Add Codex GPT-5.5 support (#1617)

Integrated into release/v3.7.1

open-sse/config/providerRegistry.ts

@@ -140,6 +140,16 @@ const KIMI_CODING_SHARED = {
 const buildModels = (ids: readonly string[]): RegistryModel[] =>
   ids.map((id) => ({ id, name: id }));
 
+const GPT_5_5_CONTEXT_LENGTH = 1050000;
+const GPT_5_5_CODEX_CAPABILITIES = {
+  targetFormat: "openai-responses",
+  toolCalling: true,
+  supportsReasoning: true,
+  supportsVision: true,
+  supportsXHighEffort: true,
+  contextLength: GPT_5_5_CONTEXT_LENGTH,
+} as const;
+
 const CHAT_OPENAI_COMPAT_MODELS: Record<string, RegistryModel[]> = {
   deepinfra: buildModels([
     "Qwen/Qwen3-Coder-480B-A35B-Instruct",
@@ -373,10 +383,10 @@ export const REGISTRY: Record<string, RegistryEntry> = {
     },
     models: [
       { id: "codex-auto-review", name: "Codex Auto Review", targetFormat: "openai-responses" },
-      { id: "gpt-5.5-xhigh", name: "GPT 5.5 (xHigh)", targetFormat: "openai-responses" },
-      { id: "gpt-5.5-high", name: "GPT 5.5 (High)", targetFormat: "openai-responses" },
-      { id: "gpt-5.5", name: "GPT 5.5 (Medium)", targetFormat: "openai-responses" },
-      { id: "gpt-5.5-low", name: "GPT 5.5 (Low)", targetFormat: "openai-responses" },
+      { id: "gpt-5.5-xhigh", name: "GPT 5.5 (xHigh)", ...GPT_5_5_CODEX_CAPABILITIES },
+      { id: "gpt-5.5-high", name: "GPT 5.5 (High)", ...GPT_5_5_CODEX_CAPABILITIES },
+      { id: "gpt-5.5", name: "GPT 5.5 (Medium)", ...GPT_5_5_CODEX_CAPABILITIES },
+      { id: "gpt-5.5-low", name: "GPT 5.5 (Low)", ...GPT_5_5_CODEX_CAPABILITIES },
       { id: "gpt-5.5-mini", name: "GPT 5.5 Mini", targetFormat: "openai-responses" },
       { id: "gpt-5.4", name: "GPT 5.4", targetFormat: "openai-responses" },
       { id: "gpt-5.4-mini", name: "GPT 5.4 Mini", targetFormat: "openai-responses" },

open-sse/executors/codex.ts

@@ -202,6 +202,26 @@ type EffortLevel = (typeof EFFORT_ORDER)[number];
 const CODEX_FAST_WIRE_VALUE = "priority";
 const CODEX_RESPONSES_WS_URL = "wss://chatgpt.com/backend-api/codex/responses";
 
+function splitCodexReasoningSuffix(model: unknown): {
+  baseModel: string;
+  effort: EffortLevel | null;
+} {
+  const modelId = typeof model === "string" ? model : "";
+  for (const level of EFFORT_ORDER) {
+    if (modelId.endsWith(`-${level}`)) {
+      return {
+        baseModel: modelId.slice(0, -`-${level}`.length),
+        effort: level,
+      };
+    }
+  }
+  return { baseModel: modelId, effort: null };
+}
+
+export function getCodexUpstreamModel(model: unknown): string {
+  return splitCodexReasoningSuffix(model).baseModel;
+}
+
 function stringifyCodexInstructionContent(content: unknown): string {
   if (typeof content === "string") {
     return content.trim();
@@ -469,10 +489,8 @@ function consumeResponsesStoreMarker(body: Record<string, unknown>): unknown {
   return marker;
 }
 
-function isCodexResponsesWebSocketRequired(model: string, credentials: unknown): boolean {
-  const normalizedModel = String(model || "")
-    .trim()
-    .toLowerCase();
+export function isCodexResponsesWebSocketRequired(model: string, credentials: unknown): boolean {
+  const normalizedModel = getCodexUpstreamModel(model).trim().toLowerCase();
   if (normalizedModel === "gpt-5.5") return true;
   const providerSpecificData =
     credentials && typeof credentials === "object"
@@ -626,7 +644,7 @@ export class CodexExecutor extends BaseExecutor {
       true,
       input.credentials
     )) as Record<string, unknown>;
-    transformedBody.model = input.model;
+    transformedBody.model = getCodexUpstreamModel(transformedBody.model || input.model);
     delete transformedBody.stream;
     delete transformedBody.stream_options;
 
@@ -988,16 +1006,13 @@ export class CodexExecutor extends BaseExecutor {
     delete body.messages;
     delete body.prompt;
 
