Fable 5 is back on Claude subscriptions. This time you pay per token for it.
Anthropic yanked its most capable model off Pro, Max, and Team plans in June when demand outran the servers. It has returned, but not as a bundled perk. Fable 5 now runs on metered usage credits at its full $10 input and $50 output list rate, the steepest price Anthropic charges and exactly double what Opus 4.8 costs at standard tier. Below is what a credit actually buys, where the number bites, and the one figure Anthropic still refuses to print.
Photo by Pawel Czerwinski on Unsplash
Quick read
- Fable 5 lists at $10 input and $50 output per million tokens, the top of Anthropic's price sheet and twice Opus 4.8's standard rate.
- On subscriptions it moved from included access to metered usage credits after a June capacity pull, now billed at that same API rate.
- A 10M-in, 2M-out month runs $200, against $100 on Opus 4.8 and $110 on GPT-5.6 Sol. It earns the premium only on long, hard tasks, where a SWE-bench Verified lead cuts retries.
The full Fable 5 rate card
| Line item | Rate per 1M tokens |
|---|---|
| Input | $10.00 |
| Output | $50.00 |
| Cache read | $1.00 |
| Cache write (5-min) | $12.50 |
| Cache write (1-hour) | $20.00 |
| Batch API (input / output) | $5.00 / $25.00 |
From Anthropic's official pricing docs. Context window is 1M tokens with no long-context surcharge; output caps at 128K per request. The Batch API takes 50% off both input and output. API model ID is claude-fable-5.
The perk became a purchase
Fable 5 shipped in early June as the headline model on Anthropic's paid tiers, free to use inside your plan for a short window. Then it disappeared. Anthropic said demand was "difficult to predict" and that keeping it fully available was straining capacity, so it pulled the model from Pro, Max, and Team while it sorted the servers out. That is the version of Fable 5 most subscribers last saw: gone.
The redeployment kept the model but changed the deal. Per Anthropic's own note, Fable 5 stayed included for up to 50% of weekly usage limits through the first week of July, and after that its usage bills against metered credits rather than counting as bundled plan access. The company frames the shift as temporary and says it wants to fold Fable 5 back into subscriptions as a standard inclusion once capacity catches up. For now, though, using it on a plan means spending down a credit balance at the API rate.
A note on the calendar, because the reporting is messier than it should be. Anthropic's redeployment page still reads "through July 7." Several outlets reported the company then extended included access to around July 12, with credits taking over the next day. We could not confirm that extension on an Anthropic-owned page, so treat the exact cutover as mid-July and check your own plan's usage screen rather than a date in a headline.
What $10 and $50 come to in real work
Rates are abstract until you attach them to a job. Take two: a single heavy task at 100K input and 20K output, and a light production month at 10M input and 2M output. We price both on Fable 5 and set them next to Opus 4.8 at its standard $5/$25 and GPT-5.6 Sol at $5/$30, the two models most people would weigh Fable against.
| Workload | Fable 5 | Opus 4.8 | GPT-5.6 Sol |
|---|---|---|---|
| One task: 100K in / 20K out | $2.00 | $1.00 | $1.10 |
| One month: 10M in / 2M out | $200 | $100 | $110 |
| Same month on Batch API | $100 | $50 | n/a |
List rates, no caching, rounded. Fable 5 $10/$50, Opus 4.8 $5/$25 standard, GPT-5.6 Sol $5/$30. Batch halves Fable and Opus; GPT-5.6 Sol's batch tier is priced separately and omitted here.
The pattern holds at any scale because it is baked into the rate: Fable 5 is 2.0x Opus 4.8 and about 1.8x GPT-5.6 Sol, task after task. That is the number to keep in your head. Whatever you were spending on Opus for a workflow, Fable will roughly double it, and the Batch discount only resets the comparison one tier down without closing the gap.
It costs exactly what Opus 4.8 Fast mode costs
Here is the coincidence that reframes the whole decision. Opus 4.8 has two rates. Standard is $5/$25. Its Fast mode tier, the one that runs at about 2.5x speed and lives behind /fast in Claude Code, is priced at $10/$50. That is Fable 5's rate to the cent.
So the real question is not "is Fable worth double Opus." It is "at the same $10/$50, do I want the faster Opus 4.8 or the smarter Fable 5." If your bottleneck is latency on tasks Opus already handles well, Fast mode buys speed at that price. If your bottleneck is that the task is genuinely hard and long, Fable buys capability at the same price. You are choosing what the premium tier gives you, not whether to pay one.
