diff --git a/.changeset/neon-gpt-codex-only.md b/.changeset/neon-gpt-codex-only.md
new file mode 100644
index 00000000..9773f729
--- /dev/null
+++ b/.changeset/neon-gpt-codex-only.md
@@ -0,0 +1,9 @@
+---
+"@upstash/box": patch
+"@upstash/box-cli": patch
+---
+
+Neon AI Gateway: GPT-5.4 and newer (GPT-6 Astra, GPT-5.6 Sol/Terra/Luna, GPT-5.5,
+GPT-5.4/Mini/Nano) are offered for Codex only. Neon refuses tool calls with reasoning
+for these models on chat completions, which OpenCode needs; Codex uses the Responses
+API. The CLI picker no longer lists them under OpenCode.
diff --git a/packages/cli/src/__tests__/models.test.ts b/packages/cli/src/__tests__/models.test.ts
index 61767f40..a524003a 100644
--- a/packages/cli/src/__tests__/models.test.ts
+++ b/packages/cli/src/__tests__/models.test.ts
@@ -125,6 +125,21 @@ describe("MODEL_OPTIONS_BY_AGENT", () => {
     expect(values(Agent.OpenCode)).toContain(NeonModel.Claude_Haiku_4_5);
     expect(values(Agent.OpenCode)).toContain(NeonModel.Claude_Opus_5_5);
     expect(values(Agent.Codex).some((v) => v.startsWith("neon/claude-"))).toBe(false);
+    // GPT-5.4 and newer refuse tools with reasoning on Neon chat completions: Codex only.
+    for (const model of [
+      NeonModel.GPT_6_Astra,
+      NeonModel.GPT_5_6_Sol,
+      NeonModel.GPT_5_6_Terra,
+      NeonModel.GPT_5_6_Luna,
+      NeonModel.GPT_5_5,
+      NeonModel.GPT_5_4,
+      NeonModel.GPT_5_4_Mini,
+      NeonModel.GPT_5_4_Nano,
+    ]) {
+      expect(values(Agent.Codex)).toContain(model);
+      expect(values(Agent.OpenCode)).not.toContain(model);
+    }
+    expect(values(Agent.OpenCode)).toContain(NeonModel.GPT_5_Nano);
   });
 
   it("includes Cursor models", () => {
diff --git a/packages/cli/src/models.ts b/packages/cli/src/models.ts
index ca09af92..dd0bbe95 100644
--- a/packages/cli/src/models.ts
+++ b/packages/cli/src/models.ts
@@ -280,17 +280,11 @@ export const MODEL_OPTIONS_BY_AGENT: Record<
       ],
     },
     {
-      // Chat-completions models only: GPT-5.5 Pro and GPT-5.3 Codex are Responses-only on Neon.
+      // Chat-completions models only. GPT-5.5 Pro and GPT-5.3 Codex are Responses-only on
+      // Neon, and GPT-5.4 and newer refuse tools with reasoning on chat completions, so
+      // those run on Codex only.
       label: "Neon AI Gateway",
       options: [
-        { value: NeonModel.GPT_6_Astra, label: "GPT-6 Astra (Neon)" },
-        { value: NeonModel.GPT_5_6_Sol, label: "GPT-5.6 Sol (Neon)" },
-        { value: NeonModel.GPT_5_6_Terra, label: "GPT-5.6 Terra (Neon)" },
-        { value: NeonModel.GPT_5_6_Luna, label: "GPT-5.6 Luna (Neon)" },
-        { value: NeonModel.GPT_5_5, label: "GPT-5.5 (Neon)" },
-        { value: NeonModel.GPT_5_4, label: "GPT-5.4 (Neon)" },
-        { value: NeonModel.GPT_5_4_Mini, label: "GPT-5.4 Mini (Neon)" },
-        { value: NeonModel.GPT_5_4_Nano, label: "GPT-5.4 Nano (Neon)" },
         { value: NeonModel.GPT_5_Mini, label: "GPT-5 Mini (Neon)" },
         { value: NeonModel.GPT_5_Nano, label: "GPT-5 Nano (Neon)" },
         { value: NeonModel.GPT_OSS_120B, label: "GPT OSS 120B (Neon)" },
diff --git a/packages/python-sdk/CHANGELOG.md b/packages/python-sdk/CHANGELOG.md
index 91580266..fe78e2bb 100644
--- a/packages/python-sdk/CHANGELOG.md
+++ b/packages/python-sdk/CHANGELOG.md
@@ -4,6 +4,9 @@ All notable changes to `upstash-box` (Python) are documented here.
 
