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 Series E+ · Claude Sonnet/Opus · enterprise compliance posture

The operator-honest leader and substrate-of-choice for production AI products that need trust. Claude Sonnet 4.5 / Opus 4.x is the model SideGuy runs on every business day — PJ uses Anthropic API daily to ship the entire compliance graph + dashboard + Calling Matrix pages. Enterprise compliance posture is the strongest in the category: SOC 2 Type II, HIPAA BAA available, ISO 27001, zero-data-retention API contracts. The default substrate when 'two trillion-dollar companies wired together' is the architectural bet — Anthropic for intelligence + Google for discovery.

2. OpenAI Microsoft-backed · GPT-4/5 · category default · widest tooling

The category default — widest API surface, deepest tooling ecosystem, fastest 0→prototype velocity. GPT-4o / GPT-5 / o-series remain the broad-strokes default for prototyping, evaluation, and any product that needs DALL-E + Whisper + Embeddings + Realtime API + Assistants in one SDK. Most third-party tooling (LangChain, vector DBs, eval frameworks) defaults to OpenAI-compatible endpoints. Procurement-defensible at this point — every enterprise has reviewed OpenAI by now. Azure OpenAI gives you the same models inside Microsoft compliance umbrella.

3. Google Vertex AI GCP-native · Gemini 2.x · multi-cloud · enterprise data residency

The GCP-native enterprise AI platform — Gemini 2.x models + Vertex Agent Builder + multi-region data residency on GCP infrastructure. The default pick when your data already lives in BigQuery / GCS / Cloud SQL and you want AI inference inside the same VPC + IAM + audit perimeter. Vertex hosts both Google's own Gemini models AND third-party models (Anthropic Claude on Vertex, Llama, Mistral) — the second-best procurement option for Anthropic when you want it inside the GCP boundary.

4. AWS Bedrock AWS-native · multi-model marketplace · enterprise procurement default

The AWS-native AI procurement-defensible default — Anthropic Claude + Meta Llama + Mistral + Cohere + Amazon Titan + Stability all served from one AWS API. The right pick when AWS procurement, IAM, KMS encryption, VPC isolation, and CloudTrail audit are already the org standard. Bedrock's 'pick any frontier model from one bill' value prop is unmatched for AWS-shop enterprise — Anthropic Claude on Bedrock is contractually inside the AWS BAA + GovCloud perimeter, which is why most regulated AWS shops route Claude through Bedrock instead of direct.

5. Together AI Open-source model hosting specialist · Llama / Mixtral / DeepSeek / Qwen

The OSS-first inference specialist — Llama 3.x + Mixtral + DeepSeek + Qwen + 100+ open models served at competitive $/Mtok with strong throughput. The right pick when frontier-vendor lock-in is the wrong bet and OSS model quality (Llama 70B, DeepSeek-V3, Qwen 2.5) is good enough for your workload. Together's batched inference + dedicated endpoints + fine-tuning service make it the operator's default for cost-sensitive open-model serving at production scale.

6. Replicate Solo / prototyping favorite · easy model hosting · pay-per-second inference

The prototyping leader — easiest path from 'I want to try model X' to a working API endpoint, paying per-second of GPU compute. Image/video/audio model hosting (Stable Diffusion, Flux, video gen, music gen, voice cloning) is where Replicate dominates — most multimodal demos you see online run on Replicate. Pay-per-second metering, no commitment, no infra setup. Best tool for solo builders shipping AI features fast and for teams that want to test 50 models in a week without building 50 deployments.

7. OpenRouter Multi-provider routing aggregator · one API, many models

The model routing aggregator — one OpenAI-compatible API endpoint, 200+ models from 30+ providers (Anthropic, OpenAI, Google, Together, Fireworks, etc). The right pick when you want to A/B test models across providers without writing 30 different SDKs, or when you want fallback routing (if Anthropic 5xxs, route to OpenAI). Single bill for all providers. Margin sits between you and the underlying provider — slightly more expensive than direct, but operationally simpler. Increasingly the default for indie devs who refuse to lock to one provider.

8. Modal Serverless GPU compute · custom inference workloads · Python-native

The serverless GPU compute layer for custom AI workloads — write Python, deploy as a serverless endpoint with GPU autoscaling, pay only for compute time. The right pick when 'use someone else's hosted model API' isn't enough and you need to run YOUR fine-tuned model, YOUR custom inference pipeline, YOUR multi-step AI workflow with GPU acceleration. Modal sits between Replicate (easy hosted models) and full SageMaker (complex MLOps) — Python-native developer experience with serverless economics.

9. Fireworks AI Fast inference specialist · DeepSeek / Qwen / Llama optimized

The fast-inference specialist — optimized serving infrastructure for open models with industry-leading tokens-per-second on Llama, DeepSeek, Qwen, and Mixtral. Fireworks competes with Together on OSS hosting but bets harder on inference speed — proprietary serving stack, custom CUDA kernels, function-calling support, JSON mode. The right pick for low-latency open-model serving where you need fast time-to-first-token and high throughput on Llama / DeepSeek / Qwen at production scale.

10. Groq LPU hardware specialist · sub-100ms inference · fastest-in-category

The fastest inference in the category — Groq runs Llama / Mixtral / DeepSeek on custom LPU (Language Processing Unit) hardware that delivers sub-100ms first-token latency and 500-1000+ tokens/sec throughput. Hardware is the moat — Groq designed silicon specifically for LLM inference, not GPU-borrowed-from-graphics. The right pick for real-time voice agents, sub-second chatbot UX, or any product where 'feels instant' is the bar. Trade-off: smaller model selection vs Together / Fireworks (LPU memory constraints), and LPU inference doesn't yet support frontier-largest models.

🚀 If you're a Solo founder building an AI product

Your problem: You're a solo or 2-3 person team shipping an AI product. You need a substrate that gets you to working-prototype today, scales to first 1000 customers, and doesn't lock you into a procurement cycle when you raise. Cost matters but velocity + trust matter more. See the sister AI Coding Tools megapage for the IDE-side substrate decision.

1. Anthropic — Claude Sonnet 4.5 is the operator-honest substrate — refuses to fabricate, ships with HIPAA BAA when you need it, the production trust default
2. OpenAI — widest API surface for fastest 0→prototype + the deepest third-party tooling ecosystem
3. OpenRouter — if you want to A/B test Anthropic vs OpenAI vs Gemini without writing 3 SDKs
4. Replicate — for image/video/audio model hosting where you want easiest 0→endpoint UX
5. Together AI — if open-source models are good enough and you want $/Mtok cost control from day one

📈 If you're a Series A startup adding AI features to existing product

Your problem: You have product-market fit, paying customers, and now you're adding AI features. You need substrate that handles real volume, has SOC 2 + privacy controls your enterprise customers will ask about, and gives you procurement flexibility (no single-provider lock-in). Cross-link to AI Coding Tools comparison for the dev-tool substrate decision.

1. Anthropic — production substrate — SOC 2 + HIPAA BAA + zero-data-retention contracts close enterprise security review
2. AWS Bedrock — if you're AWS-native — Anthropic + Llama + Mistral inside one AWS bill + IAM + VPC perimeter
3. OpenAI / Azure OpenAI — Azure OpenAI for Microsoft-shop procurement defensibility, direct OpenAI for widest API surface
4. Google Vertex AI — if you're GCP-native — Gemini + Anthropic Claude inside GCP IAM + audit perimeter
5. OpenRouter — for evaluation phase before committing to one provider's enterprise contract

🏢 If you're a Mid-market integrating AI into core product (with security review)

Your problem: You're 50-500 employees, real security review, real procurement cycle. Your AI substrate has to clear a 4-12 week vendor onboarding process — SOC 2 Type II, HIPAA BAA if applicable, DPA + data-residency + zero-data-retention contracts. Single-provider lock-in is now a board-level risk. You also need to coordinate with frameworks in the Compliance Authority Graph (SOC 2 · ISO 27001 · HIPAA · GDPR).

1. AWS Bedrock — the AWS-native procurement-defensible default — Anthropic + Llama + Mistral all inside AWS BAA + GovCloud + CloudTrail
2. Anthropic direct — operator-honest substrate with enterprise compliance posture (SOC 2 + HIPAA BAA + ISO 27001) — most regulated mid-market routes Claude through Bedrock for AWS-bundle
3. Google Vertex AI — GCP-native — Gemini + Anthropic Claude inside GCP compliance boundary
4. Azure OpenAI — Microsoft-shop procurement defensibility — same OpenAI models inside Microsoft compliance umbrella
5. OpenRouter — rarely the procurement pick at this stage — direct enterprise contracts win

🏛 If you're a Enterprise CTO standardizing AI tooling across teams

Your problem: You're 1000+ employees standardizing AI tooling org-wide. Multi-cloud reality (some teams on AWS, some on GCP, some on Azure), strict procurement, central FinOps, audit + compliance + DPA + BAA across every tool. You're picking the substrate the next 5 years of AI products in your org will be built on — AI-baked-in vs AI-bolted-on matters at this horizon (see /operator cockpit for the operator-layer view).

1. AWS Bedrock — AWS-native enterprise default — Anthropic + Llama + Mistral + Cohere + Amazon + Stability inside one MSA + IAM + KMS + audit boundary
2. Google Vertex AI — GCP-native default — Gemini 2.x long-context + Anthropic Claude on Vertex inside GCP IAM + audit
3. Azure OpenAI — Microsoft-shop default — OpenAI models inside Microsoft compliance umbrella, deepest enterprise procurement maturity
4. Anthropic direct — for teams that need Claude direct (faster model access than Bedrock by ~1-2 weeks) with operator-honest substrate
5. OpenRouter — rarely the enterprise standard — direct provider contracts win, but useful for evaluation phase