feat(zenmux): route models through protocol-specific SDKs by llc1123 · Pull Request #1254 · anomalyco/models.dev

llc1123 · 2026-03-23T15:14:24Z

Summary

switch ZenMux's default SDK route from the Anthropic path to the OpenAI-compatible path
add per-model provider overrides for Anthropic, MiniMax, and OpenAI model families where a protocol-specific SDK works better
clean up ZenMux model metadata by correcting limits/capabilities and removing deprecated or unnecessary entries

Why

The biggest change in this PR is switching ZenMux's default routing to the OpenAI-compatible path. This affects every ZenMux model that does not define its own [provider] override.
In practice, the Anthropic Messages path tends to give worse cache behavior for many non-Anthropic models. We observed this clearly on GLM-5, and that behavior was one of the main reasons for moving the default route back to OpenAI-compatible.
The other reason is protocol compatibility. Some reasoning models, especially Kimi/GLM-style models, need interleaved reasoning_content to be preserved across turns. Interleaved reasoning_content lost for non-Claude models on Anthropic SDK opencode#14638 documents the concrete failure mode for Kimi K2.5 on the Anthropic SDK path: prior assistant reasoning_content is dropped on replay because @ai-sdk/anthropic only serializes Anthropic thinking blocks. Using the OpenAI-compatible path for these models avoids that incompatibility.
Anthropic and MiniMax models are explicitly routed through @ai-sdk/anthropic, while OpenAI models are explicitly routed through @ai-sdk/openai, so each family uses the SDK/API path that best matches its protocol and feature set.
This PR also fixes incorrect model metadata (for example Kimi K2.5 and Claude Opus 4), removes Claude 3.5 Sonnet because ZenMux has already taken it offline, and removes Gemini 3 Pro Image Preview because opencode does not need image generation support here.

Validation

I manually tested all updated model configurations and confirmed they work correctly.

Copilot

Pull request overview

This PR updates the ZenMux provider and model metadata so models are routed through protocol-appropriate SDKs/endpoints (notably Anthropic + MiniMax), while aligning capabilities/limits with observed runtime behavior.

Changes:

Switch ZenMux’s default SDK configuration to OpenAI-compatible (@ai-sdk/openai-compatible) using https://zenmux.ai/api/v1.
Add per-model [provider] overrides to route OpenAI models via @ai-sdk/openai and Anthropic/MiniMax models via @ai-sdk/anthropic (with the /api/anthropic/v1 endpoint).
Adjust/clean up model metadata (e.g., add interleaving where needed, tweak a capability flag, remove deprecated/unsupported model entries, update limits).

Reviewed changes

Copilot reviewed 37 out of 37 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
providers/zenmux/provider.toml	Changes ZenMux default SDK + base API to OpenAI-compatible `/api/v1`.
providers/zenmux/models/z-ai/glm-4.7.toml	Adds interleaving config for reasoning output (`reasoning_content`).
providers/zenmux/models/z-ai/glm-4.7-flash-free.toml	Adds interleaving config for reasoning output (`reasoning_content`).
providers/zenmux/models/openai/gpt-5.toml	Adds per-model provider override to use `@ai-sdk/openai` with ZenMux `/api/v1`.
providers/zenmux/models/openai/gpt-5.4.toml	Adds per-model provider override to use `@ai-sdk/openai` with ZenMux `/api/v1`.
providers/zenmux/models/openai/gpt-5.4-pro.toml	Adds per-model provider override to use `@ai-sdk/openai` with ZenMux `/api/v1`.
providers/zenmux/models/openai/gpt-5.4-nano.toml	Adds per-model provider override to use `@ai-sdk/openai` with ZenMux `/api/v1`.
providers/zenmux/models/openai/gpt-5.4-mini.toml	Adds per-model provider override to use `@ai-sdk/openai` with ZenMux `/api/v1`.
providers/zenmux/models/openai/gpt-5.3-codex.toml	Adds per-model provider override to use `@ai-sdk/openai` with ZenMux `/api/v1`.
providers/zenmux/models/openai/gpt-5.3-chat.toml	Adds per-model provider override to use `@ai-sdk/openai` with ZenMux `/api/v1`.
providers/zenmux/models/openai/gpt-5.2.toml	Adds per-model provider override to use `@ai-sdk/openai` with ZenMux `/api/v1`.
providers/zenmux/models/openai/gpt-5.2-pro.toml	Adds per-model provider override to use `@ai-sdk/openai` with ZenMux `/api/v1`.
providers/zenmux/models/openai/gpt-5.2-codex.toml	Adds per-model provider override to use `@ai-sdk/openai` with ZenMux `/api/v1`.
providers/zenmux/models/openai/gpt-5.1.toml	Adds per-model provider override to use `@ai-sdk/openai` with ZenMux `/api/v1`.
providers/zenmux/models/openai/gpt-5.1-codex.toml	Adds per-model provider override to use `@ai-sdk/openai` with ZenMux `/api/v1`.
providers/zenmux/models/openai/gpt-5.1-codex-mini.toml	Adds per-model provider override to use `@ai-sdk/openai` with ZenMux `/api/v1`.
providers/zenmux/models/openai/gpt-5.1-chat.toml	Adds per-model provider override to use `@ai-sdk/openai` with ZenMux `/api/v1`.
providers/zenmux/models/openai/gpt-5-codex.toml	Adds per-model provider override to use `@ai-sdk/openai` with ZenMux `/api/v1`.
providers/zenmux/models/moonshotai/kimi-k2.5.toml	Updates capability flag (`temperature = false`) to match runtime behavior.
providers/zenmux/models/minimax/minimax-m2.toml	Adds per-model provider override to use `@ai-sdk/anthropic` with ZenMux `/api/anthropic/v1`.
providers/zenmux/models/minimax/minimax-m2.7.toml	Adds per-model provider override to use `@ai-sdk/anthropic` with ZenMux `/api/anthropic/v1`.
providers/zenmux/models/minimax/minimax-m2.7-highspeed.toml	Adds per-model provider override to use `@ai-sdk/anthropic` with ZenMux `/api/anthropic/v1`.
providers/zenmux/models/minimax/minimax-m2.5.toml	Removes interleaving and adds provider override to `@ai-sdk/anthropic` on `/api/anthropic/v1`.
providers/zenmux/models/minimax/minimax-m2.5-lightning.toml	Removes interleaving and adds provider override to `@ai-sdk/anthropic` on `/api/anthropic/v1`.
providers/zenmux/models/minimax/minimax-m2.1.toml	Adds per-model provider override to use `@ai-sdk/anthropic` with ZenMux `/api/anthropic/v1`.
providers/zenmux/models/google/gemini-3-pro-image-preview.toml	Removes unsupported/deprecated model entry from ZenMux catalog.
providers/zenmux/models/anthropic/claude-sonnet-4.toml	Adds per-model provider override to use `@ai-sdk/anthropic` with ZenMux `/api/anthropic/v1`.
providers/zenmux/models/anthropic/claude-sonnet-4.6.toml	Adds per-model provider override to use `@ai-sdk/anthropic` with ZenMux `/api/anthropic/v1`.
providers/zenmux/models/anthropic/claude-sonnet-4.5.toml	Adds per-model provider override to use `@ai-sdk/anthropic` with ZenMux `/api/anthropic/v1`.
providers/zenmux/models/anthropic/claude-opus-4.toml	Updates output limit and adds provider override to `@ai-sdk/anthropic` on `/api/anthropic/v1`.
providers/zenmux/models/anthropic/claude-opus-4.6.toml	Adds per-model provider override to use `@ai-sdk/anthropic` with ZenMux `/api/anthropic/v1`.
providers/zenmux/models/anthropic/claude-opus-4.5.toml	Adds per-model provider override to use `@ai-sdk/anthropic` with ZenMux `/api/anthropic/v1`.
providers/zenmux/models/anthropic/claude-opus-4.1.toml	Adds per-model provider override to use `@ai-sdk/anthropic` with ZenMux `/api/anthropic/v1`.
providers/zenmux/models/anthropic/claude-haiku-4.5.toml	Adds per-model provider override to use `@ai-sdk/anthropic` with ZenMux `/api/anthropic/v1`.
providers/zenmux/models/anthropic/claude-3.7-sonnet.toml	Adds per-model provider override to use `@ai-sdk/anthropic` with ZenMux `/api/anthropic/v1`.
providers/zenmux/models/anthropic/claude-3.5-sonnet.toml	Removes deprecated model entry from ZenMux catalog.
providers/zenmux/models/anthropic/claude-3.5-haiku.toml	Adds per-model provider override to use `@ai-sdk/anthropic` with ZenMux `/api/anthropic/v1`.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

