model "claude-opus-4.6" not supported for vision

Description

The Claude Opus 4.6 is latest anthropic model. I expected that it would support vision capabilities:
https://ai.azure.com/catalog/models/claude-opus-4-6

Plugins

opencode-beads

OpenCode version

1.2.6

Steps to reproduce

Start opencode
Select claude-opus-4.6 model from Github Copilot provider
Drop an image
See the message "model "claude-opus-4.6" not supported for vision"

Screenshot and/or share link

(I'm pasting Figma link here, but dropping screenshot provides the same result)

Operating System

macOS Tahoe 26.2

Terminal

iTerm2

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

[Github Copilot] Anthropic models can't see images? #5291: Same error message pattern (model "claude-*" not supported for vision) with Anthropic models via the GitHub Copilot provider — the prior issue covered claude-haiku/sonnet/opus 4.5 and was closed with 18 comments. This may be a continuation of the same underlying problem with the newer 4.6 model.

Can you try to add this (adding modalities)

#5291 (comment)

@rekram1-node

@rekram1-node I added modalities which did not help. I also have "Editor Preview Features" disabled in my org. I'm talking to org admin to enable that

Also, what's weird to me is that vision works for "GPT-5.2", but I can't get Anthropic models to work. Same provider

Here is my config;

{
  "$schema": "https://opencode.ai/config.json",
  "default_agent": "mayor",
  "provider": {
    "github-copilot": {
      "models": {
        "claude-opus-4.6": {
          "modalities": {
            "input": ["text", "image"],
            "output": ["text"]
          }
        },
        "claude-opus-4.5": {
          "modalities": {
            "input": ["text", "image"],
            "output": ["text"]
          }
        },
        "claude-opus-41": {
          "modalities": {
            "input": ["text", "image"],
            "output": ["text"]
          }
        },
        "claude-sonnet-4.6": {
          "modalities": {
            "input": ["text", "image"],
            "output": ["text"]
          }
        },
        "claude-sonnet-4.5": {
          "modalities": {
            "input": ["text", "image"],
            "output": ["text"]
          }
        },
        "claude-sonnet-4": {
          "modalities": {
            "input": ["text", "image"],
            "output": ["text"]
          }
        },
        "claude-haiku-4.5": {
          "modalities": {
            "input": ["text", "image"],
            "output": ["text"]
          }
        }
      }
    }
  },
  "plugin": ["opencode-beads"],
  "skills": {
    "paths": ["{env:HOME}/.config/opencode/skills-private"]
  },
  "permission": {
    "bash": {
      "*": "allow",
      "bd *": "ask",
      "bd ready*": "allow",
      "bd list*": "allow",
      "bd show*": "allow"
    },
    "skill": {
      "*": "allow"
    }
  },
  "agent": {
    "github": {
      "description": "GitHub-focused agent for repository management, PRs, issues, and automation",
      "prompt": "You are a GitHub expert. Help with repository management, pull requests, issue triage, branch management, and GitHub automation. Use GitHub tools to perform operations. Always provide clear feedback about what was done and next steps.",
      "tools": {
        "bash": true,
        "edit": true,
        "write": true,
        "read": true,
        "grep": true,
        "glob": true,
        "github_*": true,
        "gh_*": true
      }
    }
  },
  "instructions": [".opencode/github-rules.md"],
  "command": {
    "village:work": {
      "description": "Trigger the village work loop for your role",
      "template": "Claim the next bead assigned to you and start working on it.\n\nUse this workflow:\n1. Claim deterministically by calling `village_claim` (single in_progress guard):\n   - `existing in_progress: <id> | ...` => continue that bead\n   - `claimed: <id> | ...` => start that bead\n   - `no ready beads for worker` / `no ready beads for overseer` => report and wait\n2. Read the bead details and handoff packet: `bd show <id> --json`\n3. Load skills listed under `## Skills`\n4. Do the work for your role (worker: implement + local commit; overseer: run checks + approve/return)\n5. Update bead status/comments and repeat from step 1\n\nImportant: do not claim work via `bd ready` + `bd update ... in_progress`; use `village_claim` so the single in_progress guard is enforced."
    },
    "village:orphans": {
      "description": "Report and optionally fix unassigned village beads",
      "template": "Find orphan and suspect-assignee beads from open + in_progress work.\n\nBehavior:\n1. Graceful preflight:\n   - If `bd` is unavailable, return: `bd not available; cannot inspect beads.`\n   - If beads DB is missing in this repo, return: `No .beads database found; nothing to inspect.`\n2. Read data:\n   - Run `bd list --status open --json`\n   - Run `bd list --status in_progress --json`\n   - Combine both lists.\n3. Filter and track scope:\n   - `ignored_epics`: `issue_type == \"epic\"`.\n   - `scanned_non_epic`: all non-epic items.\n4. Build groups from `scanned_non_epic` only:\n   - `orphans`: assignee is missing/empty.\n   - `suspect`: assignee is present but not `worker` or `overseer`.\n5. Build ignored-epic observability groups:\n   - `ignored_epics_unassigned`: epics with missing/empty assignee.\n   - `ignored_epics_suspect`: epics whose assignee is present but not `worker` or `overseer`.\n   - `ignored_epics_attention`: union of the two groups above (deduplicated).\n6. Report output (concise):\n   - Print counts: `Scanned (non-epic): <n> | Ignored epics: <n> | Ignored epics (unassigned): <n> | Orphans: <n> | Suspect: <n>`.\n   - Print up to 20 total issue rows for non-epic `orphans` + `suspect` with: `id | title | status | assignee`.\n   - If `ignored_epics_attention` is non-empty, print a separate `Ignored epics` section (do not mix with orphan/suspect rows), listing up to 5 rows as: `id | title | status | assignee | reason` where reason is `unassigned` or `suspect assignee`.\n\nArgs:\n- `/village:orphans` -> report only.\n- `/village:orphans fix` -> after report, auto-assign only non-epic `orphans` (never `suspect`, never epics).\n\nFix mode rules (`fix` only):\n- For each non-epic orphan, inspect lowercase title + description.\n- If it matches any keyword `review|verify|verification|check|checks|overseer|approve`, assign `overseer`.\n- Otherwise assign `worker`.\n- Apply with `bd update <id> --assignee <worker|overseer>`.\n- Print a short `Applied` section listing each updated issue and chosen assignee.\n\nSafety:\n- Never reassign issues that already have an assignee.\n- Never modify epics."
    }
  }
}

