Codex App: GPT-5.3 Codex Spark fails with "Unsupported parameter: reasoning.summary"

This happens for 5.3-Codex Spark on the Windows Desktop App as well.

See Feedback ID: 019f488e-6564-7222-8c23-d9cd75a5367f

Extra detail on same regression: now seeing it specifically after desktop merged into ChatGPT app. Regular model quota exhausted + spark quota available. In merged desktop, 5.3-codex-spark requests still include unsupported
easoning.summary, failing with 400. CLI path appears unaffected in my repro. This is tied to app-only merge changes, not backend model availability. (I also added this context to #31969.)
Platform: Windows 11

Additional reproduction:

1. Fully quit Codex App.
2. Start a new CLI session using gpt-5.3-codex-spark.
3. The session works normally in CLI.
4. Open Codex App.
5. The App discovers the CLI session and automatically adds it to the workspace.
6. Return to the same CLI session and send another prompt.
7. The session now fails with:

Unsupported parameter: 'reasoning.summary' ...

This suggests the App mutates or injects config into an existing shared CLI session, rather than the bug being limited to prompts initiated from the App UI.

I investigated the error in Codex Desktop.

My global ~/.codex/config.toml does not contain either model_reasoning_summary or any other setting that explicitly enables reasoning summaries.

However, the archived session data shows that Codex created the Spark turn with:

{
  "model": "gpt-5.3-codex-spark",
  "reasoning_effort": "xhigh",
  "summary": "auto"
}

This suggests that summary: "auto" is being added by Codex itself, most likely through the model catalog defaults rather than the user configuration.

The most likely regression point appears to be PR #12873 merged in commit ba41e84a5

A particularly relevant implementation commit appears to be 9eef0a2

That change made model_reasoning_summary optional and introduced a fallback to the model catalog:

let reasoning_summary = session_configuration
    .model_reasoning_summary
    .unwrap_or(model_info.default_reasoning_summary);

This creates the following failure path:

No reasoning-summary setting in config.toml ↓ Codex uses default_reasoning_summary from the model catalog ↓ Spark receives reasoning.summary = "auto" ↓ The Spark endpoint rejects the unsupported parameter

There was also a related change in commit 3425fdb, which changed the TUI reasoning-summary default to auto.

My current hypothesis is that PR #12873 introduced the client-side behavior that made this failure possible, while a later server-side or model-catalog metadata change caused it to surface again for gpt-5.3-codex-spark.

This would also explain why the same Codex installation could previously use Spark successfully and then begin failing without any corresponding change in the local config.toml.

The important distinction is that this is not a reasoning-effort issue. Spark supports reasoning effort values such as xhigh. The rejected parameter is specifically:

reasoning.summary

Ideally, Codex should only send reasoning.summary when the selected model explicitly supports reasoning summaries. A possible fix would be to make the resolved summary conditional on the model capability instead of relying only on default_reasoning_summary.

For example, the client should effectively behave like:

let reasoning_summary = if model_info.supports_reasoning_summaries {
    session_configuration
        .model_reasoning_summary
        .unwrap_or(model_info.default_reasoning_summary)
} else {
    ReasoningSummary::None
};

This would preserve reasoning summaries for models that support them without requiring users to disable summaries globally just to use Spark.

^{Written with help of GPT-5.6 Sol Extra High}

On windows, I have resolved the issue by modifying the file "C:\Users\\< user >\\.codex\models_cache.json". Find "slug": "gpt-5.3-codex-spark" and change supports_reasoning_summaries to false from its current value true.

The model all config, that is working for me on latest codex desktop on windows 11.

{
"slug": "gpt-5.3-codex-spark",
"display_name": "GPT-5.3-Codex-Spark",
"description": "Ultra-fast coding model.",
"default_reasoning_level": "high",
"supported_reasoning_levels": [
{
"effort": "low",
"description": "Fast responses with lighter reasoning"
},
{
"effort": "medium",
"description": "Balances speed and reasoning depth for everyday tasks"
},
{
"effort": "high",
"description": "Greater reasoning depth for complex problems"
},
{
"effort": "xhigh",
"description": "Extra high reasoning depth for complex problems"
}
],
"shell_type": "shell_command",
"visibility": "list",
"supported_in_api": false,
"priority": 26,
"additional_speed_tiers": [],
"service_tiers": [],
"availability_nux": null,
"upgrade": null,
"base_instructions": "...",
"model_messages": {
"instructions_template": "...",
"instructions_variables": {
"personality_default": "",
"personality_friendly": "..."
}
},
"supports_reasoning_summaries": false,
"default_reasoning_summary": "none",
"support_verbosity": true,
"default_verbosity": "low",
"apply_patch_tool_type": "freeform",
"web_search_tool_type": "text",
"truncation_policy": {
"mode": "tokens",
"limit": 10000
},
"supports_parallel_tool_calls": true,
"supports_image_detail_original": false,
"context_window": 128000,
"max_context_window": 128000,
"comp_hash": "2911",
"effective_context_window_percent": 95,
"experimental_supported_tools": [],
"input_modalities": [
"text"
],
"supports_search_tool": true,
"use_responses_lite": false
},

