Support for reasoning_effort parameter in UI

I couldn't find references of reasoning effort in the code at the moment, but multiple models support it, including OpenAI models, Gemini, Deepseek, and more.
Would be ideal to have it implemented in UI for supported models, similar to how /models command works.

In a similar vein, supporting things like ultrathink from claude code would be great to dynamically increase the reasoning effort.

Might need some tuning thresholds for specific popular models, and perhaps just a % of the limit for other models which don't have their own hardcoded.

Since claude has 32k as the max thinking, a good starting place could be mirroring it with:

HIGHEST - 100% thinking budget
MIDDLE - 31.25% thinking budget
BASIC - 12.5% thinking budget
DEFAULT - 0% of course

Reference for claude code's thinking token thresholds and keywords:

HIGHEST (token: 31,999)

    "think harder"

    "think intensely"

    "think longer"

    "think really hard"

    "think super hard"

    "think very hard"

    "ultrathink"

MIDDLE (token: 10,000)

    "think about it"

    "think a lot"

    "think deeply"

    "think hard"

    "think more"

    "megathink"

BASIC (token: 4,000)

    "think"

NONE (token: 0)

    (empty)

Need a flag

I'm also looking for such a flag. Here is an example for llm by Simon Willison

llm -m openai/o3 "ping"

llm -m openai/o3 -o reasoning_effort high "ping"

Why I want this

from Aider leaderboard

Think about the gap between 65% (Sonnet 4, ultrathink) and 84% (o3-high) 🧠

from Codex

Also, here is PR in Codex - openai/codex#314

Can be considered fixed since d8dc23b ?

Not sure that works:

"provider": {
    "openai": {
      "models": {
        "gpt-5": {
          "options": {
            "reasoningEffort": "high"
          }
        }
      }
    },
    "anthropic": {
      "models": {
        "claude-sonnet-4-20250514": {
          "options": {
            "thinking": {
              "type": "enabled",
              "budget_tokens": 32000
            }
           }
        }
      }
    }
  },

When using Claude Sonnet I now get:

AI_UnsupportedFunctionalityError: 'thinking requires a budget' functionality not supported.

The error goes away if I remove the options. Maybe I'm passing them wrong?

EDIT: Nevermind, just change budget_tokens to budgetTokens. That's what the ai package used by opencode expects.

This is the config I use and it's verified working in openai dashboard logs:

{
    "$schema": "https://opencode.ai/config.json",
    "provider": {
        "openai": {
            "models": {
                "gpt-5": {
                    "options": {
                        "reasoningSummary": "auto",
                        "include": [
                            "reasoning.encrypted_content"
                        ]
                    }
                }
            }
        }
    },
    "agent": {
        "gpt-5-max": {
            "model": "openai/gpt-5",
            "description": "Agent that uses high reasoning effort for complex problems",
            "reasoningEffort": "high",
            "textVerbosity": "low"
        }
    }
}

I think this issue can be closed since there at least 2 possible ways to enable high reasoningEffort.

I think this issue can be closed since there at least 2 possible ways to enable high reasoningEffort.

I believe the core of this issue is different. You've describe ways to persistently enable a thinking budget ("always use high reasoning effort") but this issue is asking for a way to set it dynamically ("use high reasoning effort just for this next request"). Claude Code does it through env variables (MAX_THINKING_TOKENS) and through message keywords ("ultrathink").
There isn't currently a way to set reasoning effort for a single request, like there is in Claude Code.

I would like to mention this problem, related to private endpoint, where reasoningEffort is rejected and is epxected reasoning_effort. #1859 (comment)

[automated] Closing due to 90+ days of inactivity. Feel free to reopen if you still need this!

no reasoning effort control is the only reason I don't use opencode. gpt-5.2-xhigh is my main model and there is no way to use it with my codex subscription through opencode

I open sourced my config.
I spent a lot of time configuring my open code setup. up. I'm still not sure it's perfect but it works good at this point.

The trick here is to be aware that there's a big difference between the provider and the agents.

providers:
https://github.com/pascalandy/dotfiles/blob/main/dot_config/opencode/opencode.json.tmpl

agents:
https://github.com/pascalandy/dotfiles/blob/main/dot_config/opencode/config.json

feat: change toxikat toolOutput color to anomalyco#450

why this is closed? why is it hard to support choosing reasoning effort from the UI? You can only set it first time you start using model, why can't we have a menu option to just switch it in UI?

OpenCode stores model thinking efforts and user preferences in:

macOS and Linux: ~/.local/state/opencode
Windows: %USERPROFILE%\.local\state\opencode

See #6156 (comment)

Ok so apparently this is being kept a secret (I have no other explanation why everyone keeps saying configs e.t.c),

but reasoning in opencode is switched with ctrl+T.

You are welcome everyone stumbling upon it.

but reasoning in opencode is switched with ctrl+T.

https://opencode.ai/docs/tui/#thinking

why is this closed? here's how llama.cpp manages fine grained control: ggml-org/llama.cpp#21445

Navigation Menu

Repository navigation

Support for reasoning_effort parameter in UI #450

Need a flag

Why I want this

from Codex

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Support for reasoning_effort parameter in UI #450

Description

Activity

johnpyp commented on Jun 30, 2025

pascalandy commented on Jul 12, 2025

Need a flag

Why I want this

from Codex

asafyish commented on Aug 11, 2025

jamesponddotco commented on Aug 15, 2025

asafyish commented on Aug 16, 2025

carterworks commented on Aug 21, 2025

f5epcim commented on Aug 22, 2025

rekram1-node commented on Dec 28, 2025

jackfruitsandwich commented on Jan 22, 2026

pascalandy commented on Jan 22, 2026

maxevilmind commented on May 15, 2026

sekedus commented on Jun 10, 2026

maxevilmind commented on Jun 18, 2026

sekedus commented on Jun 18, 2026

disarticulate commented on Jun 26, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions