Your One AI Vendor Just Became a Line Item Worth Cutting

Six months ago, picking one AI vendor and building on it looked like the safe call. Now it looks like a bet you should revisit. A price war, a wave of open-weight launches, and a big swing in where enterprise traffic actually goes have all landed in the same stretch. Let me catch you up on what changed and what to do about it.

The traffic already moved

Start with the number that should get your attention. Chinese AI models peaked at 46% of US enterprise token usage in a single week this summer, per a CNBC read of OpenRouter traffic. Eighteen months ago that figure was 4.5%. Since February, Chinese-origin models have carried at least 30% of enterprise token volume every week on the largest neutral router.

This is not about sentiment. It runs on economics. Open-weight models from labs like Moonshot and Z.ai cost a fraction of the frontier options and hold up fine on production work like extraction, summarization, and agent glue. Teams are moving spend, not making a statement.

The price war made single-vendor a liability

The vendors lit the fuse themselves. In one 48-hour window, OpenAI dropped a three-model GPT-5.6 family, xAI priced a coding API at two dollars per million tokens, and Meta shipped Muse Spark. OpenAI splitting one flagship into tiers was a defensive pricing move, and the rest piled on with per-token discounts.

Even Microsoft is talking down its own suppliers. Its sales team is now coached to compare Copilot against Claude, calling the rival model slower and less accurate inside Office apps. The company that leaned on OpenAI and Anthropic for years is swapping their models out of Word and Excel to cut costs. When your biggest partner starts hedging, the single-vendor stack stops looking like a foundation.

Open weights you can own and adapt

The open-weight side got a real product, not just cheap tokens. Thinking Machines shipped Inkling, a 1-trillion-parameter open model built to be fine-tuned rather than used off the shelf. The company says flat out it is "not the strongest model available today." The pitch is customization, and the proof is a Bridgewater project where a tuned open model scored 84.7% on financial reasoning while costing roughly a fourteenth as much to run.

There is a cost you may not see on the invoice. Satya Nadella warned that enterprises on proprietary models pay twice: once in subscription fees, and again by handing over business knowledge baked into thousands of prompts and corrections. Moonshot's Kimi K3 is expected to close the gap with Anthropic's Opus 4.8 at open-weight prices. The quality argument for locking in is getting thinner.

Routing is a systems problem, not a switch

Before you tell your team to route everything to the cheapest model, read the fine print. IBM researchers ran 417 tasks and found Claude Sonnet 4.6 cost $79 while GPT-4.1 cost $155, nearly double, despite lower sticker pricing. Caching flipped the math. Sonnet's lower cache-read pricing won out over its higher base rate and longer reasoning steps.

So routing by pricing sheet optimizes against the wrong numbers. The IBM team treats it as an optimization across cost, quality, and latency at once, not a simple "hard task, big model" classifier. Their difficulty-only baseline landed at higher cost than their tuned router for the same accuracy. Two more traps to name: switching models can break prompt caching and inflate your bill, and hidden reasoning tokens add a "thinking tax" that eats the savings you expected.

The deep cut

The savings are real, but they live in your workload, not in the price list. IBM's router hit a 21% cost cut and 9% lower latency versus running Opus alone, with only a 4% accuracy drop. That gap only shows up when you measure on your own prompts. Before any migration, run an A/B on your actual traffic through a router, keep a flagship model as a deterministic fallback for the hard cases, and cap reasoning tokens so the thinking tax does not erase the win. And if you are regulated, note the compliance wall: public aggregators like OpenRouter can push you toward pricier US-cloud endpoints, but you can run open weights on US-hosted managed APIs or your own VPC so tokens never leave. The move is not "pick the cheap model." It is measure, route with a safety net, and stop treating your vendor choice as permanent.

Three questions for your team

1. If our main vendor doubled prices or got pulled by export rules next quarter, how many days would it take us to route around them? If the answer is more than a few, we are too locked in.
2. Have we run our real prompts through a router to compare actual cost, including caching and reasoning tokens, or are we trusting the pricing page?
3. Which of our workloads are "good enough" for a cheaper open-weight model, and which few genuinely need the frontier model as a fallback?