The right AI provider for a coding agent depends on three questions more than on any benchmark: how you want to pay, where your code is allowed to go, and which agent you already use. If you pay for ChatGPT, start with that subscription. If your company runs on AWS or Google Cloud, use Bedrock or Vertex AI. If code must not go to a model company, run the model yourself, on your own machine or on rented GPUs. If you want to try many models with one key, use OpenRouter.
The model matters too, but it changes every few weeks. These three questions don’t.
The short version
- You already pay for ChatGPT: use the ChatGPT subscription in Codex CLI or OpenCode. No API key, no per-token bill.
- You already pay for Claude Pro, Max or Ultra: use it in Claude Code, or in an app that runs Claude Code for you, such as Contenox. Max and Ultra also include monthly API credits for an Anthropic API key in other agents; they don’t cover Claude Code itself. See Anthropic.
- You already pay for GitHub Copilot: OpenCode signs in with it; Contenox only after you enable unofficial sign-ins. See GitHub Copilot subscription.
- Your company is on AWS: Bedrock, billed and governed like everything else in the account.
- Your company is on Google Cloud: Vertex AI, same idea.
- Code must not go to a model company: run the model yourself with Ollama on your own computer, or vLLM on a team server or rented GPUs.
- You want to try many models: OpenRouter, one key for most of them.
- You want DeepSeek’s models: use the DeepSeek API directly, in DeepSeek’s own agent DeepSeek Harness or any agent in the table below, or through OpenRouter. Read where its data is processed first.
How do you want to pay?
| Subscription | API key | Your cloud account | Run the model yourself | |
|---|---|---|---|---|
| Examples | ChatGPT, Claude, GitHub Copilot and Ollama Cloud plans | Anthropic, OpenAI, Gemini, DeepSeek, Mistral, OpenRouter | AWS Bedrock, Vertex AI | Ollama or vLLM on your computer, a team server, or rented GPUs |
| You pay | A flat monthly fee | Per token | Per token, on the cloud bill | The hardware you own, or the GPU hours you rent |
| When you hit the limit | Wait for the reset, or buy more | Your own spend limit stops you | Account quotas | Your GPUs are busy: requests queue and slow down, until you add capacity |
| Good for | Daily use by one person | Uneven use, teams, automation | Companies with existing cloud rules | Private code, steady heavy volume, models no provider offers |
Running the model yourself doesn’t mean owning a GPU. You can rent one, or a whole cluster, from a GPU cloud by the hour and run vLLM on it. You then pay for the time the machines run, not per token, which gets cheaper the more of that time your team keeps them busy.
Agent sessions use far more tokens than chat, because the whole conversation is resent every step. That makes a subscription attractive for daily use, and per-token billing painful when a session runs away. AI coding costs explains where the tokens go, and prompt caching how to keep the bill down on per-token providers.
Where is your code allowed to go?
Everything the agent reads, including your files, command output and sometimes secrets, is sent to whoever runs the model.
- A provider’s API or subscription: to that provider, under their terms for your plan. Read the data-use terms. Google’s Gemini API terms, for example, say free-tier prompts may be used to improve Google’s products and may be read by human reviewers, and tell you not to send confidential information on the free tier.
- DeepSeek: its privacy policy says data is processed and stored in the People’s Republic of China.
- OpenRouter: to whichever provider serves the request. You can restrict routing to providers with zero data retention.
- Bedrock or Vertex AI: your company’s existing cloud agreements apply, in the region you choose.
- Ollama or vLLM on your own computer or server: nowhere. The model runs on hardware you control.
- vLLM on rented GPUs: to the machines you rent, which you control, under the GPU provider’s terms. No model company sees it.
- Ollama Cloud: to Ollama’s servers. It is a hosted service, like any other provider.
If your employer or client has rules about this, they decide the shortlist before quality does.
Which providers work in my agent?
| Provider | Claude Code | Codex CLI | OpenCode | Contenox |
|---|---|---|---|---|
| ChatGPT subscription | No | Yes | Yes | Yes |
| Anthropic | Yes, plan or key | No | Key | Plan (runs Claude Code) or key |
| GitHub Copilot subscription | No | Codex VS Code extension, Pro+ and Max | Yes | Unofficial sign-in, off by default |
| OpenAI API | No | Yes | Yes | Yes |
| Gemini | No | No | Yes | Yes |
| DeepSeek | Works, unsupported | Yes | Yes | Yes |
| Qwen (Alibaba Cloud Model Studio) | Works, unsupported | Yes | Yes | Yes |
| Mistral | No | Yes | Yes | Yes |
| OpenRouter | Claude models only | Yes | Yes | Yes |
| AWS Bedrock | Claude models | OpenAI models | Yes | Yes |
| Vertex AI | Claude models | Not documented | Yes | Gemini models |
| Ollama | Works, unsupported | Yes | Yes | Yes |
| vLLM | Works, unsupported | Yes | Yes | Yes |
“Works, unsupported” means the provider documents a setup, but Anthropic doesn’t support Claude Code with non-Claude models, so some features may break. Each provider page has the exact steps for every agent. For the choice of agent itself, see the AI coding harness comparison.
Three patterns stand out:
- Claude Code is for Claude. It runs Claude models through Anthropic, Bedrock, Vertex AI or Microsoft Foundry. Other models only work through setups Anthropic doesn’t support.
- Codex CLI is for OpenAI first. It takes other providers only if they speak OpenAI’s Responses API, which rules out some, such as Gemini through its own API.
- OpenCode takes almost anything. Of the three, it is the one to pick if model choice is your priority.
Which model, then?
Once the provider is settled, pick the model by testing two or three on your own work, not by leaderboards. How to choose a model for coding has a one-afternoon method. A common setup is two: a strong model for planning and hard bugs, and a faster, cheaper one for routine edits. Given how far apart session costs are, that split matters: see AI coding costs.
Can I use more than one?
Yes, and it is common: a subscription for daily work, an API key or OpenRouter for when the subscription runs out, a local model for private code. OpenCode and Contenox let you connect several providers and switch per session; with Claude Code or Codex you switch by changing settings or environment variables.
Related guides
- AI coding costs: subscription, API key or local, in depth
- How to choose a model for coding: test models on your own tasks
- How to run local LLMs for coding: the local option in detail
- AI coding subscription comparison: what each plan includes and which agents it works in
- Terminal coding agents: Claude Code, Codex CLI and OpenCode compared