🤖 feat: add GPT-5.6 Sol/Terra/Luna models and move gpt alias to Sol

docs/config/models.mdx

@@ -18,8 +18,11 @@ Mux ships with curated models kept up to date with the frontier. Use any custom
 | Opus 4.8               | anthropic:claude-opus-4-8     | `opus`                                                       | ✓       |
 | Sonnet 5               | anthropic:claude-sonnet-5     | `sonnet`                                                     |         |
 | Haiku 4.5              | anthropic:claude-haiku-4-5    | `haiku`                                                      |         |
-| GPT-5.5                | openai:gpt-5.5                | `gpt`, `gpt-5.5`                                             |         |
+| GPT-5.5                | openai:gpt-5.5                | `gpt-5.5`                                                    |         |
 | GPT-5.5 Pro            | openai:gpt-5.5-pro            | `gpt-pro`, `gpt-5.5-pro`                                     |         |
+| GPT-5.6 Sol            | openai:gpt-5.6-sol            | `gpt`, `sol`, `gpt-5.6-sol`                                  |         |
+| GPT-5.6 Terra          | openai:gpt-5.6-terra          | `terra`, `gpt-5.6-terra`                                     |         |
+| GPT-5.6 Luna           | openai:gpt-5.6-luna           | `luna`, `gpt-5.6-luna`                                       |         |
 | GPT-5.4 Mini           | openai:gpt-5.4-mini           | `gpt-mini`                                                   |         |
 | GPT-5.4 Nano           | openai:gpt-5.4-nano           | `gpt-nano`                                                   |         |
 | Codex 5.3              | openai:gpt-5.3-codex          | `codex`, `codex-5.3`                                         |         |

src/common/constants/codexOAuth.ts

@@ -94,6 +94,10 @@ export const CODEX_OAUTH_ALLOWED_MODELS = new Set<string>([
   "gpt-5.2",
   "gpt-5.4-mini",
   "gpt-5.5",
+  // GPT-5.6 tiers ship in both Codex and the public API (GA July 9, 2026).
+  "gpt-5.6-sol",
+  "gpt-5.6-terra",
+  "gpt-5.6-luna",
   "gpt-5.2-codex",
   "gpt-5.3-codex",
   "gpt-5.3-codex-spark",

src/common/constants/knownModels.ts

@@ -85,11 +85,12 @@ const MODEL_DEFINITIONS = {
     aliases: ["haiku"],
     tokenizerOverride: "anthropic/claude-3.5-haiku",
   },
-  // GPT alias tracks the latest stable GPT-5 tier.
+  // Previous flagship tier; kept selectable via its explicit id alias.
+  // The bare `gpt` alias moved to GPT_56_SOL when GPT-5.6 went GA (July 9, 2026).
   GPT: {
     provider: "openai",
     providerModelId: "gpt-5.5",
-    aliases: ["gpt", "gpt-5.5"],
+    aliases: ["gpt-5.5"],
     warm: true,
     tokenizerOverride: "openai/gpt-5",
   },
@@ -101,6 +102,37 @@ const MODEL_DEFINITIONS = {
     warm: true,
     tokenizerOverride: "openai/gpt-5",
   },
+  // GPT-5.6 family - previewed June 26, 2026; generally available July 9, 2026 across
+  // ChatGPT, Codex, and the API (verified callable via plain API keys). New naming
+  // system: the number is the generation, while Sol/Terra/Luna are durable capability
+  // tiers. Cache writes bill at 1.25x input; cache reads keep the 90% discount.
+  // All tiers support a native "max" reasoning effort (see the thinking policy).
+  // The API also offers `reasoning.mode: "pro"`, not yet implemented in Mux (planned
+  // as a thinking-slider toggle; tracked in issue #3704).
+  // GPT-5.6 Sol - flagship tier. $5/M input, $30/M output. Bare `gpt` alias lives here:
+  // Sol succeeds gpt-5.5 as the flagship at the same list price.
+  GPT_56_SOL: {
+    provider: "openai",
+    providerModelId: "gpt-5.6-sol",
+    aliases: ["gpt", "sol", "gpt-5.6-sol"],
+    warm: true,
+    tokenizerOverride: "openai/gpt-5",
+  },
+  // GPT-5.6 Terra - balanced tier; GPT-5.5-competitive at half the cost.
+  // $2.50/M input, $15/M output.
+  GPT_56_TERRA: {
+    provider: "openai",
+    providerModelId: "gpt-5.6-terra",
+    aliases: ["terra", "gpt-5.6-terra"],
+    tokenizerOverride: "openai/gpt-5",
+  },
+  // GPT-5.6 Luna - fastest/cheapest tier. $1/M input, $6/M output.
+  GPT_56_LUNA: {
+    provider: "openai",
+    providerModelId: "gpt-5.6-luna",
+    aliases: ["luna", "gpt-5.6-luna"],
+    tokenizerOverride: "openai/gpt-5",
+  },
   // GPT Mini alias tracks the latest stable GPT-5 mini tier.
   GPT_54_MINI: {
     provider: "openai",