On windows, I have resolved the issue by modifying the file "C:\Users\< user >\.codex\models_cache.json". Find "slug": "gpt-5.3-codex-spark" and change supports_reasoning_summaries to false from its current value true. The model all config, that is working for me on latest codex desktop on windows 11. { "slug": "gpt-5.3-codex-spark", "display_name": "GPT-5.3-Codex-Spark", "description": "Ultra-fast coding model.", "default_reasoning_level": "high", "supported_reasoning_levels": [ { "effort": "low", "description": "Fast responses with lighter reasoning" }, { "effort": "medium", "description": "Balances speed and reasoning depth for everyday tasks" }, { "effort": "high", "description": "Greater reasoning depth for complex problems" }, { "effort": "xhigh", "description": "Extra high reasoning depth for complex problems" } ], "shell_type": "shell_command", "visibility": "list", "supported_in_api": false, "priority": 26, "additional_speed_tiers": [], "service_tiers": [], "availability_nux": null, "upgrade": null, "base_instructions": "...", "model_messages": { "instructions_template": "...", "instructions_variables": { "personality_default": "", "personality_friendly": "..." } }, "supports_reasoning_summaries": false, "default_reasoning_summary": "none", "support_verbosity": true, "default_verbosity": "low", "apply_patch_tool_type": "freeform", "web_search_tool_type": "text", "truncation_policy": { "mode": "tokens", "limit": 10000 }, "supports_parallel_tool_calls": true, "supports_image_detail_original": false, "context_window": 128000, "max_context_window": 128000, "comp_hash": "2911", "effective_context_window_percent": 95, "experimental_supported_tools": [], "input_modalities": [ "text" ], "supports_search_tool": true, "use_responses_lite": false },

I came to the same solution if I use the desktop version of Codex, however it resets each time you ask a new prompt (for me).

I found when I installed the Codex extension within VSCode I could run the Codex 5.3 Spark without the problems, if I had both programs open the Codex app would show the model running and answering, but if I click into that chat on the Codex app it will ruin the thread until I re-send the prompt through VSCode

Additional macOS reproduction with concrete session metadata:

• Product: Codex Desktop
• Bundled Codex version: 0.144.0-alpha.4
• Platform: macOS arm64
• Auth/provider: ChatGPT auth, built-in openai provider
• Model: gpt-5.3-codex-spark
• Reproduces in a newly created task with a trivial prompt
• Other listed models work normally

I removed the global model override and removed model_reasoning_summary from ~/.codex/config.toml, then reproduced again.

The failed rollout's turn_context contains:

{
  "model": "gpt-5.3-codex-spark",
  "reasoning_effort": null,
  "reasoning_summary": null,
  "service_tier": null
}

The task still immediately receives:

Unsupported parameter: 'reasoning.summary' is not supported with the 'gpt-5.3-codex-spark' model.

The local app-server model/list response advertises Spark normally as GPT-5.3-Codex-Spark with default reasoning effort high, but the public model-list response does not expose a reasoning-summary capability.

This further supports the existing diagnosis that the Desktop request path/model catalog is injecting or falling back to a reasoning summary even when the task has no summary override.

For macOS users, you can fix this by updating your ~/.codex/models_cache.json file similar to Windows. Change supports_reasoning_summaries to false for the gpt-5.3-codex-spark model. Because the Codex app fetches the model config and overwrites this file on every startup, you'll need to make the file read-only (chmod 444 ~/.codex/models_cache.json) after modifying it so your changes aren't lost when you restart the app.

