AI Competition Is About Packaging

Meta just spent about 10 months rebuilding its artificial intelligence team, launched a new model called Muse Spark this week, and still faces the same question Wall Street asked before the launch: where does the money come from. CNBC reported on April 9 that Meta’s first major model in over a year is also its first big test of whether a proprietary model can turn billions in spending into revenue. (cnbc.com) That question keeps coming up because the model itself is no longer the whole product. In enterprise software, the sale is increasingly the wrapper around the model: who gets access, where it runs, what data rules apply, and how the bill changes when usage spikes. (cnbc.com) (help.openai.com) (learn.microsoft.com) Meta’s shift makes that visible. Muse Spark is proprietary, unlike the earlier Llama family that Meta released as open models, and CNBC said Meta plans to offer paid application programming interface access after an initial private preview with selected partners. (cnbc.com) OpenAI is making the same turn from a different direction. Its Codex coding agent now uses a pricing system that varies by plan and, for many business customers, by token usage, which means customers are charged by how much text goes in and out rather than by a flat message count. (help.openai.com) OpenAI updated that Codex pricing on April 2, 2026 for new and existing ChatGPT Business customers and new ChatGPT Enterprise customers, and the new table charges different credit amounts for input tokens, cached input tokens, and output tokens. The company says average Codex usage runs about $100 to $200 per developer per month, with higher costs when teams use larger models or fast mode. (help.openai.com) That is price discrimination in plain English. A light user can stay on a cheaper tier, a heavy engineering team can burn more credits on bigger models, and OpenAI captures more revenue without offering one single sticker price to everyone. (help.openai.com) Microsoft is pushing the same logic inside software companies already pay for. Microsoft’s documentation says Microsoft 365 Copilot can use Anthropic models in Word, Excel, and PowerPoint for customers in the European Union, the European Free Trade Association, and the United Kingdom, with administrators controlling that setting in the Microsoft 365 admin center. (learn.microsoft.com) That changes the competition from “whose model wins the benchmark” to “whose model gets placed in the workflow.” If a finance team is already writing in Microsoft Word and building forecasts in Microsoft Excel, Microsoft can swap in different model providers behind the screen while keeping the customer inside Microsoft 365. (learn.microsoft.com) (windowsreport.com) Governance is part of the package too. Microsoft says Anthropic processing for those Microsoft 365 app features occurs outside the Microsoft European Union Data Boundary, and that detail matters because large companies buy legal assurances and admin controls along with raw model output. (learn.microsoft.com) Meta’s problem is that it has spent heavily before building the tollbooth. CNBC said Meta told investors in January it expects $115 billion to $135 billion in capital expenditures in 2026, nearly double its 2025 figure, while analysts are still waiting to see whether Muse Spark becomes a paid service rather than another expensive demonstration. (cnbc.com) So the race is starting to look less like a sprint for the smartest chatbot and more like a fight over packaging. The winners may be the companies that can bundle a model into a spreadsheet, meter it like cloud software, and give procurement teams a policy page thick enough to sign. (cnbc.com) (help.openai.com) (learn.microsoft.com)