How to Pick an AI Model and Vendor Without Overpaying

Most small businesses still pick an AI vendor the way they pick an accounting platform: once, for everything, and then they stop thinking about it. That approach made sense two years ago when the practical choice was one of three frontier labs and the prices moved together. It does not make sense now. As of late July 2026, OpenAI's GPT-5.6 Sol runs $5 per million input tokens and $30 per million output, according to TLDL's pricing tracker, while DeepSeek's V4 Flash runs $0.054 and $0.242, per OpenRouter's own analysis. That is roughly 100 to 150 times apart on output cost for models solving many of the same everyday tasks. Picking one vendor for every job in that spread is not simplicity, it is a tax.

I want to steelman the single-vendor approach first, because plenty of businesses are right to run it. If you send a few hundred requests a month, standardizing on one frontier API is the correct call: one contract, one support line, one quality ceiling that is never the weak link, and no engineering time spent managing a router between models. For a ten-person team with light, occasional AI use, the savings from shopping around are trivial next to the hours it would cost to chase them. Simplicity has real value, and I would tell a client that directly.

The case falls apart once volume or repetition enters the picture, and the market has made that obvious rather than subtle. OpenRouter's own data shows open-weight models are not a discount toy anymore: DeepSeek V4 Pro scored 80.6 percent on SWE-bench Verified, with the cheaper Flash variant close behind at 79 percent, and OpenRouter puts the capability gap between open-weight and closed frontier labs at three to six months, a gap that is holding steady rather than widening. Businesses have noticed. According to Digital Applied's Q2 2026 analysis, Chinese open-weight providers combined pulled more than 45 percent of all tokens moving through OpenRouter as of April 2026, up from under 2 percent a year earlier, with Xiaomi's MiMo alone taking 21.1 percent of weekly tokens, about three times OpenAI's 7.5 percent share. That is not marketing. That is money changing hands, at scale, because the quality gap is small and the price gap is not.

Here is the math a business owner should actually run. Say a company processes 500 million tokens a month of routine work: ticket triage, first-draft summaries, form extraction, the routine bulk of the work that does not need a frontier model's judgment. Split that 50/50 input to output and GPT-5.6 Sol, at $5 and $30 per million tokens, comes to roughly $8,750 a month. Route the identical workload to a hosted open-weight model like DeepSeek V4 Flash, at $0.054 and $0.242 per million, and the same 500 million tokens costs about $74 a month, better than a 100x drop, without touching the harder work that still deserves a frontier model's reasoning. You do not have to guess at this. Both price lists are public and your own token volume is yours to log.

Self-hosting is the next question, and here the free-market case gets more honest, not less. DoiT's cost breakdown puts an 8-GPU H100 node at roughly $12,600 a month on spot pricing and up to $71,800 on-demand depending on the cloud, and finds the break-even against metered API pricing lands around "a few dozen developers' worth" of steady usage, not before. Below that, a node sits underutilized and a metered API wins outright. The honest reason many businesses should not self-host is not ideology, it is that the fully loaded cost includes a second node for redundancy and the operations staff to run it, and for most companies that headcount does not exist and should not be hired just to save on inference.

None of this happened because a regulator mandated it. DeepSeek, Alibaba's Qwen, Xiaomi's MiMo, and a dozen smaller labs undercut the frontier price by building and shipping, and the frontier labs cut their own prices in response because customers had somewhere else to go. That is the whole mechanism. A rule restricting which models a business could route to, however well intended on security grounds, would remove exactly the competitive pressure producing this price collapse, and the security case deserves its own honest hearing rather than a dismissal, but it is a separate question from whether businesses should be free to shop.

The practical move is to stop treating model choice as a single procurement decision and start treating it as three: a frontier tier for the reasoning-heavy 10 percent, a cheap hosted open-weight tier for the routine 90 percent, and a self-hosting decision you revisit only once volume and staffing both justify it. If you want help mapping your own workload against that split, our team runs exactly that assessment before recommending a stack, and we're happy to look at your numbers.