20 July 2026 · 6 min read
Microsoft could save as much as $600 million by moving its Copilot off OpenAI and Anthropic models and onto Moonshot's Kimi K3, according to reporting this week. It is a useful number, because the lever behind it is the exact one the Agentic AI Cost Index measures: across 78 business processes and 26 models, which model you run is by far the biggest driver of what agentic AI costs. But the headline hides a twist worth understanding before you copy the move.
In the index, roughly 87 percent of the variance in cost per outcome comes from the model, not the process. The span between the cheapest and dearest model doing the same work is about 69 times: from around two cents per outcome on GPT-4o mini to about a dollar and a half on Claude Fable 5. At Microsoft's volume, shifting even part of Copilot's traffic to a model that is cheaper to serve turns that spread into a nine-figure line item. On that point the headline is right. The model is the single biggest lever on inference cost, for Microsoft and for a company running one workflow.
Here is the part the number skips. Kimi K3's published list price is $3 per million input tokens and $15 per million output. That is frontier pricing, the same tier as Claude Sonnet 4.6, and in the index it lands at about $0.47 per outcome, more than GPT-5.6 Terra at about $0.41. On the rate card, Kimi K3 is not a discount model.
So where does Microsoft's saving come from? Not the list price. It comes from the one thing Kimi K3 is that GPT and Claude are not: an open-weight model Microsoft can host on its own hardware. At Microsoft's scale, self-hosting strips out the per-token margin an API vendor charges, and that margin is where the $600 million lives. That is a lever hyperscalers have and almost nobody else does. If you buy tokens through an API, the list price is your price, and there Kimi K3 sits alongside the frontier models rather than below them.
Median cost per outcome in the index: GPT-4o mini about $0.02, GPT-5.6 Terra about $0.41, Kimi K3 about $0.47, Claude Fable 5 about $1.58. The cheapest option is not the newest or the most talked-about model. It is the one that finishes your work correctly for the fewest tokens.
The figure that runs a budget is not price per token. It is cost per finished, correct outcome, and two things push it above the sticker. Retries: a model that fails more often makes the agent redo the work. Reasoning: a model that thinks more, as Kimi's models tend to, burns more tokens reaching the same answer. Both raise cost per successful outcome while leaving the headline price per token untouched. And running a process as an agent already costs about 21 times a single chat turn in the index, once you count planning loops, sub-agents and tool calls. The right way to choose a model is cost per outcome on your own workflow, not the rate card.
The reason this swap is news is not the price, it is the politics. Kimi K3 is a Chinese open-weight model, and the US administration is reportedly weighing restrictions on exactly those models. The cheapest way to run a token can carry a governance cost that never appears on a rate card: export exposure, data residency, provenance, and the plain risk that a model you built on becomes something you are no longer allowed to use. The Cost Index measures token cost, which is one part of total cost. Governance is another, and it belongs in the decision alongside the dollars.
Kimi K3, GPT-5.6 Sol and Terra, Gemini 3.5 Flash, GLM-5.2, Qwen3.7 Max and the rest are all in the index as of 20 July 2026. You can see where each one lands on cost per outcome, and run the numbers on your own process rather than on a press release.
See the Agentic AI Cost Index → Run your own workflow →
For the small-business version of this, why the largest saving is usually not automating a broken process in the first place, read the companion piece on ProcessAnimation.