Fix OpenCode "Model Does Not Support Image Input" — GPT & Claude Vision Not Working

"attachment": true,
"modalities": {
  "input": ["text", "image"],
  "output": ["text"]
}

What This Problem Looks Like

• You switch to a GPT or Claude model in OpenCode, drag in a screenshot, and ask "describe this image."
• The reply is a variation of "I can't read images", "image input is not supported for this model", or the model answers about a filename (screenshot.png) instead of the actual content.
• The same image works fine when you paste it into ChatGPT or Claude directly.

This is not a model limitation and not a broken API key. gpt-5.5, claude-sonnet-4-6, claude-opus-4-7, claude-opus-4-8, and claude-haiku-4-5 all accept image input over the OpenAI-compatible Chat Completions API. The image simply never left OpenCode.

Why It Happens

OpenCode decides per model whether an attachment can be sent, and whether that attachment may be an image. It reads this from your provider config in opencode.jsonc. If your model entry looks like this:

"gpt-5.5": {
  "name": "gpt-5.5"
}

…then OpenCode has no signal that the model takes images. Depending on the build, it either refuses to attach the file or passes only the attachment name into the prompt text. The model receives "[attachment: diagram.png]" — a string, not a picture — and correctly replies that it wasn't given an image.

The fix is to tell OpenCode two things explicitly:

Once both are set, OpenCode base64-encodes the image and sends it as proper vision input — and the model reads it.

How to Fix It

1. Locate your opencode.jsonc

2. Add attachment + modalities to each vision model

Here is a working provider block for GPT and Claude on an OpenAI-compatible endpoint:

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "apimaster": {
      "name": "APIMaster.ai",
      "npm": "@ai-sdk/openai-compatible",
      "options": {
        "baseURL": "https://apimaster.ai/v1"
      },
      "models": {
        "gpt-5.5": {
          "name": "gpt-5.5",
          "attachment": true,
          "modalities": {
            "input": ["text", "image"],
            "output": ["text"]
          }
        },
        "claude-sonnet-4-6": {
          "name": "claude-sonnet-4-6",
          "attachment": true,
          "modalities": {
            "input": ["text", "image"],
            "output": ["text"]
          }
        }
      }
    }
  }
}

The full example covering claude-opus-4-7, claude-opus-4-8, claude-haiku-4-5, and reasoning variants is in the OpenCode setup guide. You can enable image input for just one model — only that model needs the two fields.

3. Use the correct base URL

The OpenAI-compatible host is:

https://apimaster.ai/v1

Do not use https://api.apimaster.ai/v1 — the api. prefix host can cause TLS / connection failures.

4. Restart OpenCode

Config changes are loaded at startup. Fully quit and reopen OpenCode, or at minimum start a new session, or the old (image-blind) config stays active.

5. Prefer local upload / base64 over public URLs

Upload a local PNG/JPG in OpenCode — it gets converted to a base64 data URL, which is the reliable path. Passing a public image URL directly can fail, because the upstream has to download it and may hit format, size, MIME, or hotlink limits:

Error while downloading file. Upstream status code: 400.

If you see that, switch to a local file / base64 data URL instead of a remote URL.

Test That It Works

In OpenCode

Switch to apimaster/gpt-5.5 or apimaster/claude-sonnet-4-6, upload an image, and ask:

Describe the main content of this image.

A correct config returns a real description. If it still only sees a file name, the attachment / modalities fields are missing or OpenCode wasn't restarted.

Directly via the API (isolate the problem)

To confirm the endpoint itself handles vision — separating an OpenCode config issue from a provider issue — send a base64 image. Never hardcode a real key; use an environment variable:

export APIMASTER_API_KEY="your APIMaster API key"

curl https://apimaster.ai/v1/chat/completions \
  -H "Authorization: Bearer $APIMASTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.5",
    "messages": [
      {
        "role": "user",
        "content": [
          { "type": "text", "text": "Describe this image." },
          {
            "type": "image_url",
            "image_url": { "url": "data:image/png;base64,YOUR_IMAGE_BASE64" }
          }
        ]
      }
    ],
    "max_tokens": 1000
  }'

Swap model to claude-sonnet-4-6 to test Claude the same way. A 200 response that describes the image means vision works at the API layer — so any remaining failure is on the OpenCode config side.

Keep Your API Key Safe

• Never put the API key inside opencode.jsonc. That file holds only the provider, models, and capabilities.
• Configure the key through OpenCode's /connect flow, the Connect Provider UI, or an environment variable.
• Don't paste the key into chats, screenshots, issues, or public repos. If it leaks, revoke and regenerate it in your provider console immediately.

How APIMaster Helps

If you want a single OpenAI-compatible endpoint where GPT and Claude vision "just work" with the config above, APIMaster is built for exactly this:

From $1 top-up, pay-as-you-go, no subscription. Full step-by-step OpenCode setup — including image input — is in the OpenCode configuration guide.

Register on APIMaster →

Related Guides

• OpenCode configuration guide — full provider + vision setup
• How to Fix Invalid API Key (OpenAI / Claude) — 401 auth errors
• Fix "api error 400 content blocked" — moderation, not vision
• All API error fix guides — full index

FAQ

Does OpenCode support image input at all? Yes. OpenCode can send images to any vision-capable model, but you must declare attachment: true and modalities.input: ["text", "image"] for that model in opencode.jsonc. Without the declaration it sends only the file name.

Which APIMaster models can read images in OpenCode? gpt-5.5, claude-sonnet-4-6, claude-opus-4-7, claude-opus-4-8, and claude-haiku-4-5 all accept image input. Add the two capability fields to each one you want to use with images.

Why does the model only see the file name? OpenCode didn't convert the attachment into image input — it put the attachment name into the prompt text instead. This is the tell-tale sign the attachment / modalities fields are missing.

I added the fields and it still fails. Now what? Fully restart OpenCode, start a new session, and confirm the active model is the one you configured. Then test the endpoint directly with the base64 curl above to check whether the problem is OpenCode or the provider.

Why did a public image URL error out? The upstream has to download remote URLs and can be blocked by size, format, MIME, or hotlink protection (Error while downloading file. Upstream status code: 400.). Use a local upload / base64 data URL instead.