Windows reproduction with a precise client-side trace:

Environment

• Windows 11
• Codex Desktop OpenAI.Codex 26.707.3748.0
• Bundled Codex runtime 0.144.0-alpha.4
• ChatGPT account authentication

Minimal reproduction

1. Fully quit and restart Codex Desktop.
2. Create a fresh projectless task.
3. Select GPT-5.3-Codex-Spark with Extra High reasoning.
4. Send exactly: Reply with exactly: STOP

Expected: STOP

Actual: the request fails before model output or tool execution:

Unsupported parameter: 'reasoning.summary' is not supported with the 'gpt-5.3-codex-spark' model.

Projectless reproduction thread: 019f52e6-d5c4-7312-9a8c-3d2141eab42a

Exact local evidence (UTC)

The rollout records the selected model but no assistant output:

2026-07-11T20:38:39.775Z model=gpt-5.3-codex-spark effort=xhigh summary=auto
2026-07-11T20:38:39.792Z user_message="Reply with exactly: STOP"
2026-07-11T20:38:40.158Z task_complete has_agent_message=false

At the same thread/turn start, the Desktop log resolves an app-owned feature override:

2026-07-11T20:38:07.525Z Concurrent reasoning summaries thread-start config resolved featureOverride=true threadStartConfig=true
2026-07-11T20:38:25.316Z Reasoning summary turn-start config resolved concurrentReasoningSummariesFeatureOverrideEnabled=true summary=detailed

The structured Codex log then records:

{
  "type": "error",
  "code": "unsupported_parameter",
  "message": "Unsupported parameter: 'reasoning.summary' is not supported with the 'gpt-5.3-codex-spark' model.",
  "param": "reasoning.summary"
}

Root-cause trace

1. No active TOML contains model_reasoning_summary, reasoning_summary, or reasoning.summary.
2. A full-restart experiment with model_reasoning_summary = "none" still logged summary=detailed; the experiment was reverted.
3. The live model catalog fetched for client 0.144.0 declares:

``text
slug=gpt-5.3-codex-spark
supports_reasoning_summaries=true
default_reasoning_summary=none
``

1. The Desktop bundle applies the enabled concurrent_reasoning_summaries override after thread/config defaults and sets the per-turn value to detailed before turn/start.
2. Runtime 0.144.0-alpha.4 builds a reasoning request only when model metadata says summaries are supported, so the incorrect catalog capability allows the Desktop-injected detailed value into the request: request builder at the exact runtime commit.

The documented model_supports_reasoning_summaries = false escape hatch cannot disable this in runtime 0.144.0-alpha.4. Its model override only handles true, and the bundled tests explicitly assert that false does not disable support: override implementation, no-op test. This also conflicts with the current configuration reference, which says the key can force Codex to send or not send reasoning metadata.

Requested fix

Any one of these would unblock Spark, but all three inconsistencies should be reconciled:

1. Do not force summary=detailed for Spark in Desktop.
2. Correct Spark's catalog capability to supports_reasoning_summaries=false while the backend rejects reasoning.summary.
3. Either restore model_supports_reasoning_summaries=false as a real disable override or update the public documentation and provide a supported disable mechanism.

This is independent of repositories and skills: it reproduces in a fresh projectless task before agent output, tool calls, or repository inspection.

Related open reports: #31846, #31969, and CLI report #26652.

Same here — terminal path has trouble with models >5.3, but VS Code keeps working with 5.3, so I’m currently using that as a workaround.

im seeing this since the last codex update: Unsupported parameter: 'reasoning.summary' is not supported with the 'gpt-5.3-codex-spark' model. using the cli works just fine

having the same issue.

Same issue here: Unsupported parameter: 'reasoning.summary' is not supported with the 'gpt-5.3-codex-spark' model. I have this config:

`model_reasoning_effort = "medium"

[agents]
max_threads = 4
max_depth = 1
`

having the same issue on windows desktop, meanwhile the cli version is unaffected

Same issue. Feedback ID 019f5b9c-0be4-7b91-b237-dc8e53de7f07

Also hitting this, with an additional scenario not covered above: the error occurs after switching models mid-conversation, so an in-flight task gets stuck with no way to continue.

Version: 26.707.62119 (released Jul 12, 2026)
Subscription: Pro
Platform: Darwin 25.5.0 arm64 arm (macOS, Apple Silicon)

Steps to reproduce:

1. Start a task in the Codex App using GPT-5.6 Sol.
2. Continue the conversation ("Continued from task") and switch the model to GPT-5.3 Codex Spark.
3. Send any prompt (in my case an i18n review request; context was also automatically compacted at this point).

Actual behavior:
Every request fails immediately with:

Unsupported parameter: 'reasoning.summary' is not supported with the 'gpt-5.3-codex-spark' model.

