Pay-as-you-go. Only pay for what you use.

Inputs: your current monthly spend, and the share of your spend that today goes to a frontier-class model. The remainder is assumed to already be on a small model and is left untouched: it is not scored, so no routing fee is counted on it. Scoring it too costs $0.0007 a request, which at the 65% case below would take about 5 points off the band. The share starts at 100%: every step on a frontier model, the case the headline quotes. Pick your frontier model above. The picker lists only Anthropic’s models, because this estimate uses Anthropic’s price list; with another provider’s models, the band depends on that provider’s prices.

What Kongen does: on the frontier-share traffic, the routing layer downshifts a portion of requests to a small model when prompt complexity allows. Frontier-only requests stay on the frontier model. The decision is per-request, not per-account.

The math (in plain English):

• Frontier spend = total spend × frontier share.
• Of that, we assume 40 to 78% can be downshifted to a small model; this is an assumption, not a measurement. The remainder stays on the frontier model. If you already send the easier steps to a small model, less of what remains may be able to move, so below a 100% share the estimate can be much too high.
• Small-model cost is 1/4 of frontier cost, input and output, on Claude Opus 5.5 (1/5 on Claude Opus 4.8, 1/10 on Claude Fable 5.1), with Claude Haiku 4.5 as the small model, at Anthropic’s base (uncached) list prices as listed on 5 October 2026. The estimate assumes no prompt caching. Workloads that rely on caching can save less: a cache read is billed at 10% of the base input price or less, and on Opus 5.5 and Fable 5.1 the frontier model’s cache discount is deeper than the small model’s (a cache read on Haiku 4.5 costs 1/2 of one on Opus 5.5 and 2/5 of one on Fable 5.1).
• We then subtract the Kongen routing fee (1 KT, $0.0007 per request) and apply a 15% reroute allowance: the share of downshifted requests that fall back to the frontier model because the small model's answer wasn't good enough. For those requests you pay for both answers: the small model's, which is discarded, and the frontier model's.
• The low end of the range assumes 40% of frontier requests could move; the high end assumes 78%.
• The headline band this site quotes is 21 to 43%, against sending every step to a frontier model, at the estimator’s defaults; shorter requests save less. That is what this calculator returns at its starting position (a $2,000 monthly spend, all of it on Claude Opus 5.5 today). On the same assumptions the estimate is 23 to 48% on Claude Opus 4.8 and 29 to 57% on Claude Fable 5.1: only the model’s prices change, and the 40 to 78% downshift is the same assumption for every model. If 35% of your spend already goes to a small model, the estimate on Opus 5.5 is 13 to 28%. Move the sliders and you get your own numbers, which will differ. Until 2026-10-06 the headline was 14 to 30%, quoted for that case (65% of spend on a frontier model) instead of for every step on a frontier model, and computed without charging for a small-model answer that a fallback discards. Earlier bands are withdrawn: 20 to 40%, quoted until 2026-10-05 and computed on an earlier frontier list price ($75 per million output tokens); and, before 2026-09-20, 23 to 44% elsewhere on this site and 21 to 42% in this calculator, which disagreed with each other and with what the calculator itself computed.

List prices used:

• Frontier output: $20 / 1M tokens for Claude Opus 5.5 (the default), $25 for Claude Opus 4.8 and $50 for Claude Fable 5.1 (input: $4, $5 and $10), Anthropic’s base API list prices as listed on 5 October 2026.
• $25 is also the list price of Claude Opus 5, 4.7, 4.6 and 4.5, and $50 that of Claude Fable 5, at the same input prices, so the same figures apply to them.
• Small-model output: $5 / 1M tokens for Claude Haiku 4.5 (input: $1), Anthropic’s base API list price as listed on 5 October 2026.
• Average request: 1,000 tokens, all priced at the output rate, so about $0.02 on Opus 5.5, $0.025 on Opus 4.8 and $0.05 on Fable 5.1 (assumed). The band holds for requests of about that size on any of the three: 1,000 output tokens, or about 3,600 tokens at nine parts input to one part output. Shorter or mostly-input requests save less, because the $0.0007 fee is a larger share of them: a 1,000-token request that is 90% input gives about 12 to 34% at the starting position, and below about 520 tokens of that mix the low end reaches zero on Opus 5.5 (about 380 on Opus 4.8, 170 on Fable 5.1).
• Routing fee: $0.0007 per request, list price.

The model names and their prices are Anthropic’s, given only as the basis of this estimate. Naming them does not mean Anthropic endorses Kongen or this estimate.