· AI
Picking an AI Model Is a Math Problem, Not a Loyalty Decision
The AI vendor decision that actually costs money is not which frontier lab to sign with. It is whether you build your own judgment about model fit in house or buy it from someone who has already made the mistakes.
Most businesses treat "which AI model do we use" as a vendor decision, made once, the way you pick a payroll provider. Sign with one lab, run everything through it, stop thinking about it. That instinct made more sense when the practical choice was between three frontier labs whose prices moved together. It does not hold up now that open-weight models, hosted by independent providers, sit far below frontier pricing for a large share of everyday work. Standing still on a single vendor is not simplicity. For most companies it is an unexamined tax.
The honest case for standing still
I want to steelman that instinct before I argue past it, because part of it is right. Nobody outside your company fully understands your workflow, your data, or where a wrong answer actually costs you money versus where it is merely annoying. Building that judgment in house means nobody else's incentives are in the room: a vendor who recommends a tool has reasons to recommend more tooling than you need. And switching cost is real. Standing up a second model, testing it against your own quality bar, and wiring it into whatever pipeline already exists takes engineering hours that a small team may not have free this quarter. Simplicity has a price, and sometimes it is worth paying.
What the data says about building this judgment alone
The trouble is what actually happens when companies try to build that judgment from scratch, with no outside reference point. Fortune's coverage of MIT NANDA's 2025 "State of AI in Business" study, based on interviews with 150 leaders, a survey of 350 employees, and an analysis of 300 public AI deployments, found that buying AI tools from specialized vendors or building through partnerships succeeded about 67 percent of the time. Internal builds, done alone, succeeded about a third as often. That gap is not a knock on any one team's competence. It is what you would expect when a decision is rare and expensive to get wrong: the people who have made it fifty times across fifty companies have a calibration that a team making it for the first time, on its own data, cannot match on attempt one.
The actual price gap, with real numbers
Here is the math a business owner can run this afternoon. Anthropic's own pricing page lists Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens, with Opus 5.5 at $4 and $20 for the heaviest reasoning work (claude.com/pricing). Together AI, a hosting provider for open-weight models, lists DeepSeek V4.1 Flash at $0.30 input and $1.20 output per million tokens on its own pricing page (together.ai/pricing). Take a mid-sized company running 200 million input tokens and 40 million output tokens a month of routine work: ticket drafts, document summaries, first-pass extraction, the kind of task that rarely needs a frontier model's full reasoning. Run all of it through Sonnet 5.5 and the bill is $800 a month. Route the same volume to DeepSeek V4.1 Flash and it costs about $108. That is not a rounding difference. It is 86 percent of the bill, every month, for work that likely did not need the more expensive model in the first place.
None of this required a mandate or a standards body. DeepSeek and the other open-weight labs undercut frontier pricing by shipping competitive models, hosting providers like Together built the infrastructure to serve them cheaply, and the frontier labs responded by holding their own prices in check rather than losing customers outright. That is competition doing the thing it does, without anyone voting on it.
What this does not mean
It does not mean route everything to the cheapest model. Some tasks genuinely need Opus-tier reasoning, and a wrong answer there costs far more than the token savings from a cheaper model that gets it wrong. The actual job is sorting your own workload into which bucket each task falls into, and testing a cheaper model against your own quality bar before trusting it with anything that matters, not against a vendor's published benchmark.
That sorting job is exactly the kind of decision the MIT NANDA numbers say goes better bought than built alone: not a slide deck, a real side-by-side test of your own work against both price tiers, done by someone who has seen where this choice goes wrong elsewhere. That is the engagement we run at MojoAI before recommending a stack to a client. If you are still paying one vendor's price for every task regardless of what the task needs, let's run your numbers before your next renewal.
Sources
References used in this article. Links also appear alongside the relevant claims.
Let's make it happen.
You don't need a polished brief. A couple of lines about where you want your company to go is plenty, and we'll come back with what we'd tackle first.
We scope the work around your goals and timeline before quoting anything, so you know exactly what you're getting.