Anthropic · OpenAI · Google Vertex AI · AWS Bedrock · Together AI · Replicate · OpenRouter · Modal · Fireworks AI · Groq.
One question: which one is right for your stage?

1. Anthropic Direct vendor · maximal lock-in to Anthropic models · also available via Bedrock + Vertex

Direct Anthropic API = single-vendor lock-in to Claude models — but Anthropic Claude is also served by AWS Bedrock and Google Vertex AI, which softens lock-in via cloud-procurement defensibility. Switching from Anthropic to OpenAI mid-product = real engineering work (different API shape, different model behaviors, different prompt engineering). The operator-honest path: use Anthropic SDK directly for fastest model access + accept the lock-in trade-off, OR route Anthropic Claude through OpenRouter / Bedrock / Vertex for procurement flexibility.

2. OpenAI Direct vendor · maximal lock-in to OpenAI models · also available via Azure OpenAI + OpenRouter

Direct OpenAI API = single-vendor lock-in to GPT + DALL-E + Whisper + Embeddings — but OpenAI models are also served by Azure OpenAI and routed via OpenRouter, softening lock-in. OpenAI SDK is the de facto standard — many third-party tools (LangChain, vector DBs, eval frameworks) speak OpenAI-compatible API natively. Switching from OpenAI to Anthropic mid-product = real engineering work (different API shape, different model behaviors).

3. Google Vertex AI Multi-model marketplace inside GCP · Anthropic Claude + Gemini + Llama + Mistral

Multi-model marketplace inside one GCP API — Gemini (Google's models) + Anthropic Claude on Vertex + Llama + Mistral via Vertex Model Garden. The right pick when you want optionality across multiple providers but inside the GCP IAM + audit + procurement boundary. Lock-in trade-off: you're locked to GCP-cloud-procurement, not to any single model vendor. Vertex Model Garden lets you self-deploy open-source models in your own GCP project for additional flexibility.

4. AWS Bedrock Multi-model marketplace inside AWS · Anthropic + Llama + Mistral + Cohere + Amazon + Stability

Multi-model marketplace inside one AWS API — Anthropic Claude + Meta Llama + Mistral + Cohere + Amazon Titan + Stability all served from one Bedrock endpoint. The right pick when you want optionality across multiple providers but inside the AWS IAM + KMS + VPC + CloudTrail + procurement boundary. Lock-in trade-off: you're locked to AWS-cloud-procurement, not to any single model vendor. Bedrock's multi-model breadth is the strongest in the category — same API, different models, one bill.

5. Together AI OSS-first · open-weight models = path to self-host = lowest lock-in

OSS-first hosting means the lowest lock-in story in the category — the underlying weights are downloadable, so if Together's pricing or posture changes you can move the same model to Modal / your own GPUs / another OSS host. Switching cost is near-zero for the model layer (same Llama 70B everywhere). Together's value is fast inference + dedicated endpoints + fine-tuning service on top of open weights — operationally easier than self-hosting, but not architecturally locked-in.

6. Replicate Public model marketplace · prototyping · low lock-in via OSS weights

Public model marketplace + open weights = low lock-in for OSS models, but moderate lock-in for the proprietary 'cog'-packaged deployments. If you use Replicate to host an open-source model, the underlying weights are downloadable. If you build a custom Replicate Cog deployment with proprietary code, switching cost is moderate. Best for prototyping where speed-of-evaluation beats long-term lock-in considerations.

7. OpenRouter Multi-provider routing aggregator · explicit anti-lock-in product

Explicit anti-lock-in product — one OpenAI-compatible API endpoint, 200+ models from 30+ providers, automatic fallback routing. Switching from one provider to another is a config change, not a code change. The right pick when 'never lock-in to one model vendor' is the architectural priority. Trade-off: you can't get direct enterprise contracts (BAA, custom DPA, custom rate limits) through OpenRouter that you'd get going direct to a single provider.

8. Modal Serverless GPU platform · model-agnostic · run any model = zero lock-in to model layer

Serverless GPU compute platform — model-agnostic by design, you run whatever model your code defines. Lock-in is at the platform layer (Modal SDK + Modal infrastructure), not the model layer. Switching from Modal to Runpod / Lambda Labs / Replicate / your own GPUs = real engineering work for the platform layer, but the underlying model + inference code is portable.

