Closed source models

Closed source models are typically accessed through a vendor API or managed platform. Weights are not published for general download; you buy capability, latency SLAs and operational simplicity rather than full self-host control.

Executive summary

Study closed models as production systems: routing, reasoning-effort controls, tool calling, structured outputs, caching, safety layers, evaluation quality and cost per successful task — not as reconstructible neural architectures.

When to use this page

• Preferring managed inference over owning GPU clusters.
• Comparing provider features (tools, structured output, context, data residency).
• Framing buy-vs-build for a client who cannot self-host frontier MoE models.

Models in this category

Model	Publisher	Highlights	Guide
GPT-5.6	OpenAI	Sol / Terra / Luna tiers, routed reasoning, Responses API	GPT-5.6
Claude 5	Anthropic	Fable / Opus / Sonnet, Constitutional AI, MCP, computer use	Claude
Gemini 3.x	Google	Native multimodal, Search grounding, Live/TTS, Vertex AI	Gemini
Grok 4.5	xAI	Coding RL, long-running agents, reasoning effort	Grok
Amazon Nova	AWS	Bedrock-native, Guardrails, Sonic, Nova Act, SageMaker customise	Nova

How to read a closed-source guide

1. What kind of system — API product vs single homogeneous model.
2. Family lineup — tiers, effort controls, specialised variants.
3. Public disclosures — system cards and evals; what remains unknown.
4. Production contract — tools, caching, structured outputs, latency and cost.
5. Safety and deployment controls — classifiers, fallbacks, enterprise features.

What to do next

Start with GPT-5.6 or Claude, compare categories in the AI model landscape 2026, or read the LLM technical reading map.