src/common/types/thinking.test.ts

@@ -22,6 +22,16 @@ describe("getThinkingDisplayLabel", () => {
     expect(getThinkingDisplayLabel("max", "mux-gateway:openai/gpt-5.2")).toBe("XHIGH");
   });
 
+  test("keeps xhigh and max distinguishable on GPT-5.6 (native max effort)", () => {
+    expect(getThinkingDisplayLabel("xhigh", "openai:gpt-5.6-sol")).toBe("XHIGH");
+    expect(getThinkingDisplayLabel("max", "openai:gpt-5.6-sol")).toBe("MAX");
+    expect(getThinkingDisplayLabel("max", "openai:gpt-5.6-terra")).toBe("MAX");
+    expect(getThinkingDisplayLabel("max", "mux-gateway:openai/gpt-5.6-luna")).toBe("MAX");
+    // Option labels (settings dropdowns) must diverge too
+    expect(getThinkingOptionLabel("xhigh", "openai:gpt-5.6-sol")).toBe("xhigh");
+    expect(getThinkingOptionLabel("max", "openai:gpt-5.6-sol")).toBe("max");
+  });
+
   test("returns MAX for xhigh/max when no model specified (default)", () => {
     expect(getThinkingDisplayLabel("xhigh")).toBe("MAX");
     expect(getThinkingDisplayLabel("max")).toBe("MAX");

src/common/types/thinking.ts

@@ -38,8 +38,13 @@ export function getThinkingDisplayLabel(level: ThinkingLevel, modelString?: stri
     const normalized = modelString.trim().toLowerCase();
     const withoutPrefix = normalized.replace(/^[a-z0-9_-]+:\s*/, "");
 
-    // OpenAI: both xhigh and max resolve to "xhigh" reasoning effort
-    if (normalized.startsWith("openai:") || withoutPrefix.startsWith("openai/")) return "XHIGH";
+    if (normalized.startsWith("openai:") || withoutPrefix.startsWith("openai/")) {
+      // GPT-5.6 has a native "max" reasoning effort distinct from xhigh, so both
+      // labels must stay distinguishable in settings dropdowns and the slider.
+      if (level === "max" && openaiSupportsNativeMaxEffort(modelString)) return "MAX";
+      // Other OpenAI models: both xhigh and max resolve to "xhigh" reasoning effort
+      return "XHIGH";
+    }
 
     // Anthropic Opus 4.7+: xhigh is a distinct effort level from max
     if (level === "xhigh" && anthropicSupportsNativeXhigh(modelString)) return "XHIGH";
@@ -237,6 +242,24 @@ export function anthropicSupportsNativeXhigh(modelString: string): boolean {
   );
 }
 
+/**
+ * Whether the given OpenAI model supports the native "max" reasoning effort
+ * (GPT-5.6 family — Sol/Terra/Luna, GA July 9, 2026 — including version-suffixed
+ * ids like `gpt-5.6-sol-2026-07-09`). OpenAI recommends max effort "for demanding
+ * tasks that need more exploration and verification".
+ *
+ * Only valid on the Responses API path: the @ai-sdk/openai Responses schema
+ * accepts arbitrary effort strings, but the Chat Completions schema rejects
+ * anything outside its enum — callers must clamp to "xhigh" there.
+ *
+ * Note: GPT-5.6 also supports `reasoning.mode: "pro"` (more model work for
+ * reliability, single final answer). Mux does not implement it yet — planned as
+ * a toggle inside the thinking slider; tracked in issue #3704.
+ */
+export function openaiSupportsNativeMaxEffort(modelString: string): boolean {
+  return /^gpt-5\.6-(?:sol|terra|luna)(?!-[a-z])/.test(stripModelProviderPrefixes(modelString));
+}
+
 /**
  * Whether the given Anthropic model rejects `thinking: { type: "disabled" }`.
  *
@@ -276,6 +299,9 @@ export const OPENAI_REASONING_EFFORT: Record<ThinkingLevel, string | undefined>
   medium: "medium",
   high: "high",
   xhigh: "xhigh", // Maps 1:1 to OpenAI's reasoning effort value
+  // Shared clamp for models without a native max effort. GPT-5.6 (see
+  // openaiSupportsNativeMaxEffort) overrides this to "max" in buildProviderOptions
+  // on the Responses API path.
   max: "xhigh",
 };
 

