AI & Machine Learning · Engineering, IT & AI
Should you build or buy Foundation Model APIs (LLM & Multimodal)?
Foundation model APIs sell metered access to frontier large-language and multimodal models: OpenAI (GPT-5.x), Anthropic (Claude Opus/Sonnet/Haiku), Google (Gemini 3.x), Mistral, and xAI (Grok), plus resellers and gateways (AWS Bedrock, Azure OpenAI, OpenRouter) that pass through provider pricing. Flagship rates run roughly $1-10 per million input tokens and $10-50 per million output tokens, with cheap commodity tiers (flash/mini/nano) at $0.10-0.50. This is the layer every AI chat, agent, copilot, and generation feature is built on.
Buy — this is the least ambiguous verdict in the index. Training your own frontier model costs hundreds of millions per run, and the companies that tried with real funding (Inflection $1.3B, Adept, Character.AI) all stopped and licensed their way out. The honest build-side question isn't training, it's self-hosting open weights (Llama, DeepSeek, Qwen) — real for data-sovereignty buyers and very-high-volume workloads, but for most organizations metered APIs are cheaper than carrying idle GPUs plus the MLOps team to run them. What you should own is the layer above: routing (cheap model for easy calls, frontier for hard ones), evals, and the product logic. Keep vendor switching cheap via a gateway; the price war is doing the rest of the work for you.
Build it, buy it, or bridge?
When building makes sense
Never train from scratch unless you are a frontier lab. The narrow legitimate build-side question is self-hosting open weights (Llama, DeepSeek, Qwen): justified for hard data-sovereignty requirements (the pattern behind Chinese SOE deployments and regulated-finance VPC installs) or workloads that keep dedicated GPUs saturated around the clock. Below that utilization, metered APIs are cheaper than idle capacity plus the seven operational surfaces a self-hosted deployment makes you own — serving, registry, observability, upgrades, auth, evals, and incident response. And remember self-hosting is still consuming a lab's model: it changes where the model runs, not whether you built it.
When buying makes sense
Always, for the model layer itself. Five capable vendors are in an active price war (flagship input prices down ~96% since GPT-4), gateways route between them with near-zero markup, and the capability gap between what you can rent and what you could ever train is widening. Buy metered, set a routing policy that sends easy calls to commodity tiers and hard ones to frontier models, and put your engineering into the layers that compound: evals, retrieval, and product logic. If your board asks whether you should 'own your own model,' the observed record — Inflection, Adept, and Character.AI burning nine figures before licensing their way out — is the answer.
The desk read
Build-versus-buy analysis for Foundation Model APIs (LLM & Multimodal) is being written. In the meantime, the framework that drives every B4 call is on the B4 Index page.
Vendors in Foundation Model APIs (LLM & Multimodal)
Each file covers what the product is, its funding history, and when the index last verified it alive.
Frequently asked
What are foundation model APIs?
Metered access to frontier large-language and multimodal models — OpenAI's GPT-5.x, Anthropic's Claude, Google's Gemini, Mistral, xAI's Grok — plus resellers (AWS Bedrock, Azure OpenAI) and gateways (OpenRouter) that pass through provider pricing. You pay per million tokens in and out.
Should we train our own model instead of paying OpenAI or Anthropic?
No. Frontier-class training runs cost $200-500M+ in compute alone, and the well-funded companies that tried — Inflection ($1.3B raised), Adept, Character.AI — all stopped and licensed their way out. Fine-tune or self-host open weights if you must; train from scratch, no.
When does self-hosting open-weight models (Llama, DeepSeek) beat the APIs?
Two cases with real observed adoption: hard data-sovereignty requirements (the pattern behind Chinese state-enterprise DeepSeek deployments and regulated-finance private installs), and workloads that keep dedicated GPUs saturated. Below constant saturation, metered APIs are cheaper than idle GPU capacity plus the MLOps team to run it.
How fast are prices falling?
OpenAI's flagship input price fell roughly 96% from GPT-4 (~$30/MTok) to the GPT-5.x tier ($1.25-5); Mistral Large fell 75%; Anthropic's Opus fell 67%. Meanwhile every vendor added commodity tiers at $0.10-0.50/MTok for volume routing. Premium reasoning tiers hold their price — the market is bifurcating, not uniformly deflating.
How do we avoid lock-in to one lab?
Route through a gateway (OpenRouter, Bedrock) with a policy layer: cheap models for easy requests, frontier models for hard ones, and evals that let you swap vendors on evidence. Gateways charge ~5% or pass through list price, so the insurance is nearly free.