Every provider, one endpoint, the cheapest route wins every request.

Run on 2026-10-26: 7,400+ requests across 7 models and two Auto routers (7,420 in all), each sent once (one try per prompt) from our cloud servers.

• Routes: Foundry (Dispatch on its default cheapest routing, and Foundry Auto), OpenRouter (the same models, and openrouter/auto, on our own key) and buying direct from each provider at its list price.
• Models: DeepSeek V4.1 Flash, Qwen3 235B, GLM-4.7 and Kimi K2.6 on 330 prompts per route (Qwen 305: it skips the 25 tool-calling prompts on every route); Claude Haiku 4.5, GPT-6 Luna and Gemini 3.1 Flash Lite on a 50-prompt sample per route. GPT-6 Luna has no direct route in this test and Gemini 3.1 Flash Lite no Foundry route in this test.
• Prompts: a fixed set of 330 original prompts across coding, chat, long documents, summaries, extraction, classification, translation, math, reasoning and tool calls, plus two 5-turn coding agent sessions, three 4-turn support conversations and a repeat-prompt test (one long-document prompt sent 6 times in a row).
• Answers capped at 200 tokens for the route comparisons (60 for the repeat test). For the main chart every single model also ran once at each prompt's own answer length, bought direct (GPT-6 Luna through Foundry at the provider's list price), the same length the two Autos ran at. Thinking was turned off or set to its lowest level, the same on every route.
• Cost: what the buyer pays, fees included. OpenRouter: the cost on its receipt plus its 5.5% fee on credit purchases. Direct: the provider's billed cost at list price. Foundry: the charge on its receipt. Per-task comparisons add back prompt-cache discounts so caches warmed by earlier runs do not count; the caching saving is measured on its own in the repeat-prompt test.
• Speed: streamed requests; time to first token and output tokens per second from the stream.
• Quality: gpt-oss-120b as a blind judge (temperature 0) with a rubric per task type and reference answers where a task has one; a failed request scores 1.
• Auto: both routers got the same 330 prompts at each prompt's own answer length; OpenRouter's auto left 119 of its answers empty (its picks reason before answering), judged as given. On the 220 prompts not used to tune Foundry Auto: Foundry Auto $0.108 per 1,000 tasks at quality 8.40; OpenRouter Auto $0.388 at 6.08.
• Monthly bills are projected from measured costs: prices are per token, so per-task costs, percentages and multipliers hold at any spend.
• Every request is in results.csv, including the metrics where Foundry does not lead.

Sources, chart by chart. Charts 7 to 9 use no new runs.

• Auto against OpenRouter's auto: the Auto rows of results.csv and their judge scores.
• One model, every provider: each route's list price on the price sheet the run used (providers unnamed), next to what Foundry and OpenRouter charged in this run for the same prompts.
• Reliability: the router's gate test (each of 4 providers forced to fail 3 ways, whole and streamed: 24 requests, all answered), 8 live requests on staging with one provider's key refused (all answered by another provider or the backup), and every request this run sent to Foundry (2,241, 0 route errors).
• Cost per task, monthly bill, caching and speed: the single-model rows of results.csv.
• Fees and features: OpenRouter's pricing page and FAQ as read on 2026-10-26.
• Modes: the router's fixed test set (its gate test), drawn as written there.