src/common/utils/ai/providerOptions.test.ts

@@ -415,6 +415,45 @@ describe("buildProviderOptions - OpenAI", () => {
     expect(openai!.parallelToolCalls).toBe(true);
   });
 
+  describe("GPT-5.6 native max reasoning effort", () => {
+    test.each(["gpt-5.6-sol", "gpt-5.6-terra", "gpt-5.6-luna"])(
+      "sends native max effort for %s on the Responses API",
+      (modelId) => {
+        const result = buildProviderOptions(`openai:${modelId}`, "max");
+        const openai = getOpenAIOptions(result);
+
+        expect(openai).toBeDefined();
+        expect(openai!.reasoningEffort).toBe("max");
+      }
+    );
+
+    test("clamps max to xhigh on the Chat Completions wire format", () => {
+      const result = buildProviderOptions("openai:gpt-5.6-sol", "max", undefined, undefined, {
+        openai: { wireFormat: "chatCompletions" },
+      });
+      const openai = getOpenAIOptions(result);
+
+      expect(openai).toBeDefined();
+      expect(openai!.reasoningEffort).toBe("xhigh");
+    });
+
+    test("keeps the max→xhigh clamp for non-GPT-5.6 models", () => {
+      const result = buildProviderOptions("openai:gpt-5.5", "max");
+      const openai = getOpenAIOptions(result);
+
+      expect(openai).toBeDefined();
+      expect(openai!.reasoningEffort).toBe("xhigh");
+    });
+
+    test("non-max levels stay on the shared mapping for GPT-5.6", () => {
+      const result = buildProviderOptions("openai:gpt-5.6-sol", "high");
+      const openai = getOpenAIOptions(result);
+
+      expect(openai).toBeDefined();
+      expect(openai!.reasoningEffort).toBe("high");
+    });
+  });
+
   describe("store option", () => {
     test("should include store: false when muxProviderOptions sets store to false", () => {
       const result = buildProviderOptions("openai:gpt-5", "medium", undefined, undefined, {

src/common/utils/ai/providerOptions.ts

@@ -23,6 +23,7 @@ import {
   GEMINI_THINKING_BUDGETS,
   OPENAI_REASONING_EFFORT,
   OPENROUTER_REASONING_EFFORT,
+  openaiSupportsNativeMaxEffort,
 } from "@/common/types/thinking";
 import { isGeminiFlashThinkingLevelModelName } from "@/common/utils/thinking/policy";
 import { resolveModelForMetadata } from "@/common/utils/providers/modelEntries";
@@ -353,8 +354,6 @@ export function buildProviderOptions(
 
   // Build OpenAI-specific options
   if (formatProvider === "openai") {
-    const reasoningEffort = OPENAI_REASONING_EFFORT[effectiveThinking];
-
     // Mux always sends the latest conversation history explicitly. OpenAI's
     // previous_response_id is an alternative state-management path, not an additive one.
     // Chaining it on top of explicit history double-counts prior turns and caused GPT-5.4
@@ -374,6 +373,15 @@ export function buildProviderOptions(
     const truncationMode = openaiTruncationMode ?? "disabled";
     const shouldSendReasoningSummary = supportsOpenAIReasoningSummary(capModelName);
 