Where the premium earns itself
Anthropic pitches Fable 5 at "the most demanding reasoning and long-horizon agentic work," and says the longer and harder the task, the wider its lead. The independent numbers back the direction even where they do not match Anthropic's framing exactly.
| SWE-bench Verified | Score |
|---|---|
| Claude Fable 5 | 95.0% |
| Claude Opus 4.8 | 88.6% |
| Grok 4.5 | 86.6% |
SWE-bench Verified figures as listed by third-party evaluator vals.ai. Anthropic's own announcement leaned on Cognition's FrontierCode evaluation instead, where it says Fable 5 ranks highest among frontier models even at medium effort, and did not publish an SWE-bench number of its own.
One score you may see quoted deserves a flag. A SWE-bench Pro figure near 80% has circulated for Fable 5, described as an 11-point lead over the field, but it traces to vendor-reported material and has been openly contested for lacking independent confirmation. We are leaving it out of the table rather than lend it weight. The Verified result and the FrontierCode framing are enough to make the case: on hard, long tasks Fable 5 does clear Opus 4.8, and that gap is the entire argument for the second $5 per million.
The token count you do not see on the invoice
Fable 5 runs on the newer Claude tokenizer, the same one under Opus 4.7 and later, Sonnet 5, and Mythos 5. Anthropic says it turns the same text into roughly 30% more tokens, with code and JSON near the top of that range and plain prose near the bottom. Because you pay per token, that inflation is a real cost that never appears as a line on the rate card.
Be precise about who it bites. Against Opus 4.8 it does not, since they share the tokenizer, so our $200-versus-$100 month is a clean token-for-token read. Where it matters is against GPT-5.6, Grok 4.5, or DeepSeek, which count on their own scales, and against your memory of what older Claude models cost on the same source text. If you are budgeting Fable off a spreadsheet built for Sonnet 4.6, add the 30% before you sign off.
The floor it towers over
Pull the camera back and Fable 5 sits alone at the top of a market that has spent the year racing the other way. The same 10M-in, 2M-out month, priced across the board:
| Model | Input | Output | 10M in / 2M out |
|---|---|---|---|
| DeepSeek V4-Pro | $0.44 | $0.87 | $6 |
| Grok 4.5 | $2 | $6 | $32 |
| Sonnet 5 (intro) | $2 | $10 | $40 |
| Opus 4.8 (standard) | $5 | $25 | $100 |
| GPT-5.6 Sol | $5 | $30 | $110 |
| Fable 5 | $10 | $50 | $200 |
Per million tokens, list rates, no caching, rounded. DeepSeek V4-Pro reflects its promotional $0.435/$0.87 and different tokenizer; Sonnet 5 intro runs through Aug 31. Non-Claude models count tokens on their own scales, so treat cross-vendor rows as close rather than exact.
Fable 5 does that month for what DeepSeek V4-Pro would charge for more than thirty of them. That is not a knock on Fable; it is the shape of the decision. The cheap seats exist and they are very cheap. Reaching for the priciest model only pays off when the task is hard enough that a wrong answer costs more than the token bill, which is a smaller slice of real work than the leaderboard makes it feel.
So is it worth double Opus?
Reach for Fable 5 when the work is long-horizon and the failure mode is expensive: multi-hour agent runs, gnarly refactors across a large codebase, research syntheses where a subtle miss propagates. That is the ground where its Verified lead and its widening advantage on longer tasks turn into fewer retries, and fewer retries is where a pricier model ends up the cheaper one on the invoice that counts.
For everything else, route down. Opus 4.8 standard handles most of what people assume needs the top tier at half the price, Sonnet 5 covers general and agentic work for a fifth, and the DeepSeek and Grok end of the board eats routine classification and extraction for pocket change. On a subscription, remember the credits are real money the moment included access ends, so the model picker is now a spending decision, not a free toggle.
And keep one eye on the unpublished number. Anthropic has not said how usage credits convert to dollars or tasks beyond the $10/$50 rate, so anyone quoting a neat cost-per-run is guessing. Price your own token mix against that list rate, treat the result as a floor, and revisit when Anthropic documents the conversion or restores Fable to bundled access.
Sources
- - Anthropic pricing docs (Fable 5 rate card, cache, batch): platform.claude.com
- - Introducing Claude Fable 5 and Mythos 5 (docs): platform.claude.com/models
- - Anthropic news, Fable 5 & Mythos 5 launch: anthropic.com
- - Anthropic news, Redeploying Fable 5 (50% of weekly limits): anthropic.com/redeploying-fable-5
- - BleepingComputer (temporary status, subscription return): bleepingcomputer.com
- - vals.ai SWE-bench Verified leaderboard: vals.ai
- - TokenCost pricing page: tokencost.app/pricing