rekram1-node · 2026-03-23T19:59:58Z

Can u provide a bit more info for this change, evidence it works, etc

Kinda a big pr and idk this provider

llc1123 · 2026-03-24T02:16:53Z

@rekram1-node Added more detail to the PR description, but the short version is:

The biggest change here is switching ZenMux's default SDK route back to the OpenAI-compatible path. That affects every ZenMux model without its own [provider] override.
The reason is mostly a mix of cache behavior and protocol compatibility. In practice, the Anthropic Messages path has been worse for many non-Anthropic models, and I observed the difference especially clearly on GLM-5.
There is also a concrete compatibility issue for reasoning models like Kimi/GLM-style models: anomalyco/opencode#14638 documents that Kimi K2.5 breaks on the Anthropic SDK path because prior assistant reasoning_content gets dropped on replay.
Anthropic and MiniMax are explicitly routed through @ai-sdk/anthropic, and OpenAI models are explicitly routed through @ai-sdk/openai, so each family uses the SDK/API path that matches its protocol better.
The rest of the PR is metadata cleanup: fixing incorrect limits/capabilities, removing Claude 3.5 Sonnet because ZenMux already took it offline, and removing Gemini 3 Pro Image Preview because opencode does not need image generation support here.

For validation: I manually tested all updated model configs and confirmed they work correctly.

llc1123 · 2026-03-28T03:26:58Z

Is there anything blocking this PR? Hope this one can be merged asap.

…MiniMax M2.1/M2.5 - Add [provider] npm = "@ai-sdk/anthropic" to match -free variants - Remove incorrect [interleaved] section per PR anomalyco#1254 design decision MiniMax models should route through @ai-sdk/anthropic with per-model provider overrides, not rely on [interleaved] for reasoning content handling.

llc1123 added 2 commits March 23, 2026 21:48

feat(zenmux): route models through protocol-specific SDKs

8dc90a4

fix(zenmux): align model metadata with runtime support

0f95849

Copilot AI review requested due to automatic review settings March 23, 2026 15:14

Copilot started reviewing on behalf of llc1123 March 23, 2026 15:14 View session

Copilot AI reviewed Mar 23, 2026

View reviewed changes

rekram1-node merged commit 798538f into anomalyco:dev Mar 30, 2026
4 of 5 checks passed

This was referenced Mar 31, 2026

fix(opencode): MiniMax models missing [provider] override and incorrectly have [interleaved] LYY/models.dev#1

Open

fix(opencode): MiniMax models missing [provider] override and incorrectly have [interleaved] #1300

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat(zenmux): route models through protocol-specific SDKs#1254

feat(zenmux): route models through protocol-specific SDKs#1254
rekram1-node merged 2 commits intoanomalyco:devfrom
llc1123:dev

llc1123 commented Mar 23, 2026 •

edited

Loading

Uh oh!

Copilot AI left a comment

Uh oh!

rekram1-node commented Mar 23, 2026

Uh oh!

llc1123 commented Mar 24, 2026 •

edited

Loading

Uh oh!

llc1123 commented Mar 28, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

llc1123 commented Mar 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Why

Validation

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

rekram1-node commented Mar 23, 2026

Uh oh!

llc1123 commented Mar 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

llc1123 commented Mar 28, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

llc1123 commented Mar 23, 2026 •

edited

Loading

llc1123 commented Mar 24, 2026 •

edited

Loading