Claude Fable 5 is the most expensive model Anthropic has ever shipped. For two weeks, it's also free.
Anthropic put its “Mythos-class” intelligence into general release yesterday, June 9, as Claude Fable 5. The price is $10 in, $50 out per million tokens, twice what Opus 4.8 costs and the steepest rate card Anthropic has ever published for a model you can actually call. Then there is the twist: it is bundled free on every paid Claude subscription until June 22, after which the meter starts. We break down the real per-token cost, what the 2x premium buys against Opus 4.8 and the frontier field, and where the free window quietly turns into a bill.
Photo by Martin Martz on Unsplash
Two prices, and one of them is zero
Fable 5 lists at $10 input and $50 output per million tokens, exactly 2x Opus 4.8 down every billing line, with the full 1M context at standard rate and no long-context surcharge. Anthropic reports it near 80% on SWE-Bench Pro, roughly 11 points over Opus 4.8, though that figure is vendor-reported and not yet independently checked. The catch worth circling: it is free on Pro, Max, Team, and Enterprise plans through June 22, then the meter switches on June 23.
The price, line by line
Anthropic kept Opus list prices flat for three releases running, so a jump was overdue. Fable 5 doubles it. Every number on the rate card is exactly twice the matching Opus 4.8 line, which makes the comparison clean: you are paying a 100% premium, and the only question is whether the model earns it. Here is the full sheet, the same way Anthropic publishes it.
| Billing line | Fable 5 / Mythos 5 | Opus 4.8 |
|---|---|---|
| Input | $10.00 | $5.00 |
| Output | $50.00 | $25.00 |
| Cache read (hit) | $1.00 | $0.50 |
| Cache write (5min / 1hr) | $12.50 / $20.00 | $6.25 / $10.00 |
| Batch (input / output) | $5.00 / $25.00 | $2.50 / $12.50 |
One detail worth pocketing: the cache-read line is the cheapest way to feed Fable 5 a large, stable prompt. At $1 a million, a long shared system prompt replayed across calls costs a tenth of fresh input. If you are going to pay frontier rates, prompt caching is where you claw some back. The model IDs are claude-fable-5 on the Claude API and Vertex, and anthropic.claude-fable-5 on Bedrock.
What the doubled price is supposed to buy
A 2x price only makes sense if the model clears work Opus 4.8 cannot. Anthropic's pitch is the agentic-coding gap. On SWE-Bench Pro, the harder of the two SWE-Bench tracks, it reports Fable 5 around 80%, against a confirmed 69.2% for Opus 4.8. Here is the field on that benchmark, with one caveat in bold: only the Opus number is independently established. Fable's is Anthropic's own figure, published the day the model shipped.
SWE-Bench Pro, percent resolved (higher is better)
Fable 5 figure is Anthropic-reported and not yet independently verified. Opus 4.8, GPT-5.5, and Gemini 3.1 Pro figures reflect published SWE-Bench Pro results as tracked on tokencost.
Anthropic also claims state-of-the-art on FrontierCode and the Hebbia finance benchmark, and says Fable 5 is the first generally available model to clear 90% on a widely used analytics benchmark. Take the specific decimals with a grain of salt until third parties run their own evals, which usually takes a week or two. The shape of the claim, a clear lead on the hardest agentic tasks, is consistent enough across sources to plan around. The flat reasoning-and-chat lead is narrower, which is exactly why this is not a blanket upgrade.
The free window has a hard edge on June 23
Here is the part that catches people. From launch through June 22, Fable 5 is included at no extra charge on Pro, Max, Team, and seat-based Enterprise plans. You can route real work to the most expensive model Anthropic sells and pay nothing beyond your usual subscription. Then June 23 arrives, and on those same plans the model moves to credit or list-price billing until Anthropic has the capacity to fold it back in as a standard feature. There is no promised date for that.
Two things follow. If you are on a subscription, the next twelve days are a free trial of a $10/$50 model, so this is the moment to throw your hardest problems at it and see if the coding lead is real for your stack. If you are on the API, none of this applies and you pay list price from day one. The free window is a consumer-plan promotion, not an API discount, and reading it the wrong way is an easy way to be surprised by an invoice.
The quiet win: no long-context tax
Buried under the headline price is a genuinely good policy. Fable 5 carries a 1M-token context window, and every one of those tokens bills at the flat $10/$50 rate. Send a 900K-token request and it costs the same per token as a 9K one. That is not how the rest of the frontier prices long context. GPT-5.5 applies a 2x input and 1.5x output surcharge on the whole session once you cross 272K tokens. Gemini 3.1 Pro doubles its input rate, with output rising about 1.5x, past 200K. Anthropic dropped its own long-context surcharge back in March, and Fable 5 inherits that.
For a model aimed at long agentic runs and whole-repository reasoning, this matters more than it looks. The workloads that need a frontier coder are the same ones that pack the context window, and those are exactly the requests where a surcharge would bite. Fable 5 is expensive per token but honest about it, with no cliff waiting at 200K or 272K to double the bill mid-session.
What a month of real traffic costs
Numbers make the premium concrete. Assume a 70/30 input-output split, no caching, and run the same volume through Fable 5, Opus 4.8, GPT-5.5, and Gemini 3.1 Pro at list rates. This is what lands on the bill.
| Monthly volume | Fable 5 | Opus 4.8 | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|
| 10M tokens | $220 | $110 | $125 | $50 |
| 100M tokens | $2,200 | $1,100 | $1,250 | $500 |
| 1B tokens | $22,000 | $11,000 | $12,500 | $5,000 |
At a billion tokens a month you are handing Anthropic about $11,000 more than Opus 4.8 would cost, and roughly 4x what Gemini 3.1 Pro charges. The premium is real money at scale. Whether it pays back comes down to one question: does the coding lead cut your retry rate enough to cover the gap? On a hard agent that reruns failed tasks, a few percent fewer failures can. On bulk generation, it almost never will. Plug your own split into the calculator before you switch the default.
The safety leash, and why Mythos 5 stays locked
Fable 5 ships with a guardrail that affects what you get for your money. In high-risk domains, think cybersecurity, biology, chemistry, model distillation, it blocks and hands the request down to Opus 4.8 instead. Anthropic says at least 95% of Fable sessions run entirely on Fable's own output, so most users will never see the fallback. But it means the model you pay $10/$50 for occasionally answers at Opus quality on sensitive prompts, which is part of why some of Fable's coding scores reportedly trail the unrestricted version by a hair.
That unrestricted version is Mythos 5: same model, same price, safeguards lifted. It is not on the menu. Anthropic keeps it behind Project Glasswing for vetted partners only, the same gate that held Claude Mythos Preview. There is no signup, so for everyone outside that program, Fable 5 is the ceiling. One more practical note: Fable 5 requires 30-day data retention for safety monitoring, so zero-data-retention setups cannot use it. If your compliance posture depends on ZDR, this model is off the table regardless of price.
Reach for it on the hard problems, default to Opus 4.8
The honest read is that Fable 5 is a specialist, not a replacement. If you run agentic coding where a failed task costs real time, or you are chasing the last few points on a benchmark that decides a deal, the lead it claims over Opus 4.8 can justify the doubled rate. The free window through June 22 makes testing that nearly costless on a subscription, so there is no reason to guess. Route your gnarliest tickets at it for two weeks and measure the success rate against Opus.
For everything else, Opus 4.8 remains the smart default at half the price, and Gemini 3.1 Pro is cheaper still if you can live with a lower coding ceiling. The mistake would be flipping every workload to Fable 5 because it tops a chart, then watching the bill double for work that never needed the extra headroom. Put the rates side by side first. The full pricing table has Fable 5, Opus 4.8, and every frontier model in one view, so you can match the model to the task instead of the headline.
Sources
- Anthropic: Introducing Claude Fable 5 and Mythos 5 - positioning, free-access window through June 22, safety fallback
- Anthropic pricing docs - $10 / $50 list, cache and batch rates, full 1M context at standard pricing
- Anthropic models overview - model IDs, 1M context, 128K max output, June 9 release, availability
- GitHub Copilot changelog - same-day GA in Copilot, 30-day data retention requirement
- TechCrunch: Anthropic releases Claude Fable 5 - most powerful publicly released model, benchmark framing