-    const effortLevels = ["none", "low", "medium", "high", "xhigh"];
     let modelEffort: string | null = null;
     let cleanModel = typeof body.model === "string" ? body.model : model;
-    for (const level of effortLevels) {
-      if (typeof cleanModel === "string" && cleanModel.endsWith(`-${level}`)) {
-        modelEffort = level;
-        body.model = cleanModel.slice(0, -`-${level}`.length);
-        cleanModel = body.model;
-        break;
-      }
+    const splitModel = splitCodexReasoningSuffix(cleanModel);
+    if (splitModel.effort) {
+      modelEffort = splitModel.effort;
+      body.model = splitModel.baseModel;
+      cleanModel = body.model;
     }
 
     const explicitReasoning = normalizeEffortValue(body?.reasoning?.effort);

src/shared/constants/modelSpecs.ts

@@ -18,6 +18,23 @@ export interface ModelSpec {
 }
 
 export const MODEL_SPECS: Record<string, ModelSpec> = {
+  "gpt-5.5-mini": {
+    maxOutputTokens: 128000,
+    contextWindow: 400000,
+    supportsThinking: true,
+    supportsTools: true,
+    supportsVision: true,
+  },
+
+  "gpt-5.5": {
+    maxOutputTokens: 128000,
+    contextWindow: 1050000,
+    supportsThinking: true,
+    supportsTools: true,
+    supportsVision: true,
+    aliases: ["gpt-5.5-xhigh", "gpt-5.5-high", "gpt-5.5-medium", "gpt-5.5-low", "gpt-5.5-none"],
+  },
+
   // ── Gemini 3 Flash series ───────────────────────────────────────
   "gemini-3-flash": {
     maxOutputTokens: 65536,

src/shared/constants/pricing.ts

@@ -11,6 +11,14 @@ const GPT_5_3_CODEX_PRICING = {
   cache_creation: 5.0,
 };
 
+const GPT_5_5_PRICING = {
+  input: 5.0,
+  output: 30.0,
+  cached: 0.5,
+  reasoning: 30.0,
+  cache_creation: 5.0,
+};
+
 const CLAUDE_OPUS_4_PRICING = {
   input: 15.0,
   output: 75.0,
@@ -167,6 +175,14 @@ export const DEFAULT_PRICING = {
 
   // OpenAI Codex (cx)
   cx: {
+    // GPT 5.5
+    "gpt-5.5": GPT_5_5_PRICING,
+    "gpt5.5": GPT_5_5_PRICING,
+    "gpt-5.5-xhigh": GPT_5_5_PRICING,
+    "gpt-5.5-high": GPT_5_5_PRICING,
+    "gpt-5.5-medium": GPT_5_5_PRICING,
+    "gpt-5.5-low": GPT_5_5_PRICING,
+    "gpt-5.5-none": GPT_5_5_PRICING,
     // GPT 5.4
     "gpt-5.4": {
       input: 5.0,
@@ -584,6 +600,7 @@ export const DEFAULT_PRICING = {
 
   // OpenAI
   openai: {
+    "gpt-5.5": GPT_5_5_PRICING,
     "gpt-4o": {
       input: 2.5,
       output: 10.0,

tests/unit/context-manager.test.ts

@@ -26,6 +26,10 @@ test("getTokenLimit: detects gemini", () => {
   assert.equal(getTokenLimit("gemini", "gemini-2.5-pro"), 1048576);
 });
 
+test("getTokenLimit: uses GPT-5.5 Codex model context", () => {
+  assert.equal(getTokenLimit("codex", "gpt-5.5"), 1050000);
+});
+
 test("getTokenLimit: default fallback", () => {
   assert.equal(getTokenLimit("unknown"), 128000);
 });

tests/unit/executor-codex.test.ts

@@ -7,6 +7,8 @@ import {
   getCodexModelScope,
   getCodexRateLimitKey,
   getCodexResetTime,
+  getCodexUpstreamModel,
+  isCodexResponsesWebSocketRequired,
   parseCodexQuotaHeaders,
 } from "../../open-sse/executors/codex.ts";
 import {
@@ -58,6 +60,10 @@ test("Codex helper functions isolate rate-limit scopes and parse quota headers",
 
   assert.equal(getCodexModelScope("codex-spark-mini"), "spark");
   assert.equal(getCodexModelScope("gpt-5.3-codex"), "codex");
+  assert.equal(getCodexModelScope("gpt-5.5-xhigh"), "codex");
+  assert.equal(getCodexUpstreamModel("gpt-5.5-xhigh"), "gpt-5.5");
+  assert.equal(isCodexResponsesWebSocketRequired("gpt-5.5-xhigh", {}), true);
+  assert.equal(isCodexResponsesWebSocketRequired("gpt-5.5-mini", {}), false);
   assert.equal(getCodexRateLimitKey("acct-1", "codex-spark-mini"), "acct-1:spark");
   assert.equal(quota.usage5h, 100);
   assert.equal(quota.limit7d, 5000);
@@ -276,8 +282,8 @@ test("CodexExecutor.transformRequest lets model suffix beat connection reasoning
 test("CodexExecutor.transformRequest keeps gpt-5.5 as the model and applies xhigh reasoning", () => {
   const executor = new CodexExecutor();
   const result = executor.transformRequest(
-    "gpt-5.5",
-    { model: "gpt-5.5", input: [], reasoning_effort: "xhigh" },
+    "gpt-5.5-xhigh",
+    { model: "gpt-5.5-xhigh", input: [] },
     false,
     {}
   );

