IMR R&D · Touchstone · cost analysis

What the bench costs

Every result on Touchstone is produced on our own GX10 at zero marginal cost. This is what the same work would have cost on OpenRouter — priced from our real token counts against live API rates.

$0.00

our actual cost — 594 agentic supportflow runs on the local GX10

~$4.76

the same 594 runs on OpenRouter, model-for-model at cloud rates

~$487

if every run used a frontier model (Sonnet 4.6) instead

per-model — our roster vs its OpenRouter equivalent

Model (local)	OpenRouter equivalent	Runs	In/Out $/M	~$/run	~Total
qwen3.6-35b-a3b	qwen3.5-35b-a3b ~	246	0.14 / 1.00	$0.012	$2.97
gemma-4-26b	gemma-3-27b-it ~	100	0.08 / 0.16	$0.005	$0.48
qwen3-coder-30b	qwen3-coder-30b-a3b =	31	0.07 / 0.27	$0.010	$0.32
devstral-small-2	devstral-2512 ~	33	0.40 / 2.00	$0.009	$0.31
qwen3.5-4b	qwen3.5-9b (proxy)	88	0.10 / 0.15	$0.003	$0.26
mistral-small-3.1	mistral-small-3.1-24b =	7	0.35 / 0.56	$0.031	$0.22
gpt-oss-20b	gpt-oss-20b =	61	0.03 / 0.14	$0.002	$0.09
qwen3-next-80b	qwen3.5-122b-a10b (proxy)	22	0.26 / 2.08	$0.004	$0.08
granite-4.1-30b	granite-4.x (proxy)	6	0.10 / 0.40	$0.005	$0.03

= exact match on OpenRouter · ~ closest-size equivalent · proxy = price stand-in

anchored to real payments

$0.0035 / bench

gemini-2.5-flash-lite — measured: $0.03 for 8 supportflow benches via OpenRouter.

$0.82 / bench

Claude Sonnet 4.6 — measured: $6.56 for 8 benches (one 80-step hard bench was ~$5 of it).

Method & honesty. Costs use live OpenRouter prices (pulled today) × our real completion-token counts per model. Input tokens (the bigger driver in an agentic re-send loop) aren't logged locally yet, so they're estimated at ~15× completion — a ratio calibrated so the formula reproduces the measured Sonnet bill above. Treat per-model figures as ±~2×; the headline ($0 vs cloud) is robust. Next: log prompt tokens in the engine for exact costing.

The point. Bulk roster benching is effectively free on owned hardware — the GX10 has paid for itself many times over for this volume of agentic work (every run is dozens of model calls, not one). We spend real money only to buy the frontier ceiling line selectively, where $6 a run is worth knowing the top mark. Cloud-for-everything is the expensive way to get the same board.