Clicking "continue" fails with the same error, so the continued task is effectively bricked on this model — the app keeps sending reasoning.summary even though the selected model rejects it.

Expected behavior:
The app should omit reasoning.summary (or whatever unsupported parameters apply) when the selected model doesn't support it, so switching to GPT-5.3 Codex Spark mid-conversation works.

Same issue here.

Same issue.

Windows 11 reproduction persists on the latest Desktop build

I can still reproduce this on:

• OS: Windows 11
• Codex Desktop: 26.707.9981.0
• Bundled Codex runtime: 0.144.2
• Model: gpt-5.3-codex-spark
• Reasoning effort: Extra High / xhigh
• Failed session: 019f628d-ead8-7de1-b4e7-fedceff7c8f5

I fully restarted Codex Desktop and created a fresh task. Every Spark request still fails immediately with:

Unsupported parameter: 'reasoning.summary' is not supported with the 'gpt-5.3-codex-spark' model.

Same issue.

Yeah, I've tried manually editing the json config, but that doesn't seem to fix it on MacOS.

You'd think that a company striving to be worth a trillion dollars would be more keen on fixing obvious bugs. Maybe they could get a Claude subscription. /s

Same issue here as well. What worked for me was changing spark in ~/.codex/models_cache.json to:

"supports_reasoning_summaries": false,
"default_reasoning_summary": "none"

and then running chmod 444 ~/.codex/models_cache.json (so codex doesn't change it when it restarts)

one caveat - you may need to quit and reopen the codex/chatgpt electron app

What worked for me was changing spark in ~/.codex/models_cache.json to: `` "supports_reasoning_summaries": false, "default_reasoning_summary": "none" ` and then running chmod 444 ~/.codex/models_cache.json` (so codex doesn't change it when it restarts)

This worked for me on Mac

Same issue - windows

Thanks for the bug report. This will be fixed in the next release of the app.

Thanks @etraut-openai

Not to be greedy or anything but would it be possible to get a reset since we have been unable to use this model until now?

Thanks @etraut-openai Not to be greedy or anything but would it be possible to get a reset since we have been unable to use this model until now?

Is it fixed for you? What platform are you on? For me (Windows) the issue still has not been fixed yet.

Additional Windows desktop reproduction:

• Platform: Windows 10 x64
• Product: Codex desktop app after a recent app update
• Model: gpt-5.3-codex-spark
• Exact error:

``
Unsupported parameter: 'reasoning.summary' is not supported with the 'gpt-5.3-codex-spark' model.
``

The local cache workaround described in this thread was attempted:

1. Fully quit the Codex app.
2. In %USERPROFILE%\.codex\models_cache.json, change the Spark entry's supports_reasoning_summaries value to false.
3. Restart the app and start a fresh Spark task.

Observed behavior:

• Spark appeared to work briefly and began a coding task.
• The app then displayed reconnecting 5/5.
• After reconnecting, the same unsupported reasoning.summary error returned.
• The cache workaround is therefore not durable for this Windows desktop path.

This suggests that the app may reintroduce or re-resolve the unsupported summary parameter after reconnect, session restoration, or model-catalog refresh.

Please ensure reasoning.summary is omitted during final request serialization whenever the final selected model is gpt-5.3-codex-spark, regardless of cached metadata, prior thread/model state, reconnects, or restored sessions. A Windows desktop regression test covering a fresh Spark turn followed by reconnect/session restoration would be useful.

@martinmclee nope not working for me.

Looks like we just got the reset. Thanks so much!

still not working 019f69d4-7bd3-7ab3-8356-23156a805028

After the recent update with the desktop app on Windows, the 5.3-codex-spark model is gone from the model selector, replaced by the "5.2" model. Can I take it that they have retired the model? Meanwhile I can see the "GPT-5.3-Codex-Spark usage limits", and the spark model is still in the cli (while 5.2 is not in cli)

Thanks for the bug report. This will be fixed in the next release of the app.

Also please look into the issue of "■ Error running remote compact task: Codex ran out of room in the model's context window. Start a new thread or clear earlier history before retrying." with the spark model, it happens for sessions with spark and there is no way to recover such session.

Same error on Version 26.707.91948

Same issue here on macOS with the official Codex app. Selecting gpt-5.3-codex-spark causes the prompt to fail immediately with:

Unsupported parameter: 'reasoning.summary' is not supported with the 'gpt-5.3-codex-spark' model.

I am using the official app, not a custom API provider. Please ensure the app omits reasoning.summary when Spark is selected.