for Mythos-class is genuinely unsettled right now. I wouldn’t write it into a Q3 plan without a fallback model named in the same sentence.

Comparing cost with Opus 4.8

This is the comparison that actually decides most routing, so here’s the table I keep open.

Model	Input / 1M	Output / 1M	Relative to Opus 4.8	Access (mid-June 2026)
Claude Opus 4.8	$5	$25	baseline	Available
Claude Opus 4.8 (Fast)	$10	$50	2x (speed, not capability)	Available
Claude Fable 5	$10	$50	2x	Suspended
Claude Mythos 5	$10	$50	2x	Restricted + suspended

Opus 4.8 pricing is confirmed at $5 per million input tokens and $25 per million output tokens in standard mode, with fast mode at $10 and $50 for roughly 2.5x faster output. Worth noting: Opus 4.8 Fast Mode lands on the same $10/$50 sticker as Fable 5. So if your reason for reaching past Opus was latency rather than raw capability, Fast Mode gets you there at the same price without a Mythos-class model at all.

When Opus 4.8 is still enough

Most of the time. I mean that. Opus 4.8 is a strong model on its own — Anthropic reports it’s around four times less likely than its predecessor to allow flaws in code it has written to pass unremarked, scores 84% on Online-Mind2Web, and is the first model to break 10% on the Legal Agent Benchmark all-pass standard. For drafting, standard coding, analysis, most agent loops — you will not feel the ceiling. And it’s half the price and actually purchasable today, which right now is not a small thing.

When Fable 5 is worth the premium

The premium pays off on the longest, hardest, most multi-step tasks — the ones where Anthropic says the longer and more complex the task, the larger Fable 5’s lead over other models. Stripe reported Fable 5 compressed months of engineering into days, performing a codebase-wide migration in a 50-million-line Ruby codebase in a day that would otherwise have taken a team over two months. If a task is that shape — long-horizon, ambiguous, high-value — paying 2x per token to finish in a fraction of the wall-clock time is obviously worth it. The 2x sticker is noise next to the labor it replaces.

Why output-heavy workloads can change the ROI

The lever that moves your bill most isn’t the input rate — it’s output length, because output costs 5x input on these models. As one cost breakdown put it bluntly: output tokens cost five times more than input tokens, so the length of Claude’s responses drives your bill, not the size of your prompts. A long-running Fable task that generates verbose intermediate reasoning across hours can quietly become your most expensive line. Which is the whole reason the next section exists.

Production cost controls

If you do get access back and route real workloads to a Mythos-class model, these are the dials that matter. They matter on Opus 4.8 too — same family, same controls.

Effort parameter and task budgets

The effort dial is the single biggest cost lever. Lower effort on simpler tasks, reserve max effort for the work that needs it. The standard guidance for this model family: drop the effort level on simpler tasks, cache repeated prompt content, batch non-urgent jobs, and keep max_tokens tight. Setting a tight max_tokens is the cheapest insurance you can buy against a runaway response. Do it before you ship, not after the invoice.

Fallback and retry behavior

Two fallback behaviors to plan around. First, the classifier fallback to Opus 4.8 I mentioned — it’s automatic and you get told. Second, your own retry logic. A long-horizon agent that retries a failed step three times is paying for three runs. On a $50/M output model running for hours, sloppy retry policy is where budgets actually die. Cap retries, log token spend per run, and alert on outliers. Anthropic’s own caution for dynamic agent workflows is worth quoting: a misconfigured plan that spins 500 subagents instead of 50 is a 10x cost spike, so set budget limits in your harness before enabling dynamic workflows in production.

Routing by task value and risk level

This is the part I actually believe in. Don’t default everything to the most capable model. Route by what the task is worth and how much it’ll cost to run:

Cheap, simple, high-volume → Opus 4.8 or smaller, low effort. Never a Mythos-class model.
Latency-sensitive but not capability-bound → Opus 4.8 Fast Mode (same $10/$50 as Fable, but available).
Genuinely frontier — long, ambiguous, high-value → Fable 5, if and when it’s back, max effort, with budget caps.

The model you reach for should be a function of the task, not a habit. Defaulting to the top of the menu is how you end up explaining a 4x invoice to someone who asked for a 2x model.

FAQ

How much does Claude Mythos 5 cost? $10 per million input tokens and $50 per million output tokens — the same rate as Claude Fable 5. But Mythos 5 is a restricted-access model (Project Glasswing partners only), and as of June 12, 2026 both it and Fable 5 are suspended for all customers pending a government directive. So the rate is set; the access, for now, isn’t. Always confirm against Anthropic’s official pricing page before you budget, because this situation is moving.

Is Claude Fable 5 priced the same as Mythos 5? Yes, identically: $10/$50 per million tokens. They’re the same underlying model. The difference is safeguards — Fable 5 ships with safety classifiers for general use; Mythos 5 has some of them lifted for vetted partners. Same capability, same price, different access and risk posture.

Why can Fable 5 become expensive in long-running tasks? Because output tokens cost 5x input tokens, and Fable 5 is built for tasks that run for hours and generate a lot of output. A long agent loop with verbose reasoning, plus retries, can multiply your output spend fast. The 2x-over-Opus sticker understates the gap on output-heavy work. Cap max_tokens, control effort, and watch retry behavior.

When should builders use Opus 4.8 instead? Most of the time, honestly. Opus 4.8 is half the price ($5/$25), fully available right now, and strong enough for the large majority of coding, analysis, and agent work. Reach past it only when the task is genuinely long-horizon and high-value enough that finishing faster justifies 2x per token — and if you only need speed, Opus 4.8 Fast Mode hits Fable’s price point ($10/$50) while staying available.

So where does that leave you?

The number you came for is $10/$50 per million tokens, same for Mythos 5 and Fable 5, 2x Opus 4.8. That part’s easy. The part that actually decides whether you build on it is access — and right now access is suspended, with no firm restore date. My read: treat Mythos-class as a model you design for but don’t depend on yet. Name Opus 4.8 as your working default, build your routing so the frontier tier is a swap-in rather than a foundation, and keep the budget caps on regardless of which model answers.

When access comes back, the cost question becomes the routing question I sketched above — value and risk per task, not “most powerful by default.” Until then, I’d point your production load at something you can actually call. Go check Anthropic’s model overview docs and the Fable and Mythos suspension statement yourself before you commit a single token — this one’s still in motion.

Numbers verified against Anthropic’s official announcements as of June 13, 2026. Access status especially is changing; confirm before you budget.

Previous posts: