Service Comparison: GenAI Platforms (Foundation-Model APIs)¶

Where to call a frontier or open-weights LLM in production. This is the API-and-model-routing layer, not the agent or app framework layer (see Agent frameworks comparison for that).

For underlying concepts: LLM basics, Tool use, Prompt caching.

Decision matrix¶

"Native tool use" means structured tool/function calling at the API level. "Prompt caching" means a real cache mechanism (not just retries).

Anthropic API¶

Direct access to Claude (Opus / Sonnet / Haiku families). Anthropic's first-party endpoint.

• Models: Claude 3.5/3.7/4 Opus, Sonnet, Haiku tiers; vision; computer use beta on supported models.
• Tool use: native, excellent with multi-step agentic workflows. Strong support for parallel tool calls.
• MCP: Anthropic created the Model Context Protocol; first-class client support.
• Prompt caching: ephemeral cache breakpoints (~5min TTL, ~25% write premium, ~10% read cost) and 1-hour extended cache.
• Streaming: SSE-based streaming; events include thinking, tool calls, stop reasons.
• Regions: US, EU. Workspaces for org-level isolation.
• Quotas: tier-based; enterprise has higher limits.
• No HIPAA BAA on first-party API directly - use Bedrock or Vertex for regulated workloads.

Pick Anthropic API when: you're building a Claude-first app and want the latest model features fastest (multi-step tool use, computer use, MCP).

📖 Anthropic API docs - models, messages API, tool use 📖 Claude pricing - per-token, per-tier

OpenAI API¶

The original frontier-model API. GPT-5 family, o-series reasoning models, embeddings, fine-tuning.

• Models: GPT family, o-series reasoning models, image and audio modalities (4o), embeddings, Whisper, DALL-E.
• Tool use: native function calling; structured outputs with strict JSON schemas.
• Prompt caching: automatic on prompts >1024 tokens, cached read at ~50% off; no manual breakpoints.
• Realtime API: low-latency voice and audio agents.
• Assistants API and Responses API: server-side agent loops with persistent threads.
• Regions: global; data residency available on enterprise tiers.
• Privacy: enterprise customers get data-not-used-for-training and zero-data-retention options.

Pick OpenAI API when: you need GPT-class models, o-series reasoning, voice / realtime, or built-in agent state via Assistants/Responses.

📖 OpenAI API docs - chat completions, function calling, streaming 📖 OpenAI pricing - per-model pricing

AWS Bedrock¶

Single API to many models including Anthropic Claude, Meta Llama, Mistral, AI21, Cohere, Stability, Amazon's own (Titan, Nova), and others.

• Models: Claude (full Anthropic family on Bedrock with slight version lag), Llama ¾, Mistral, Cohere Command, AI21 Jamba, Amazon Titan/Nova.
• Tool use: native; per-model availability varies.
• Prompt caching: yes for supported models (Claude family).
• Knowledge Bases: managed RAG (see vector DB comparison).
• Agents for Bedrock: managed agentic loop with tool/action groups.
• Guardrails: configurable input/output filters, PII detection.
• Privacy: VPC endpoints, PrivateLink, customer-managed KMS, HIPAA BAA, no data used for training.
• Regions: most major AWS regions; model availability varies.

Pick Bedrock when: you're on AWS, want VPC private endpoints, multi-model flexibility behind one API, or HIPAA/regulated workloads. Especially good when you want managed RAG (Knowledge Bases) or managed agents.

📖 Bedrock documentation - models, agents, knowledge bases 📖 Bedrock pricing - per-token + provisioned throughput

Azure OpenAI Service¶

OpenAI's GPT family hosted by Microsoft on Azure infrastructure.

• Models: GPT family, embeddings, Whisper. Slight lag behind OpenAI direct on newest models.
• Tool use: native.
• Prompt caching: same as OpenAI semantics.
• Privacy: VNet integration, private endpoints, customer-managed keys, HIPAA BAA.
• Regions: many Azure regions; model availability varies by region.
• Microsoft 365 integration: Copilot APIs, Graph data connectors, Azure AI Search.
• Content safety: built-in content filters configurable per deployment.

Pick Azure OpenAI Service when: you're on Azure, integrating with Microsoft 365, or need GPT in regulated regions with VNet isolation.

📖 Azure OpenAI documentation - service overview 📖 Azure OpenAI pricing - per-token

Google Vertex AI¶

Google's GenAI platform. Access to Gemini (1.5, 2.x), Anthropic Claude (via partnership), open models (Llama, Mistral, Gemma), and your own custom-deployed models.

• Models: Gemini family (text, image, video, code), Claude on Vertex, Llama ¾, Mistral, Mixtral, Gemma, plus Model Garden for many more.
• Tool use: native function calling.
• Context caching: explicit cache creation with TTL; useful for large stable contexts (e.g. 1M-token Gemini windows).
• Grounding: built-in grounding with Google Search.
• Privacy: VPC-SC, CMEK, HIPAA BAA, data not used for training.
• Regions: global; model availability varies.
• BigQuery / Vertex Vector Search proximity: easy retrieval against GCP data.

Pick Vertex AI when: you're on GCP, want Gemini's very long context windows, or want to deploy/serve custom models alongside frontier APIs through one platform.

📖 Vertex AI generative AI docs - Gemini, Claude on Vertex, Model Garden 📖 Vertex AI pricing - per-token, per-modality

Together AI¶

Specialist host for open-weights and fine-tuned models. Per-token pricing comparable to or lower than hosted-frontier APIs.

• Models: hundreds of OSS models (Llama ¾, Mixtral, Mistral, Qwen, DeepSeek, etc.).
• Tool use: yes for supported models.
• Fine-tuning: serverless LoRA fine-tunes.
• Inference: optimized vLLM-style infra.
• Pricing: per-token, generally lower than hosted frontier.

Pick Together AI when: you want OSS models in production at competitive prices without running inference servers yourself.

📖 Together AI documentation - models, inference, fine-tuning

Fireworks AI¶

Similar shape to Together: managed serving of OSS models with strong performance focus.

• Models: Llama, Mixtral, others; some proprietary Fireworks-tuned variants.
• Tool use: native function calling on supported models.
• Performance: known for low latency, particularly on smaller and quantized models.
• Fine-tuning: LoRA and full fine-tunes.

Pick Fireworks when: Together-like need but with stronger latency or specific model tunings.

📖 Fireworks AI documentation - models, function calling, tuning

Groq¶

Custom LPU (Language Processing Unit) hardware specialized for very-low-latency LLM inference.

• Models: curated OSS models (Llama family, Mixtral, Qwen, etc.).
• Latency: typically the fastest tokens/sec in the industry, often 5-10x faster than GPU-based hosting.
• Tool use: native on supported models.
• Limits: smaller catalog, no full-frontier models like Claude or GPT.

Pick Groq when: latency is the primary constraint - voice agents, interactive UIs, anything where time-to-first-token under 200ms matters.

📖 Groq documentation - models, API

Pick by scenario¶

flowchart TD
  A{Need a specific frontier model?}
  A -->|Claude| B{Where?}
  B -->|Anthropic-direct, latest features| ANT[Anthropic API]
  B -->|AWS / VPC / HIPAA| BR[Bedrock]
  B -->|GCP| VX[Vertex AI]

  A -->|GPT family| C{Where?}
  C -->|OpenAI direct, latest| OAI[OpenAI API]
  C -->|Azure / VNet / Microsoft 365| AOA[Azure OpenAI]

  A -->|Gemini| VX2[Vertex AI]

  A -->|OSS models| D{Constraint?}
  D -->|Cost| TG[Together AI]
  D -->|Latency| GQ[Groq]
  D -->|AWS-native| BR2[Bedrock]
  D -->|GCP-native| VX3[Vertex AI Model Garden]
  D -->|Self-host| SH[Self-host with vLLM<br/>see inference servers]

Cost intuition (per 1M input tokens, mid-2026 spot)¶

Numbers move; this is for relative ordering only. Confirm with vendor calculators.

Cross-references¶

• Concepts: LLM basics, Tool use, Prompt caching, Inference servers
• Topic: LLMs and GenAI
• Related comparisons: Vector databases, Agent frameworks, LLM observability
• Certs: Anthropic Architect Foundations, AWS AI Practitioner, Azure AI-102, Databricks GenAI Engineer