Fable 5 at $50 per million output tokens turns model routing into a budget line

The money is in the output column. Fable 5 charges $50 per million output tokens against $10 per million input ^[6], a five to one spread ^[18]. A briefing turn is read-heavy: you pay to push the repository and the constraints in, and what comes back is a plan. Implementation inverts that, because the deliverable is the code itself. So the division of labour Breunig describes ^[8] hands the cheap tokens to the expensive model and the expensive tokens to something else.

On his one-ninth estimate ^[9], a million output tokens of routine implementation costs about $5.60 rather than $50, leaving roughly $44 per million on the table for anything else ^[16]. That is the entire case, and it is not a subtle one. It is a per-unit difference that survives contact with a spreadsheet.

What the router actually buys is conditional competence. Breunig does not claim GLM 5.2 matches Fable; he says it is sufficient for much rote coding, particularly when it receives strong context ^[10], and he is explicit that his comparison covers the economics of routine implementation after a stronger model has prepared the brief, not parity across every task ^[11]. The brief is the asset. You pay frontier rates once to produce it, then amortise it across the calls that do the typing.

Herb Sutter's original argument, which Breunig borrows ^[2], was that when single-threaded processor gains slowed, engineers had to start attending to parallelism, memory and architecture again. The model version ran on a shorter clock. In Breunig's account of the before state, "it felt silly to waste too much time improving your coding harness or context strategies. A new model would arrive at the same price (or cheaper!) and paper over most of your problems" ^[3]. Then, he writes, "the cost was so high and Opus was good enough (as was 5.6, K3, and even GLM) for most of the code we needed" ^[4].

Price is not the only field in the routing table. Anthropic requires 30-day retention for Fable 5 and Mythos 5 traffic including API use, and a zero-data-retention workspace has to switch retention on for the covered models before it can use them ^[12]. Anthropic's materials also describe additional safeguards for cybersecurity and biology requests, which makes model identity a routing input rather than a preference ^[13]. And on June 12, three days after launch ^[19], Anthropic disabled both models after the US government directed it to block foreign nationals, a directive the company said cited national-security authorities without specific details and forced a shutdown for every customer ^[14]. Routing is what gives you somewhere to send work on that day ^[15].

Which leaves one thing unresolved in the reporting. The same source calls Fable 5 Anthropic's most capable generally available model ^[7] and reports it switched off for the entire customer base three days in ^[14], without saying it came back ^[20]. Breunig, writing 75 days after release ^[17], uses the present tense: it "was (and still is!) incredible" ^[4]. Read that as evidence access returned, and the router stops being a cost trick and starts being the record of which dependencies you can actually rely on.