How to Use the Codex App with Any Model: GLM 5.1, MiniMax M3, MiMo V2.5 Pro, OpenCode Go

The Codex app is the best coding agent interface I have used so far. Worktrees, automations, the in-app browser, the review flow. Nothing else I run feels this polished. The problem is what happens on Thursday afternoon when my ChatGPT plan hits its weekly limit and the app politely tells me to wait until Monday.

Here is the part most people miss: Codex does not care where the model comes from. The app and the CLI read the same ~/.codex/config.toml, and that file lets you define custom model providers. Any OpenAI-compatible endpoint works. So when my OpenAI quota runs dry, I switch to GLM 5.1, MiniMax M3, MiMo V2.5 Pro, or my OpenCode Go subscription and keep working in the same app.

This guide covers the exact configs for all four. Copy, paste your API key, done.

How Codex talks to other models

Codex stores its configuration in ~/.codex/config.toml. The advanced config docs describe a model_providers table where you define extra endpoints:

model = "some-model"
model_provider = "my-provider"

[model_providers.my-provider]
name = "My Provider"
base_url = "https://api.example.com/v1"
env_key = "MY_API_KEY"
wire_api = "chat"

Three things to know before you start:

The Codex app picks up whatever default model and model_provider you set in config.toml. The CLI does too, plus it supports profiles for fast switching (more on that below).

Option 1: OpenCode Go (12 models, one key)

This is the setup I use most. OpenCode Go is a $10/month subscription that bundles 12 open source models behind one OpenAI-compatible endpoint: DeepSeek V4, Qwen 3.6, GLM 5.1, MiniMax, Kimi, MiMo, and more. Instead of juggling four provider accounts, you get one key that covers all of them.

Add this to ~/.codex/config.toml:

model = "glm-5.1"
model_provider = "opencode-go"

[model_providers.opencode-go]
name = "OpenCode Go"
base_url = "https://opencode.ai/zen/go/v1"
env_key = "OPENCODE_API_KEY"
wire_api = "chat"

Then export your key:

export OPENCODE_API_KEY="your-go-key"

Swap model for any model in the Go lineup: minimax-m3, mimo-v2.5-pro, deepseek-v4-pro, qwen-3.6-plus, and so on. The full model list and usage limits are in my OpenCode Go review.

If you only set up one provider from this article, make it this one. One key, 12 models, and the limits reset every 5 hours.

Option 2: GLM 5.1 (strongest coder)

GLM 5.1 is Z.AI’s flagship and the strongest open source coding model right now. It scores 58.4% on SWE-Bench Pro, ahead of GPT-5.4 and Claude Opus 4.6, and it can grind on a single task for hours without drifting. I covered it in depth in the open source Claude alternatives roundup.

Z.AI sells a GLM Coding Plan starting at $18/month, and the coding endpoint is OpenAI-compatible:

model = "glm-5.1"
model_provider = "zai"

[model_providers.zai]
name = "Z.AI GLM"
base_url = "https://api.z.ai/api/coding/paas/v4"
env_key = "ZAI_API_KEY"
wire_api = "chat"

export ZAI_API_KEY="your-zai-key"

Option 3: MiniMax M3 (cheapest to run)

MiniMax M3 is the newest release in the MiniMax line, and like the M2.7 before it, the pitch is price. This is the model I point Codex at for long refactoring sessions where token count matters more than squeezing out the last benchmark point. MiniMax also offers a Token Plan subscription with discounted rates, which pairs well with always-on agent work.

model = "MiniMax-M3"
model_provider = "minimax"

[model_providers.minimax]
name = "MiniMax"
base_url = "https://api.minimax.io/v1"
env_key = "MINIMAX_API_KEY"
wire_api = "chat"

export MINIMAX_API_KEY="your-minimax-key"

If you are outside the international region and use the mainland China endpoint, swap the base URL for https://api.minimaxi.com/v1.

Option 4: MiMo V2.5 Pro (best for long agent runs)

MiMo V2.5 Pro is Xiaomi’s flagship, built for agent workloads with hundreds of tool calls per session and a 1M token context window. In internal testing it built a full SysY compiler in Rust over 4.3 hours and 672 tool calls. It is the model I reach for when I hand Codex a task and walk away for the afternoon.

model = "mimo-v2.5-pro"
model_provider = "mimo"

[model_providers.mimo]
name = "Xiaomi MiMo"
base_url = "https://api.xiaomimimo.com/v1"
env_key = "MIMO_API_KEY"
wire_api = "chat"

export MIMO_API_KEY="your-mimo-key"

Switching between providers without editing config.toml

You can keep all four providers in one config.toml and only change the two top-level lines (model and model_provider) when you want to switch. That works, but it gets old fast.

Two better options:

codex --config model_provider='"zai"' --config model='"glm-5.1"'

model = "glm-5.1"
model_provider = "zai"

model = "MiniMax-M3"
model_provider = "minimax"

codex --profile glm
codex --profile minimax

My setup: OpenCode Go is the default in config.toml, and I keep a glm profile for the days I want GLM 5.1 through Z.AI directly with the bigger coding plan limits.

What still needs your OpenAI account

Custom providers cover local work: the app, the CLI, the IDE extension, file edits, terminal commands, MCP servers. Some pieces stay tied to OpenAI auth:

• Codex cloud tasks and the web environment run on OpenAI infrastructure with your ChatGPT login
• The GitHub integration for cloud-delegated work expects an OpenAI-backed account
• Reasoning summaries and model_verbosity need the Responses API, which these chat-based providers do not use

In practice this has not bothered me. I use the app locally, and local is exactly where custom providers work.

Which one should you pick?

If you want the longer comparison with benchmarks and per-token pricing, the best open source LLMs roundup covers GLM 5.1, MiniMax, MiMo, and three others side by side.

Wrapping up

The Codex app stopped being an OpenAI-only tool the moment custom providers landed in config.toml. Ten lines of TOML and an API key get you GLM 5.1, MiniMax M3, MiMo V2.5 Pro, or all 12 OpenCode Go models inside the same interface you already use. My ChatGPT plan still does the heavy thinking early in the week. After that, the open source models take over, and honestly, for most coding tasks I cannot tell the difference.

Related articles

• OpenCode Go: 12 AI Coding Models for $10/Month — the subscription I use as my Codex fallback
• Best Open Source LLMs to Replace Opus 4.7 or GPT-5.5 — benchmarks and pricing for GLM 5.1, MiniMax, MiMo, and more
• Best Cheap Models for AI Coding Agents — per-token pricing breakdown
• GitHub Copilot Alternatives After the June 2026 Pricing Change — more ways to cut AI coding costs