9. Fireworks AI OSS-first · open weights = low model-layer lock-in · platform-layer lock-in

OSS-first hosting means low model-layer lock-in (same Llama 70B / DeepSeek / Qwen weights are open) — Fireworks-specific features (function-calling implementation, JSON mode, fine-tuned variants) create some platform-layer lock-in. Switching from Fireworks to Together = moderate engineering work for Fireworks-specific features, but the underlying model is portable. Lower lock-in than frontier vendors, higher than pure OpenRouter aggregation.

10. Groq LPU specialist · open-weight models on LPU = low model-layer lock-in · LPU hardware lock-in

Hardware-specialty vendor — open-weight models on LPU = low model-layer lock-in (Llama / Mixtral / DeepSeek weights are downloadable), but LPU is unique hardware so latency advantage doesn't transfer if you switch. If sub-100ms latency is the deciding factor, Groq's hardware advantage is the lock-in (you can't replicate sub-100ms LPU latency on GPU). If latency is secondary, you can move the same model to Together / Fireworks / your own GPUs.

🔓 If you're a Architect minimizing single-vendor lock-in (operator-honest substrate flexibility)

Your problem: You've watched too many vendors get acquired, change pricing 3x, or sunset products. You want architectural flexibility — the ability to switch model providers without rewriting your product. Lock-in is the architectural risk you're optimizing against. See the sister AI Coding Tools comparison for the dev-tool lock-in decision.

1. OpenRouter — explicit anti-lock-in product — one API across 200+ models from 30+ providers, switching is a config change
2. Together AI / Fireworks — OSS-first hosting = downloadable open weights = path to self-host = lowest model-layer lock-in
3. AWS Bedrock — multi-model marketplace inside AWS — locks you to AWS but not to any single model vendor
4. Google Vertex AI — multi-model marketplace inside GCP — locks you to GCP but not to any single model vendor
5. Modal — model-agnostic serverless GPU platform — run any model your code defines, zero model-layer lock-in

🏢 If you're a Enterprise CTO standardizing across teams (multi-cloud reality)

Your problem: You're 1000+ employees with multi-cloud reality — some teams on AWS, some on GCP, some on Azure. You need AI infrastructure that works inside each cloud's compliance + procurement boundary. Standardizing on one cloud's AI marketplace creates org-wide lock-in. Cross-link to /operator cockpit for the operator-layer view of multi-cloud AI procurement.

1. AWS Bedrock — multi-model marketplace inside AWS for AWS-native teams — strongest enterprise procurement bundle
2. Google Vertex AI — multi-model marketplace inside GCP for GCP-native teams — strongest GCP-native procurement
3. Azure OpenAI — OpenAI models inside Microsoft compliance umbrella for Microsoft-shop teams
4. Anthropic direct — for teams needing direct Claude access (faster than Bedrock/Vertex by 1-2 weeks on new models)
5. OpenRouter — rarely the enterprise standard — useful for evaluation phase before direct contracts

🚀 If you're a Solo founder optimizing for evaluation velocity (try many models cheap)

Your problem: You're early. You want to try Claude vs GPT vs Gemini vs DeepSeek vs Llama in a week and figure out which model fits your workload best. Writing 5 SDKs is the wrong use of time. Lock-in is secondary; evaluation speed is primary.

1. OpenRouter — one API, 200+ models, switching providers is a model-name change — fastest model-comparison velocity
2. Anthropic + OpenAI directly (both) — two SDKs covers 80% of frontier evaluation if you have the engineering capacity
3. Replicate — easiest 0→hosted-endpoint for trying new multimodal models
4. Groq — free tier + sub-100ms latency — the OSS evaluation default for latency-sensitive workloads
5. Together AI — cheapest per-token on OSS — try Llama 70B / DeepSeek / Qwen at minimal cost

🛡 If you're a Regulated mid-market with procurement gates (BAA + DPA + ZDR + audit)

Your problem: You're 50-500 employees in a regulated industry. You need direct enterprise contracts (BAA + DPA + ZDR + custom rate limits) — multi-provider routing aggregators can't broker these. Lock-in is acceptable in exchange for procurement defensibility. Cross-link to Compliance Authority Graph for the regulated-procurement framework stack.

1. AWS Bedrock — multi-model marketplace inside AWS BAA + GovCloud — single-procurement multi-vendor optionality
2. Google Vertex AI — multi-model marketplace inside GCP BAA + audit — single-procurement multi-vendor optionality
3. Anthropic direct Enterprise — ZDR + HIPAA BAA available direct + operator-honest substrate
4. Azure OpenAI — ZDR + Microsoft compliance umbrella for Microsoft-shop procurement
5. Fireworks AI — SOC 2 + HIPAA BAA on enterprise tier for OSS-first regulated workloads