tests/unit/model-capabilities-registry.test.ts

@@ -102,6 +102,12 @@ test("canonical model capability resolver merges models.dev data and keeps stati
     modelCapabilities.capThinkingBudget("antigravity/gemini-3.1-pro-high", 40000),
     32768
   );
+
+  const codexGpt55 = modelCapabilities.getResolvedModelCapabilities("codex/gpt-5.5");
+  assert.equal(codexGpt55.contextWindow, 1050000);
+  assert.equal(codexGpt55.maxOutputTokens, 128000);
+  assert.equal(codexGpt55.supportsThinking, true);
+  assert.equal(codexGpt55.supportsVision, true);
 });
 
 test("GPT OSS and DeepSeek Reasoner models support tool calling", () => {

tests/unit/t12-pricing-updates.test.ts

@@ -7,6 +7,12 @@ import { REGISTRY } from "../../open-sse/config/providerRegistry.ts";
 test("T12: pricing table includes MiniMax, GLM, Kimi and gpt-5.4 mini entries", () => {
   const pricing = getDefaultPricing();
 
+  assert.ok(pricing.cx["gpt-5.5"], "missing cx/gpt-5.5");
+  assert.ok(pricing.cx["gpt-5.5-xhigh"], "missing cx/gpt-5.5-xhigh");
+  assert.equal(pricing.cx["gpt-5.5"].input, 5.0);
+  assert.equal(pricing.cx["gpt-5.5"].cached, 0.5);
+  assert.equal(pricing.cx["gpt-5.5"].output, 30.0);
+
   assert.ok(pricing.cx["gpt-5.4"], "missing cx/gpt-5.4");
   assert.ok(pricing.cx["gpt-5.4-mini"], "missing cx/gpt-5.4-mini");
 
@@ -30,9 +36,12 @@ test("T12: pricing table includes MiniMax, GLM, Kimi and gpt-5.4 mini entries",
 });
 
 test("T12: codex catalog includes GPT 5.5 entries", () => {
-  const codexModels = REGISTRY.codex.models.map((m) => m.id);
-  assert.ok(codexModels.includes("gpt-5.5"), "missing codex/gpt-5.5");
-  assert.ok(codexModels.includes("gpt-5.5-mini"), "missing codex/gpt-5.5-mini");
+  const codexModels = new Map(REGISTRY.codex.models.map((m) => [m.id, m]));
+  assert.ok(codexModels.has("gpt-5.5"), "missing codex/gpt-5.5");
+  assert.ok(codexModels.has("gpt-5.5-mini"), "missing codex/gpt-5.5-mini");
+  assert.equal(codexModels.get("gpt-5.5")?.contextLength, 1050000);
+  assert.equal(codexModels.get("gpt-5.5")?.supportsXHighEffort, true);
+  assert.equal(codexModels.get("gpt-5.5-xhigh")?.targetFormat, "openai-responses");
 });
 
 test("T12: minimax default model list starts with M2.7", () => {

tests/unit/t31-t33-t34-t38-model-specs.test.ts

@@ -42,6 +42,8 @@ test("T33: thinkingLevel string is converted into numeric thinkingBudget", () =>
 });
 
 test("T34: max output tokens are capped by model spec", () => {
+  assert.equal(capMaxOutputTokens("gpt-5.5", 200000), 128000);
+  assert.equal(capMaxOutputTokens("gpt-5.5-xhigh", 200000), 128000);
   assert.equal(capMaxOutputTokens("gemini-3-flash", 131072), 65536);
   assert.equal(capMaxOutputTokens("gemini-3-flash"), 65536);
   assert.equal(capMaxOutputTokens("gemini-3.1-pro-high", 131072), 65535);
@@ -49,6 +51,9 @@ test("T34: max output tokens are capped by model spec", () => {
 });
 
 test("T38: modelSpecs exposes centralized helpers with alias and prefix lookup", () => {
+  assert.equal(getModelSpec("gpt-5.5").contextWindow, 1050000);
+  assert.equal(getModelSpec("gpt-5.5-high").maxOutputTokens, 128000);
+  assert.equal(getModelSpec("gpt-5.5-mini").contextWindow, 400000);
   assert.equal(typeof MODEL_SPECS["gemini-3.1-pro-high"], "object");
   assert.equal(getModelSpec("gemini-3-pro-high").maxOutputTokens, 65535);
   assert.equal(getModelSpec("gemini-3-pro-preview").maxOutputTokens, 65535);