Seems like a regression since the official copilot auth? ie: now opencode has its own client id, github can limit models/capabilities for it - see also 5.3-codex.

Would be great if this could be confirmed.

Just fyi, copilot cli has same error for opus-4.6 and sonnet-4.6 github/copilot-cli#1555

However, image does work with sonnet-4.5 in copilot cli, don't know how

I am hitting this issue too. I'm getting it on the desktop mac app.

Workaround: deny image file reads in permission config.

{
  "$schema": "https://opencode.ai/config.json",
  "permission": {
    "read": {
      "*": "allow",
      "*.png": "deny",
      "*.jpg": "deny",
      "*.jpeg": "deny",
      "*.gif": "deny",
      "*.webp": "deny",
      "*.svg": "deny",
      "*.bmp": "deny",
      "*.ico": "deny"
    }
  }
}

Workaround: deny image file reads in permission config.

{
"$schema": "https://opencode.ai/config.json",
"permission": {
"read": {
"": "allow",
".png": "deny",
".jpg": "deny",
".jpeg": "deny",
".gif": "deny",
".webp": "deny",
".svg": "deny",
".bmp": "deny",
"*.ico": "deny"
}
}
}

How is denying image reads a workaround for the problem which is that images cannot be read for a model which does support it?

Workaround: deny image file reads in permission config.
{
"$schema": "https://opencode.ai/config.json",
"permission": {
"read": {
"": "allow",
".png": "deny",
".jpg": "deny",
".jpeg": "deny",
".gif": "deny",
".webp": "deny",
".svg": "deny",
".bmp": "deny",
"*.ico": "deny"
}
}
}

How is denying image reads a workaround for the problem which is that images cannot be read for a model which does support it?

You're right, that doesn't really help with the core issue. I was trying to avoid a secondary problem where opencode becomes unusable after an image read error, but denying image reads entirely defeats the purpose if you actually need to read images. My bad.

Our workaround is remove to image modality from the unsupported models and provide an image-looker subagent with gpt-5.2 (description and prompt nicked from oh-my-opencode). The main agent will use it automatically or ask if you want to use it or tell you you can use it - but it's an improvement.

{
  "$schema": "https://opencode.ai/config.json",
  "agent": {
    "image-looker": {
      "description": "Analyze image files. Describes visual content. Use when you need analyzed/extracted data rather than literal file contents.",
      "mode": "subagent",
      "model": "github-copilot/gpt-5.2",
      "prompt": "You interpret image files that cannot be read as plain text.\n\nYour job: examine the attached file and extract ONLY what was requested.\n\nWhen to use you:\n- Media files the Read tool cannot interpret\n- Extracting specific information or summaries from documents\n- Describing visual content in images or diagrams\n- When analyzed/extracted data is needed, not raw file contents\n\nWhen NOT to use you:\n- Source code or plain text files needing exact contents (use Read)\n- Files that need editing afterward (need literal content from Read)\n- Simple file reading where no interpretation is needed\n\nHow you work:\n1. Receive a file path and a goal describing what to extract\n2. Read and analyze the file deeply\n3. Return ONLY the relevant extracted information\n4. The main agent never processes the raw file - you save context tokens\n\nFor images: describe layouts, UI elements, text, diagrams, charts\nFor diagrams: explain relationships, flows, architecture depicted\n\nResponse rules:\n- Return extracted information directly, no preamble\n- If info not found, state clearly what's missing\n- Match the language of the request\n- Be thorough on the goal, concise on everything else\n\nYour output goes straight to the main agent for continued work.",
      "temperature": 0.1
    },
  },
  "provider": {
    "github-copilot": {
      "models": {
        "claude-haiku-4.5": {
          "modalities": {
            "input": ["text"],
            "output": ["text"]
          }
        },
        "claude-opus-4.5": {
          "modalities": {
            "input": ["text"],
            "output": ["text"]
          }
        },
        "claude-opus-4.6": {
          "modalities": {
            "input": ["text"],
            "output": ["text"]
          }
        },
        "claude-sonnet-4.5": {
          "modalities": {
            "input": ["text"],
            "output": ["text"]
          }
        },
        "claude-sonnet-4.6": {
          "modalities": {
            "input": ["text"],
            "output": ["text"]
          }
        },
        "gemini-2.5-pro": {
          "modalities": {
            "input": ["text"],
            "output": ["text"]
          }
        }
      }
    }
  }
}

Enabling "Editor Preview Features" helped me 👍

@technoch1ef

Enabling "Editor Preview Features" helped me 👍

@technoch1ef just for clarity, could you confirm that it works for opus 4.6 in opencode with "Editor Preview Features"

Our org won't turn on preview features atm so it would be really appreciated, thanks

Confirming: enabling Editor Preview Features in Copilot settings made the image vision work with Opus

Another confirmation: it's configurable from the org's settings.

I do not have permission to update the Github Copilot org's settings to enable the "Editor Preview Features", and the company does not want to : is there another solution on opencode side to fix this please ?

Root cause analysis

I traced this in the opencode source on dev. The bug is in how the GitHub Copilot plugin rebuilds model capabilities and overwrites user config.

Load order in packages/opencode/src/provider/provider.ts:

Lines 1107-1161: config from opencode.json is merged into the model database. A user override like:
```
"github-copilot": {
  "models": {
    "claude-sonnet-4.6": { "modalities": { "input": ["text"], "output": ["text"] } }
  }
}
```
correctly sets capabilities.input.image = false (via model.modalities?.input?.includes("image") ?? ...).
Lines 1285-1310: the plugin models() hook runs after and fully replaces provider.models:
```
provider.models = await models(provider, { auth: pluginAuth })
```
For Copilot this calls CopilotModels.get() in packages/opencode/src/plugin/github-copilot/models.ts, which runs build() on every model.

