Low Cost Custom Models Threaten Thinking Machines

{"schema_version":"assistant_turn_v1","blocks":[{"type":"paragraph","runs":[{"text":"This pressure means model quality alone is no longer enough to justify a frontier price. Enterprises can now take open or open weight models from Meta, Mistral, and DeepSeek, fine tune them on company data, and run them through APIs or inside their own infrastructure. That gives buyers much of the same practical customization they want, while lowering token costs and reducing vendor lock in.","marks":[]}],"source_refs":[1,2,3,4,5,6,7]},{"type":"bullet_list","items":[{"runs":[{"text":"Mistral is built around this exact wedge. It offers cloud, serverless, and self hosted deployments, plus enterprise features like agents, document retrieval, connectors, and governance. The commercial pitch is not just a base model, it is a customizable stack that can live inside a customer environment.",\"marks\":[]}],"source_refs":[2,6,7]},{"runs":[{"text":"DeepSeek shows how fast price can collapse once strong open alternatives exist. Its API lists prices at $0.28 per 1M input tokens and $0.42 per 1M output tokens for DeepSeek V3.2, with even lower cached input pricing, while also promoting open source releases. That makes low cost customization credible, not theoretical.","marks":[]}],"source_refs":[3,4]},{"runs":[{"text":"The workflow has also gotten simpler. Teams increasingly start from Llama, Mistral, or similar pretrained models, then fine tune after deployment using live data. Infrastructure vendors now package training and serving behind OpenAI compatible endpoints, so switching from a frontier API to a customized open model is operationally much easier than it was two years ago.","marks":[]}],"source_refs":[5,8,9]}]},{"type":"paragraph","runs":[{"text":"Going forward, the winning labs will capture value by pairing strong models with distribution, enterprise tooling, and deployment flexibility. For Thinking Machines, that raises the bar from building a better model to building a product stack that feels meaningfully easier, safer, or more productive than a tuned open model running at a fraction of the cost.","marks":[]}],"source_refs":[1,2,5,6,9]}],"sources":[{"id":1,"url":"https://sacra.com/c/thinking-machines","label":"Thinking Machines company page","publisher":"Sacra","date":"2026-02-27"},{"id":2,"url":"https://sacra.com/c/mistral","label":"Mistral company page","publisher":"Sacra","date":"2026-02-18"},{"id":3,"url":"https://api-docs.deepseek.com/quick_start/pricing/","label":"Models and Pricing","publisher":"DeepSeek","date":"2026-02-19"},{"id":4,"url":"https://api-docs.deepseek.com/news/news250120","label":"DeepSeek-R1 Release","publisher":"DeepSeek","date":"2025-01-20"},{"id":5,"url":"https://sacra.com/research/towaki-takikawa-outerport-llm-deployment-mlops","label":"Towaki Takikawa on the rise of DevOps for LLMs","publisher":"Sacra","date":"2024-01-30"},{"id":6,"url":"https://help.mistral.ai/en/articles/316363-what-is-le-chat-enterprise","label":"What is le Chat Enterprise","publisher":"Mistral AI","date":"2026-01-01"},{"id":7,"url":"https://mistral.ai/news/mistral-code","label":"Introducing Mistral Code","publisher":"Mistral AI","date":"2025-06-04"},{"id":8,"url":"https://sacra.com/c/openpipe","label":"OpenPipe company page","publisher":"Sacra","date":"2026-03-12"},{"id":9,"url":"https://about.fb.com/news/2025/08/accelerating-indias-ai-adoption-a-strategic-partnership-with-reliance-industries-to-build-llama-based-enterprise-ai-solutions/","label":"Reliance partnership for Llama based enterprise AI solutions","publisher":"Meta","date":"2025-08-28"}]}