+    // GPT-5.6 supports a native "max" reasoning effort for demanding tasks that need
+    // more exploration and verification. The Responses schema in @ai-sdk/openai accepts
+    // arbitrary effort strings, so "max" passes through unchanged; the Chat Completions
+    // schema enum rejects it, so that path keeps the shared max→"xhigh" clamp.
+    const reasoningEffort =
+      effectiveThinking === "max" && isResponses && openaiSupportsNativeMaxEffort(capModelName)
+        ? "max"
+        : OPENAI_REASONING_EFFORT[effectiveThinking];
+
     log.debug("buildProviderOptions: OpenAI config", {
       reasoningEffort,
       shouldSendReasoningSummary,

src/common/utils/thinking/policy.test.ts

@@ -186,6 +186,31 @@ describe("getThinkingPolicyForModel", () => {
     ]);
   });
 
+  test.each(["gpt-5.6-sol", "gpt-5.6-terra", "gpt-5.6-luna"])(
+    "returns all 6 levels including native max for %s",
+    (modelId) => {
+      expect(getThinkingPolicyForModel(`openai:${modelId}`)).toEqual([
+        "off",
+        "low",
+        "medium",
+        "high",
+        "xhigh",
+        "max",
+      ]);
+    }
+  );
+
+  test("returns all 6 levels for gpt-5.6-sol behind mux-gateway with version suffix", () => {
+    expect(getThinkingPolicyForModel("mux-gateway:openai/gpt-5.6-sol-2026-06-26")).toEqual([
+      "off",
+      "low",
+      "medium",
+      "high",
+      "xhigh",
+      "max",
+    ]);
+  });
+
   test("returns 5 levels including xhigh for gpt-5.5", () => {
     expect(getThinkingPolicyForModel("openai:gpt-5.5")).toEqual([
       "off",

src/common/utils/thinking/policy.ts

@@ -19,6 +19,7 @@ import {
   THINKING_LEVEL_OFF,
   anthropicRejectsDisabledThinking,
   anthropicSupportsNativeXhigh,
+  openaiSupportsNativeMaxEffort,
   stripModelProviderPrefixes,
   type ThinkingLevel,
   type ParsedThinkingInput,
@@ -54,6 +55,7 @@ export function isGeminiFlashThinkingLevelModelName(modelName: string): boolean
  * - openai:gpt-5.3-codex / Spark variants →
  *   ["off", "low", "medium", "high", "xhigh"] (5 levels including xhigh)
  * - openai:gpt-5.2 / openai:gpt-5.5 → ["off", "low", "medium", "high", "xhigh"]
+ * - gpt-5.6 tiers (sol/terra/luna) → all 6 levels including native "max"
  * - openai:gpt-5.2-pro / openai:gpt-5.5-pro → ["medium", "high", "xhigh"] (3 levels)
  * - openai:gpt-5-pro → ["high"] (only supported level, legacy)
  * - Gemini Flash chat variants → ["off", "low", "medium", "high"]
@@ -125,7 +127,16 @@ function getExplicitThinkingPolicy(modelString: string): ThinkingPolicy | null {
     return ["medium", "high", "xhigh"];
   }
 
-  // gpt-5.2, gpt-5.5 and the gpt-5.4-mini / gpt-5.4-nano variants support 5 reasoning levels including xhigh.
+  // GPT-5.6 tiers (sol/terra/luna) support all 6 levels: OpenAI documents a native
+  // "max" reasoning effort for demanding tasks that need more exploration and
+  // verification. On the Responses API we send "max" through as-is; the Chat
+  // Completions path clamps to "xhigh" (see buildProviderOptions).
+  if (openaiSupportsNativeMaxEffort(withoutProviderNamespace)) {
+    return ["off", "low", "medium", "high", "xhigh", "max"];
+  }
+
+  // gpt-5.2, gpt-5.5 and the gpt-5.4-mini / gpt-5.4-nano variants support 5 reasoning
+  // levels including xhigh.
   if (
     /^gpt-5\.2(?!-[a-z])/.test(withoutProviderNamespace) ||
     /^gpt-5\.(?:4|5)(?:-(?:mini|nano))?(?!-[a-z])/.test(withoutProviderNamespace)

src/common/utils/tokens/models-extra.ts

@@ -261,6 +261,59 @@ export const modelsExtra: Record<string, ModelData> = {
     supported_endpoints: ["/v1/responses"],
   },
 
