b.ai posts zero‑markup model rates

AI model pricing is usually weird on purpose. You get credits, tiers, bundles, and “included usage,” but not a clean answer to a simple question — what does the model itself cost? B.AI is pushing the opposite direction. Its docs now show a public table of model prices across OpenAI, Anthropic, Google, DeepSeek, Kimi, GLM, and MiniMax, with a fixed conversion of 1 USD to 1,000,000 credits, so the numbers line up directly with dollar-denominated token pricing. ### What actually changed? The important move is not just “we have pricing.” Lots of platforms have pricing. The change is that B.AI exposed a cross-provider rate card in one place, with per-token input, cache-write, cache-read, output, and sometimes web-search charges all broken out. That makes the platform look less like a mystery bundle and more like a pass-through meter. (docs.b.ai) ### Why is the credit conversion the real trick? Because credits are usually where markup hides. If a platform says “1 million credits” but never pins credits to dollars, users cannot tell whether a model is being resold at cost, at a premium, or at some blended rate that changes by plan. B.AI’s docs pin the exchange rate exactly — 1 USD = 1,000,000 credits — which means 2.50 credits per token is just $2.50 per million tokens, and so on. (docs.b.ai) ### Do the posted numbers really look like base rates? Broadly, yes. B.AI lists GPT-5.4 at 2.50 input and 15.00 output credits per token-million, GPT-5.4 Nano at 0.20 and 1.25, Claude Sonnet 4.6 at 3.00 and 15.00, Claude Opus 4.6 at 5.00 and 25.00, Gemini 3 Flash at 0.50 and 3.00, and Gemini 3.1 Pro at 2.00 and 12.00. With the fixed credit conversion, those read like direct provider-style rates rather than padded reseller prices. (docs.b.ai) ### What else does the table reveal? It shows where costs really come from in agent systems. Output is often much pricier than input. Cache writes and reads have their own economics. Claude cache writes carry a 25% premium over standard input pricing on B.AI, while cache reads are heavily discounted. Web search is also explicitly priced per use on supported models — 10,000 credits for several GPT and Claude models, 14,000 for Gemini models. (docs.b.ai) ### Why does that matter for agents and MCP tools? Because once model costs are legible, the next hidden layer becomes tool costs. B.AI’s own ecosystem already leans into agent workflows through BAIclaw and OpenClaw integrations, so publishing model rates sets up a world where users can separate three bills: model tokens, search/tool invocations, and platform entitlements. Basically, the more agentic software gets, the less acceptable one blended mystery price becomes. (docs.b.ai) ### Does “zero markup” mean nobody makes money? No — it just shifts where the margin has to live. A platform can still charge for seats, support, reliability, orchestration, UI, analytics, enterprise controls, or bundled quotas. B.AI itself has subscription plans — Plan Pro at $200 a month and Plan Max at $2,000 a month — on top of prepaid credits. So the revenue story does not disappear. It just becomes easier to tell whether you are paying for the model or for the wrapper around it. (docs.b.ai) ### Why could this pressure other gateways? Because transparency spreads. Once one gateway publishes a clean table, users start asking every other gateway the awkward question — am I paying provider rates, or am I paying a tax hidden inside credits? That does not kill reselling, but it does compress the easy kind of margin, the kind earned by making the bill hard to parse. (docs.b.ai) ### Bottom line? B.AI’s move is small in format but big in implication. It turns model pricing from a trust-me product decision into a visible spreadsheet. And once buyers can see the base layer clearly, every extra dollar above it has to defend itself. (docs.b.ai)