The actual bug is in build() (around lines 75-85):

capabilities: {
  temperature: prev?.capabilities.temperature ?? true,
  reasoning:   prev?.capabilities.reasoning   ?? reasoning,
  attachment:  prev?.capabilities.attachment  ?? true,
  toolcall:    remote.capabilities.supports.tool_calls,
  input: {
    text:  true,
    audio: false,
    image,           // <-- derived from remote, prev is IGNORED
    video: false,
    pdf:   false,
  },

Every other capability honors prev?.capabilities.X (preserving config overrides), but input.image is unconditionally written from the Copilot /models response. The user's explicit modalities: { input: ["text"] } is silently destroyed.

Why this is wrong: the Copilot /models endpoint reports supports.vision: true at the model level for Claude, but "vision enabled" is enforced at the organization level via "Editor Preview Features". opencode conflates model-level capability with org-level entitlement, so for orgs that can't enable previews there is no working configuration.

Why the session stays broken after the error: unsupportedParts() in packages/opencode/src/provider/transform.ts (lines 267-303) strips/rewrites image parts only when model.capabilities.input.image === false. Since Copilot+Claude reports it as true, the image stays in conversation history, and every subsequent turn retriggers the same 400 from the Copilot API. Switching to a provider the user has configured as text-only finally lets unsupportedParts rewrite the image into a text error, which is why that workaround unblocks the session.

Proposed fix

Minimal, one-line change in build() in packages/opencode/src/plugin/github-copilot/models.ts:

input: {
  text:  true,
  audio: false,
  image: prev?.capabilities.input.image ?? image,
  video: false,
  pdf:   false,
},

This makes the plugin consistent with how it treats temperature, reasoning, and attachment: user/config value wins, remote default is the fallback. Users whose org has not enabled vision previews can then force text-only and opencode's unsupportedParts will correctly strip incoming images before they are sent.

A more general fix would be to re-apply config overrides after the plugin models() hook in provider.ts:1285-1310, which would protect any plugin-derived provider from silently overwriting user config in the same way.

Happy to open a PR if a maintainer confirms the approach.

@stevoland

@stevoland Confirming: enabling Editor Preview Features in Copilot settings made the image vision work with Opus

@rekram1-node

@rekram1-node may I ask why you closed this please ? many of us have their github organization managed at corporate level and cannot use the workaround to enable preview features. Can you please take a look at the proposal ? Thank you for your time

Can you try to add this (adding modalities)

#5291 (comment)

Worked with open models as well, before 1.15 image capability was derived by the model itself, but it looks like it should be explicited now

EDIT: Server is llamacpp with model router (maybe this was my issue, previously i was using plain llamacpo so the model config was preloaded instead of being loaded on request)

Navigation Menu

Repository navigation

model "claude-opus-4.6" not supported for vision #14289

Description

Plugins

OpenCode version

Steps to reproduce

Screenshot and/or share link

Operating System

Terminal

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

model "claude-opus-4.6" not supported for vision #14289

Description

Description

Plugins

OpenCode version

Steps to reproduce

Screenshot and/or share link

Operating System

Terminal

Activity

github-actions commented on Feb 19, 2026

R44VC0RP commented on Feb 19, 2026

technoch1ef commented on Feb 19, 2026

technoch1ef commented on Feb 19, 2026

stevoland commented on Feb 23, 2026

stevoland commented on Feb 23, 2026

malachibazar commented on Feb 23, 2026

soyukke commented on Feb 25, 2026

jamesforwardnwboxed commented on Feb 25, 2026

soyukke commented on Feb 25, 2026

stevoland commented on Feb 25, 2026

technoch1ef commented on Feb 26, 2026

stevoland commented on Feb 27, 2026

Leksat commented on Mar 13, 2026

juliobellano commented on Mar 30, 2026

yogeek commented on Apr 15, 2026

yogeek commented on Apr 17, 2026

technoch1ef commented on Apr 21, 2026

yogeek commented on Jun 18, 2026

andreasntr commented on Aug 30, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions