From a single provider API to a fragmented landscape of 14+ options — how switching cost became the hidden engineering problem of LLM development.
Jun 2020
OpenAI API opens — there's only one provider
GPT-3 API launches by invitation. There's no "provider switching" problem because there's only one provider. All frameworks assume OpenAI by default. LangChain is built around this assumption.
Provider lock-in is invisible
Jan 2023
LangChain launches — OpenAI-first, community integrations as separate packages
LangChain's architecture choices: OpenAI is core, all other providers are ecosystem packages. This works for a world with one dominant provider. It creates friction for every world that comes after.
Architecture reflects 2023 reality
Mar 2023
Anthropic opens Claude API — second major commercial provider
For the first time, teams have a real choice between two competitive cloud LLM providers. LangChain adds langchain-anthropic as a separate package. The switching cost pattern appears for the first time.
First real switching cost
Nov 2023
LlamaIndex v0.10 — modular architecture, per-provider packages
LlamaIndex doubles down on the modular approach. Each provider integration is a separate PyPI package (llama-index-llms-groq, llama-index-llms-anthropic, etc.). Clean architecture, but the switching friction compounds as providers multiply.
Modularity vs switching ease
Jan 2024
Groq ships inference API — 10x faster than OpenAI on same models
Groq's LPU inference is orders of magnitude faster for certain workloads. Teams want to benchmark it. For the first time, switching cost shows up in sprint planning: is the speed improvement worth 3 changes × N files?
Speed creates switching urgency
2024
Provider explosion — Mistral, Cohere, Together AI, Fireworks, Ollama local
The provider landscape goes from 2 major choices to 10+. European providers (Mistral) for GDPR compliance. Local models (Ollama) for air-gapped deployments. Each one is a separate install in LangChain and LlamaIndex. Unified frameworks become a real consideration.
14 providers, 14 separate installs
2025
SynapseKit ships with unified provider interface — 14 providers, 0 extra installs
Single constructor, string-based provider selection. The argument: provider selection is a runtime config decision, not an import decision. Trade-off: you may not get cutting-edge provider-specific parameters the day they ship.
Architecture rethink
2026
LLM Showdown #5 — 1 line vs 3 changes, measured across 5 providers
The same RAG pipeline, switched across OpenAI → Anthropic → Groq → Mistral → Ollama. SynapseKit: 1 string change each time. LangChain and LlamaIndex: new install + new import + new class, every switch. The gap is structural, not incidental.
Measured, reproducible
The architectural takeaway
LangChain and LlamaIndex's per-provider package model was correct for 2023 — it enabled fast ecosystem growth and provider-specific features. The cost shows in 2026 when there are 14 providers and switching between them is a routine production decision, not a one-time architectural choice.
The unified interface model trades off first-day access to new provider parameters for zero-friction switching. Both are real tradeoffs. Neither is categorically better. What matters is knowing which one your architecture imposes on you.