OpenCode requests unaffordable output budget for gpt-5.4 in normal serve/web sessions and fails with 402

[rosspeoples]

Description

When OpenCode is configured to use gpt-5.4 through an OpenAI-compatible provider path, normal prompts in serve/web mode can fail immediately because OpenCode requests an output budget that is too large for the available model credits.

In our repro, even a trivial prompt failed because OpenCode requested up to 32000 output tokens. The provider returned a 402 error indicating the request required more credits or fewer max tokens.

Representative error returned through our provider gateway:

This request requires more credits, or fewer max_tokens. You requested up to 32000 tokens, but can only afford 8421.

Important control case:

• the same model (gpt-5.4) worked correctly when we called the provider gateway directly and explicitly capped max_tokens to a smaller value like 512
• so the model itself was not unusable in our environment; the failure was specifically OpenCode's default request shape/output budget for that model in serve/web usage

This made gpt-5.4 unusable as the default model in our Build-integrated OpenCode deployment.

Plugins

Build-local plugins:

• secret-guard.js
• workspace-shell.js

OpenCode version

1.3.15

Steps to reproduce

1. Run OpenCode in serve/web mode.
2. Configure an OpenAI-compatible provider path for gpt-5.4.
3. Make gpt-5.4 the default model.
4. Send a trivial prompt like: What is 2 plus 2? Reply in one sentence.
5. Observe the request fails immediately with a provider-side 402 affordability error.
6. As a control, call the same provider directly with the same model and a capped max_tokens like 512.
7. Observe the direct provider call succeeds.

Screenshot and/or share link

No share link. Representative session error from OpenCode:

litellm.APIError: APIError: OpenrouterException - {"error":{"message":"This request requires more credits, or fewer max_tokens. You requested up to 32000 tokens, but can only afford 8421." ... }}

Operating System

Linux (Kubernetes pod on Fedora CoreOS host)

Terminal

Web UI / serve mode behind a reverse proxy launcher

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• This request requires more credits, or fewer max_tokens. You requested up to 32000 tokens, but can only afford 0. To increase, visit https://openrouter.ai/settings/keys and create a key with a higher monthly limit #12219: Same exact error message — "This request requires more credits, or fewer max_tokens. You requested up to 32000 tokens" — via OpenRouter with a free/limited-credit model
• Custom provider (LM Studio) ignores limit.output config, hardcodes max_tokens to 32000 #20078: Same root cause — OpenCode hardcodes max_tokens to 32000 for custom OpenAI-compatible providers, ignoring limit.output config
• @ai-sdk/openai-compatible max_tokens error for GPT 5.x #5421: Related — max_tokens parameter errors with GPT 5.x models via @ai-sdk/openai-compatible

If your issue is distinct from the above, please add context clarifying how it differs.