Amazon is routing most Alexa+ queries away from Anthropic's expensive frontier models to cheaper alternatives, and Stanford's 2026 AI Index confirms it isn't alone. The metric that now defines competitive advantage in AI isn't raw capability. It's useful output per dollar spent.
Anthropic launched Claude Opus 5 on July 24, priced at $5 per million input tokens and $25 per million output tokens. That's half the price of its flagship Fable 5 model, and Anthropic positioned it explicitly as near-frontier intelligence at half the cost. The move wasn't accidental. It was a concession to the market signal that's been building all year: enterprises are done paying frontier prices for tasks that don't need frontier intelligence.
Amazon is the clearest case study. The company rebuilt Alexa+ on a routing architecture that pulls from Amazon's own Nova models and Anthropic's Claude simultaneously, sending hard questions to the expensive models and deflecting the rest. According to reporting from The Next Web and CNBC, this is in direct response to a renegotiated contract with Anthropic that will shift Amazon to token-based billing in 2027, potentially raising what the company pays across products like Alexa for Shopping, the coding tool Kiro, and workplace assistant Quick. Some Amazon engineers are already distilling Anthropic models into leaner internal versions to front-run the pricing change. That's not a partnership fraying. That's a partner doing the math.
CNBC reported in June that two AI industry leaders described model routing as the emerging fix for enterprise overspending: match the task to the cheapest model that can handle it, and reserve the expensive calls for the work that genuinely requires them. Sounds sensible. For OpenAI and Anthropic, it's a structural threat to their business model. Their valuations were built on the assumption that enterprises would keep running nearly everything through frontier APIs. As CNBC noted, if high-volume, low-complexity queries migrate to cheaper open-source models, including Chinese models like Zhipu's, then the two dominant US AI labs end up with only the hard jobs. High-value, lower-volume, and much harder to scale into a recurring revenue story.
Glean CEO Arvind Jain told CNBC that roughly 95% of enterprise AI usage is still running on frontier models. That number will not hold. The economics are already rewriting themselves. Complexity-based routing can cut API bills by as much as 96%, according to industry cost analyses, and workhorse models cost 30 to 60 times less than frontier models per token. Once finance departments see those numbers, the decision isn't really a technical one.
Stanford's 2026 AI Index puts numbers on why this was always going to happen. The cost of inference for systems reaching GPT-3.5-level capability dropped more than 280 times between November 2022 and October 2024. Hardware costs are falling 30% per year. Energy efficiency is improving 40% per year. Smaller models are closing the gap fast: in two years, parameter counts for models scoring above 60% on the MMLU benchmark shrank by roughly 100 times. The frontier keeps moving, but the distance between frontier and good enough is compressing faster than frontier labs can monetize.
The routing stack that now makes sense #
For startups that priced their products and gross margins around the assumption that customers would accept frontier API costs as a cost of doing business, this is a real problem. AI-first SaaS companies are already running AI costs at 40 to 50 percent of revenue, compressing gross margins to 25 to 60 percent against the 75 to 85 percent that traditional software enjoys, according to analysis by Value Add VC. That gap isn't survivable at scale without either raising prices or cutting inference costs, and customers are increasingly unwilling to absorb the former. The smarter build now looks like this: a routing layer that dispatches cheap models for simple classification, retrieval, and formatting tasks, reserves mid-tier models for reasoning and summarization, and only escalates to frontier capability when the task is genuinely novel or high-stakes. That architecture is more complex to build and maintain than a single-model API call, but the unit economics make it non-optional for anyone running at volume. Frankly, the startups that wired directly to a single frontier model and passed the cost on are discovering that their competitive moat was never the model. It was always something else, and now they need to find it.
The broader implication for investors is less comfortable. The revenue assumptions embedded in Anthropic's and OpenAI's current valuations depend on sustained, high-volume consumption of premium-priced tokens. Anthropic is already responding, with Opus 5 priced to compete on the cost-efficiency dimension it previously ceded to smaller models. OpenAI is facing the same pressure. When your largest customer is building internal distilled copies of your model to avoid your pricing, and routing an increasing share of queries away from you toward cheaper alternatives, the revenue trajectory gets harder to defend. Amazon has committed up to $25 billion to Anthropic. That number was not a guarantee of margin, and it is starting to look less like a moat and more like a ceiling.
Also read: Ford's $1 Billion DRAM Bill Shows the AI Boom Is Now a Car-Buyer Problem • AI data centers have quietly become the biggest threat to grid stability America has ever built • Monday.com cuts 630 jobs and calls it an AI pivot, but the math tells a harder story