Auditing a Model You Can See But Can't Price
What didn't fit inside the Claude Fable 5 review: the cost calculation that looks right and isn't, the footnote our fact-check nearly missed, and how the cloud picture moved the day after we published
The formal review is here. It is a Tier B hands-on writeup: what we measured on a Max 5x account, what we could not verify, and the verdict. This is the notebook next to it. The parts that do not belong in a review, but that shaped it.
Claude Fable 5 came back on July 1 after its June suspension. The review’s finding is that it returned metered, and its usage reads in a different unit on every surface. Tokens in Claude Code, a percentage in the Claude app, dollars only once you cross into paid credits. You can watch your usage climb. You cannot say what a task cost.
That is the finding. Here is what surrounded it.
Field note one: the cost calculation that looks right and isn’t
Fable 5’s API price is twice Opus 4.8. Ten dollars per million input tokens against five, fifty against twenty-five. Anthropic’s pricing page also says the newer tokenizer these models use turns the same text into roughly thirty percent more tokens. The intuitive next step is to multiply. Two times the price, times about thirty percent more tokens, and you land on Fable costing something like 2.6 times Opus.
It does not hold. Opus 4.8 uses the same newer tokenizer, so the thirty percent applies to both models and cancels out of the comparison. Against Opus 4.8, the difference is the flat 2x rate, and nothing more. The token inflation only bites when you compare Fable to an older model still on the previous tokenizer. We checked the multiplied number against the pricing page before using it anywhere, and it did not survive the check. The lesson we keep relearning: the clean-sounding derived figure is not the primary source, and the primary source is usually one page away.
Field note two: the footnote our fact-check nearly missed
One of the sharper details in the review is the Enterprise seat split. Standard Enterprise seats get no included Fable allowance at all and need usage credits enabled to run the model. Premium seats are included through July 7. That is a real procurement trap, and a team assuming uniform access can find part of its staff locked out.
We nearly missed it. A first automated research pass reported that no official page defined the seat breakdown. Only a direct read of Anthropic’s own redeployment post surfaced the footnote that spells it out, word for word. The difference between “not found” and “found” was not more searching. It was opening the actual document instead of trusting a summary of it. For a publication whose entire value is primary-source accuracy, that is the whole job in a single example.
Field note three: what changed the day after we published
Reviews of moving targets go stale, and the honest thing is to say when. On July 2, the day we tested, cloud availability was still catching up to the restoration. Anthropic’s own pages did not fully agree. The pricing page listed Fable across AWS, Google Cloud, and Microsoft Foundry, while the restoration post said cloud access would return as quickly as possible, without a date.
By July 3, that resolved. AWS confirmed Fable 5 is back on Amazon Bedrock, Google Cloud lists it as generally available, and Anthropic’s documentation now names all three clouds as available. One nuance remains worth checking for your own account. On Microsoft Foundry, Fable 5 runs through the Anthropic-hosted deployment option, while the Azure-hosted general-availability path currently lists Opus 4.8 and Haiku 4.5. So cloud parity is mostly back, with one hosting caveat. If the review’s decision map still lists cloud parity as a reason to wait, read that as the July 2 snapshot. The picture moved.
What we still could not measure
None of this closes the core gap. There is no official conversion from the app’s percentage meter to tokens or dollars for an individual plan. We did not cross into paid credits, so post-July-7 billing on a real subscription workload is still unverified. We did not pull an API or console receipt. A blocked Fable request reroutes to Opus 4.8 with a notification, which means a test you believe ran on Fable can quietly become a mixed-model test unless you log which model answered. The review lists these as open, and they stay open here.
The market angle, which is the point
When a vendor moves its most capable model from included to metered, the review question changes. It stops being whether the model is good, which Fable plainly is, and becomes whether you can see what it costs before you commit. Right now, across Claude’s surfaces, the answer is not clean. The usage is visible everywhere and priced nowhere until you spend.
This is not unique to Anthropic. It is where a lot of AI pricing is heading. Consumption billing wrapped inside a subscription, with meters that measure but do not translate. The teams that stay in control will be the ones that instrument their own runs and build the conversion evidence the vendor does not publish. Test it before you budget it, and log everything while you do.
The full hands-on review, with the numbers and the boundaries, is here.