Kimi K3 vs Claude — model-routing re-price
A scheduled re-run of the “should we switch the content factory off Claude?” cost math. It re-checks live prices for Fable 5 / Opus 4.8 / Sonnet 5 / Haiku 4.5 vs Kimi K3, rebuilds the cost-per-day table, and updates the live article.
This job is the fleet doing math on itself. On 7/20/2026 we decided not to move the content
factory from the Claude ladder to Kimi K3 — routing inside Anthropic (≈$914/day modeled) beat an all-K3 switch (≈$2,385/day).
But three things were due to change around August 1, so we scheduled a one-shot to re-run the numbers instead of trusting a
two-week-old decision.
What it recomputes
- Kimi K3 open weights were due July 27 — which US hosts (Fireworks, Baseten, Modal, Together, OpenRouter)
now serve it, at what per-million-token price, under what license. - Anthropic Sonnet 5 was set to rise Sept 1 from $2/$10 to $3/$15 — confirm.
- Every tier’s live price re-checked against the primary pricing pages, then the factory-day table
(1.5B tokens in / 75M out / 80% cache hits) rebuilt for all-Fable, all-Opus, all-K3, and the routed 80/15/5 blend.
dollars per passed deliverable — not per token. Cheap output that fails review isn’t cheap.
Why it’s a one-shot, and why the timing is clean
It only needs to run once, right after the pricing dust settles. It lands Saturday around 4 AM — a weekend, when nothing
else in the fleet is scheduled and Dennis isn’t working — so it has the shared usage window entirely to itself. When it
finishes it updates the live article and the project-memory file, so the next decision starts from fresh numbers.
That’s the meta-loop: the fleet keeps re-pricing the fuel it runs on.