[Bug]: OpenCode Go grok-4.5 returns HTTP 503 while deepseek-v4-flash works

Description

OpenCode Go consistently returns HTTP 503 when calling grok-4.5 through the officially documented Chat Completions endpoint.

• Endpoint: https://opencode.ai/zen/go/v1/chat/completions
• Model: grok-4.5
• Temperature: 1
• Top-p: 0.95
• Streaming: enabled

Actual response:

HTTP 503

{"error":{"type":"server_error","message":"Upstream request failed: Endpoint is unavailable."}}

Response details:

• Time: August 6, 2026, 15:40 UTC
• CF-Ray: a26f233459ab3f1e-MIA
• CF placement: remote-ORD

The issue was reproduced with both Cherry Studio using its OpenAI-compatible client and Claude Code 2.1.223 through CC Switch.

As a control test, deepseek-v4-flash works with the same OpenCode Go subscription, API key, base URL, and Chat Completions format while grok-4.5 continues to return HTTP 503.

This appears to be model-specific rather than a complete OpenCode Go outage. It may be related to #36889 and #37815, but differs because another OpenCode Go model remains available at the same time.

Plugins

None. The OpenCode Go API was tested through external OpenAI-compatible clients.

OpenCode version

N/A — the OpenCode Go API endpoint was tested directly. Secondary reproduction used Claude Code 2.1.223.

Steps to reproduce

1. Configure an OpenAI-compatible client with an active OpenCode Go API key.
2. Send a simple message such as Hello to https://opencode.ai/zen/go/v1/chat/completions using model grok-4.5.
3. Observe HTTP 503: Upstream request failed: Endpoint is unavailable.
4. Change only the model to deepseek-v4-flash, keeping the same API key, endpoint, and API format.
5. Observe that deepseek-v4-flash responds successfully.

Screenshot and/or share link

The complete HTTP response and relevant Cloudflare request identifiers are included in the Description above.

Operating System

Windows 11

Terminal

Windows Terminal with PowerShell