All paid OpenCode Zen models fail with 'Upstream request failed' — free models work

[vicmuchina]

Description

All paid OpenCode Zen models (e.g., opencode/MiniMax-M3, opencode/deepseek-v4-flash, etc.) fail with "Error from provider (Console): Upstream request failed". Free Zen models (e.g., opencode/big-pickle, opencode/deepseek-v4-flash-free) work fine. OpenCode Go models also work fine.

Environment

• OpenCode version: latest (just upgraded)
• OS: Linux (Pop!_OS)
• Zen balance: $19.60 (from $52 total deposits, see payment history below)
• Free models work: ✅ yes
• Paid models fail: ❌ all of them

Payment History

Date	Amount
Jul 12, 1:13 PM	$10.00
Jul 7, 10:47 AM	$12.00
Jun 9, 5:18 PM	$5.00
May 4, 3:48 PM	$10.00
Apr 30, 10:16 PM	$10.00
Apr 6, 1:38 PM	$5.00

Total deposited: $52.00 | Remaining balance: $19.60

Steps to Reproduce

1. Connect OpenCode Zen provider via /connect
2. Select any paid Zen model (e.g., opencode/MiniMax-M3)
3. Send a message
4. Observe error: "Error from provider (Console): Upstream request failed"

Suspected Root Cause

This looks similar to issue #28129 where upstream provider accounts used by the Zen backend are depleted or restricted. Free models use different upstream providers (NVIDIA, Cohere, etc.) which is why they still work. The Zen backend's upstream accounts for paid models need to be refilled.

Logs

Available at ~/.local/share/opencode/log/ on request.

cc @StarpTech @fwang

[matiasfha]

Same issue here. Adding more diagnostic data:

Environment: macOS (Darwin arm64), OpenCode 1.18.3 (just upgraded from 1.17.11), Zen provider connected via API key from opencode.ai/auth

Balance: Positive ($19.60+ remaining)

Key findings:

1. Direct curl to https://opencode.ai/zen/v1/chat/completions with a paid model (both deepseek-v4-flash and claude-sonnet-5) works fine with HTTP 200 — the upstream Zen API itself is healthy.
2. The same paid model fails when called through the opencode client — same API key, same model.
3. Free models (deepseek-v4-flash-free) work fine through both curl and the client.
4. The client logs show: failed to load OpenCode provider config — HttpClientError: StatusCode: non 2xx status code (401 GET https://console.opencode.ai/api/config) — the Zen API key (sk-...) is rejected by the Console API.

So the flow seems to be:

• Free models (cost=0) → skip billing validation → call Zen API directly → ✅ works
• Paid models (cost>0) → client tries to validate via console.opencode.ai/api/config → gets 401 → returns "Upstream request failed" without ever hitting the Zen API

This matches the theory in the issue — the routing/billing validation layer between the client and the Zen API is broken, not the Zen API itself. cc @fwang @StarpTech

[StarpTech]

Hi, in case of issues, please upgrade to the latest version and share the information behind /debug command. Thank you!

[matiasfha]

Running latest versions

Version: 1.18.3 (latest)
Date: 2026-07-16T22:04:10.321Z
OS: macOS 24.6.0 (arm64)
Terminal: xterm-kitty
Session ID: n/a
Model: opencode/deepseek-v4-flash

[StarpTech]

You need to run /debug in the next session where you experience the issue.

[matiasfha]

Thas was run starting a session

This is for a new session

Version: 1.18.3 (latest)
Date: 2026-07-17T00:57:18.243Z
OS: macOS 24.6.0 (arm64)
Terminal: xterm-kitty
Session ID: ses_0926c3222ffe4OtoKTFhrv44GW
Model: opencode/deepseek-v4-flash

[royhananwar]

got same issue but with opencode go kimi 3 model

Version: 1.18.3 (latest)
Date: 2026-07-17T10:25:08.515Z
OS: macOS 24.6.0 (arm64)
Terminal: iTerm.app 3.6.10
Session ID: ses_090653d63ffeUFa2PKpVNa6AL9
Model: opencode-go/kimi-k3

[StarpTech]

@royhananwar Hello, after inspection, we see the following provider error

invalid temperature: only 1 is allowed for this model

Do you overwrite it? I tried Kimi K3, and it works fine.

[alpharald]

Kimi K3 seems to fail for me with the 'Error from provider (Console Go): Upstream request failed' message whenever I am connected to the xcode MCP.

opencode.config:

...
  "mcp": 
  {
    "xcode": {
      "type": "local",
      "command": [
        "xcrun",
        "mcpbridge"
      ]
    }
}
...

Debug Info - Previously I only started the sessions with the MCP connected where it failed, started this session without MCP and it was successful. Connected the MCP mid-session and it failed. Then I switched to DS4Flash in the same session with the MCP still connected as a test and it worked. Only Kimi K3 seems to have issues with it:

Version: 1.18.3 (latest)
Date: 2026-07-18T05:23:36.815Z
OS: macOS 25.5.0 (arm64)
Terminal: ghostty 1.3.1
Session ID: ses_08c5cd638ffejoBNfKDRgLju3K
Model: opencode-go/deepseek-v4-flash

I also just realized the macOS version is stated as 25.5.0 in the debug info but it actually is 26.5.2.

[royhananwar]

@StarpTech Hey, I didn’t change or overwrite anything.

I tried on 2 different environments, one is my laptop and the second one is my server

I just installed opencode on my server and tried kimi 3 have the same results like in my laptop.

any idea why the temperature is not 1 ?

I actually don’t know where to change the temperature model 😅