+  // GPT-5.6 family (Sol / Terra / Luna) - Previewed June 26, 2026; GA July 9, 2026.
+  // Published pricing per 1M tokens: Sol $5/$30, Terra $2.50/$15, Luna $1/$6.
+  // Prompt caching: cache reads keep the 90% discount; cache writes for GPT-5.6+ bill
+  // at 1.25x the uncached input rate (with explicit breakpoints and a 30-min minimum
+  // cache life). No long-context tiered pricing has been published for this family.
+  // OpenAI has not published official context specs in the GA announcement; we mirror
+  // GPT-5.5's window (1.05M in / 128K out) as a provisional value until the model card
+  // lands. (Preview reports suggest Sol may be larger, ~1.5M, but that's unconfirmed.)
+  "gpt-5.6-sol": {
+    max_input_tokens: 1050000,
+    max_output_tokens: 128000,
+    input_cost_per_token: 0.000005, // $5 per million input tokens
+    output_cost_per_token: 0.00003, // $30 per million output tokens
+    cache_read_input_token_cost: 0.0000005, // $0.50 per million cached input tokens (90% off)
+    cache_creation_input_token_cost: 0.00000625, // $6.25 per million tokens (1.25x input)
+    litellm_provider: "openai",
+    mode: "chat",
+    supports_function_calling: true,
+    supports_vision: true,
+    supports_reasoning: true,
+    supports_response_schema: true,
+  },
+
+  "gpt-5.6-terra": {
+    max_input_tokens: 1050000,
+    max_output_tokens: 128000,
+    input_cost_per_token: 0.0000025, // $2.50 per million input tokens
+    output_cost_per_token: 0.000015, // $15 per million output tokens
+    cache_read_input_token_cost: 0.00000025, // $0.25 per million cached input tokens (90% off)
+    cache_creation_input_token_cost: 0.000003125, // $3.125 per million tokens (1.25x input)
+    litellm_provider: "openai",
+    mode: "chat",
+    supports_function_calling: true,
+    supports_vision: true,
+    supports_reasoning: true,
+    supports_response_schema: true,
+  },
+
+  "gpt-5.6-luna": {
+    max_input_tokens: 1050000,
+    max_output_tokens: 128000,
+    input_cost_per_token: 0.000001, // $1 per million input tokens
+    output_cost_per_token: 0.000006, // $6 per million output tokens
+    cache_read_input_token_cost: 0.0000001, // $0.10 per million cached input tokens (90% off)
+    cache_creation_input_token_cost: 0.00000125, // $1.25 per million tokens (1.25x input)
+    litellm_provider: "openai",
+    mode: "chat",
+    supports_function_calling: true,
+    supports_vision: true,
+    supports_reasoning: true,
+    supports_response_schema: true,
+  },
+
   // GPT-5.4 mini - Released March 11, 2026
   // Smaller/faster gpt-5.4-mini tier with 400K context, 128K max output, and published
   // pricing of $0.75/M input, $4.50/M output, and $0.075/M cached input.

src/node/services/agentSkills/builtInSkillContent.generated.ts

@@ -3097,8 +3097,11 @@ export const BUILTIN_SKILL_FILES: Record<string, Record<string, string>> = {
       "| Opus 4.8               | anthropic:claude-opus-4-8     | `opus`                                                       | ✓       |",
       "| Sonnet 5               | anthropic:claude-sonnet-5     | `sonnet`                                                     |         |",
       "| Haiku 4.5              | anthropic:claude-haiku-4-5    | `haiku`                                                      |         |",
-      "| GPT-5.5                | openai:gpt-5.5                | `gpt`, `gpt-5.5`                                             |         |",
+      "| GPT-5.5                | openai:gpt-5.5                | `gpt-5.5`                                                    |         |",
       "| GPT-5.5 Pro            | openai:gpt-5.5-pro            | `gpt-pro`, `gpt-5.5-pro`                                     |         |",
+      "| GPT-5.6 Sol            | openai:gpt-5.6-sol            | `gpt`, `sol`, `gpt-5.6-sol`                                  |         |",
+      "| GPT-5.6 Terra          | openai:gpt-5.6-terra          | `terra`, `gpt-5.6-terra`                                     |         |",
+      "| GPT-5.6 Luna           | openai:gpt-5.6-luna           | `luna`, `gpt-5.6-luna`                                       |         |",
       "| GPT-5.4 Mini           | openai:gpt-5.4-mini           | `gpt-mini`                                                   |         |",
       "| GPT-5.4 Nano           | openai:gpt-5.4-nano           | `gpt-nano`                                                   |         |",
       "| Codex 5.3              | openai:gpt-5.3-codex          | `codex`, `codex-5.3`                                         |         |",