 ## Unreleased
 
+- Document that Neon GPT-5.4 and newer (`GPT_6_ASTRA`, `GPT_5_6_*`, `GPT_5_5`,
+  `GPT_5_4*`) run on Codex only: Neon refuses tool calls with reasoning for them
+  on chat completions, which OpenCode needs.
 - Add `NeonModel` (Neon AI Gateway, `neon/<neon-short-id>`). `infer_default_provider`
   routes `neon/gpt-*` to Codex, except `neon/gpt-oss-*`, and every other Neon model
   (GPT OSS, Claude, Gemini, Llama, Qwen, Kimi, GLM, Grok) to OpenCode.
diff --git a/packages/python-sdk/upstash_box/types.py b/packages/python-sdk/upstash_box/types.py
index ab70a295..bd79239c 100644
--- a/packages/python-sdk/upstash_box/types.py
+++ b/packages/python-sdk/upstash_box/types.py
@@ -150,8 +150,11 @@ class NeonModel(str, Enum):
 
     These run on the Codex harness (Responses API) or the OpenCode harness
     (chat completions); Box does not run Claude Code with Neon.
-    ``GPT_5_5_PRO`` and ``GPT_5_3_CODEX`` are Responses-only (Codex);
-    the other GPT models and ``GROK_4_6`` run on both; ``GPT_OSS_*``, the
+    ``GPT_5_5_PRO`` and ``GPT_5_3_CODEX`` are Responses-only (Codex), and
+    GPT-5.4 and newer (``GPT_6_ASTRA``, ``GPT_5_6_*``, ``GPT_5_5``,
+    ``GPT_5_4*``) run on Codex only, because Neon refuses tool calls with
+    reasoning for them on chat completions; ``GPT_5_MINI``, ``GPT_5_NANO``
+    and ``GROK_4_6`` run on both; ``GPT_OSS_*``, the
     ``CLAUDE_*`` models and every other non-OpenAI model except Grok are served
     through chat completions only (OpenCode).
     Neon credentials are a token plus a per-branch base URL and must be stored
diff --git a/packages/sdk/src/types.ts b/packages/sdk/src/types.ts
index 64bbcc2a..69c4a108 100644
--- a/packages/sdk/src/types.ts
+++ b/packages/sdk/src/types.ts
@@ -148,8 +148,10 @@ export enum VercelModel {
  *
  * These models run on the Codex harness (Responses API) or the OpenCode harness
  * (chat completions). Box does not run Claude Code with Neon.
- * Codex-only: `GPT_5_5_Pro`, `GPT_5_3_Codex` (Responses-only on Neon).
- * Both: the other GPT models and `Grok_4_6` (Neon serves them on both endpoints).
+ * Codex-only: `GPT_5_5_Pro`, `GPT_5_3_Codex` (Responses-only on Neon), and GPT-5.4 and
+ * newer (`GPT_6_Astra`, `GPT_5_6_*`, `GPT_5_5`, `GPT_5_4*`): Neon refuses tool calls with
+ * reasoning for them on chat completions, which OpenCode needs.
+ * Both: `GPT_5_Mini`, `GPT_5_Nano` and `Grok_4_6`.
  * OpenCode-only: `GPT_OSS_120B`, `GPT_OSS_20B`, the Claude models and every other
  * non-OpenAI model except Grok (Neon serves them through chat completions only).
  * Neon credentials are a token plus a per-branch base URL and must be stored via the