Add AI/ML API provider

package.json

@@ -18,6 +18,7 @@
     "test": "bun test",
     "validate": "bun ./packages/core/script/validate.ts",
     "compare:migrations": "bun ./packages/core/script/compare-model-migrations.ts",
+    "aimlapi:sync": "bun ./packages/core/script/sync-models.ts aimlapi",
     "anthropic:sync": "bun ./packages/core/script/sync-models.ts anthropic",
     "baseten:sync": "bun ./packages/core/script/sync-models.ts baseten",
     "deepinfra:sync": "bun ./packages/core/script/sync-models.ts deepinfra",

packages/core/src/sync/index.ts

@@ -7,6 +7,7 @@ import { AuthoredModel, AuthoredModelShape, ModelMetadata } from "../schema.js";
 import { openMissingModelIssues } from "./missing-issues.js";
 import { MissingReasoningOptionsError } from "./missing-reasoning-options.js";
 import { aiand } from "./providers/aiand.js";
+import { aimlapi } from "./providers/aimlapi.js";
 import { ambient } from "./providers/ambient.js";
 import { anthropic } from "./providers/anthropic.js";
 import { baseten } from "./providers/baseten.js";
@@ -141,6 +142,7 @@ export interface SyncResult {
 
 export const providers: {
   aiand: SyncProvider<any>;
+  aimlapi: SyncProvider<any>;
   ambient: SyncProvider<any>;
   anthropic: SyncProvider<any>;
   baseten: SyncProvider<any>;
@@ -180,6 +182,7 @@ export const providers: {
   xai: SyncProvider<any>;
 } = {
   aiand,
+  aimlapi,
   ambient,
   anthropic,
   baseten,
@@ -221,6 +224,7 @@ export const providers: {
 
 export const groups = {
   aggregators: [
+    "aimlapi",
     "crossmodel",
     "edenai",
     "empiriolabs",

packages/core/src/sync/providers/openrouter.ts

@@ -51,6 +51,36 @@ const CANONICAL_PROVIDER_PREFIXES = {
   "zai-org": { provider: "zai", metadata: "zhipuai" },
 } as const;
 
+/**
+ * The first-party provider directory for a metadata lab id.
+ *
+ * `base_model` ids are spelled with the METADATA lab (`zhipuai/glm-4.7-flash`,
+ * `meta/llama-...`) while a lab's own reasoning ladder lives under its
+ * PROVIDER directory (`providers/zai`, `providers/llama`). The two coincide for
+ * most labs and not for these, and a lookup that assumes they coincide fails
+ * silently — the caller falls back to whatever it had, which for a relay means
+ * publishing the host's enum as if it were the lab's.
+ */
+export function providerDirForMetadataLab(metadata: string): string {
+  for (const entry of Object.values(CANONICAL_PROVIDER_PREFIXES)) {
+    if (entry.metadata === metadata) return entry.provider;
+  }
+  return metadata;
+}
+
+/**
+ * The OpenRouter model-directory prefixes that carry a metadata lab, in table
+ * order — e.g. `alibaba` lives under both `alibaba/` and `qwen/` on OpenRouter,
+ * and only the latter is populated.
+ */
+export function openrouterDirsForMetadataLab(metadata: string): string[] {
+  const dirs: string[] = [];
+  for (const [prefix, entry] of Object.entries(CANONICAL_PROVIDER_PREFIXES)) {
+    if (entry.metadata === metadata) dirs.push(prefix);
+  }
+  return dirs.length > 0 ? dirs : [metadata];
+}
+
 export const OpenRouterModel = z.object({
   id: z.string(),
   name: z.string(),

packages/core/test/aimlapi.test.ts

@@ -0,0 +1,121 @@
+import { expect, test } from "bun:test";
+
+import { labEffortValues, labLadder, resolveLadder } from "../src/sync/providers/aimlapi.js";
+import { providerDirForMetadataLab } from "../src/sync/providers/openrouter.js";
+
+// The lookup that silently kept the host enum for every GLM id: `base_model`
+// names the metadata lab, the ladder lives under the provider directory, and
+// for these two they differ.
+test("maps a metadata lab to its first-party provider directory", () => {
+  expect(providerDirForMetadataLab("zhipuai")).toBe("zai");
+  expect(providerDirForMetadataLab("deepseek")).toBe("deepseek");
+  expect(providerDirForMetadataLab("no-such-lab")).toBe("no-such-lab");
+});
+
+test("a lab split across two provider directories is read from both", () => {
+  // Meta's Llama cards live under `llama/`, its Muse cards under `meta/`.
+  expect(providerDirForMetadataLab("meta")).toBe("llama");
+  expect(labLadder("meta/muse-spark-1.1")).toEqual({
+    kind: "effort",
+    values: ["minimal", "low", "medium", "high", "xhigh"],
+    source: "lab",
+  });
+  expect(resolveLadder("meta/muse-spark-1.1", "meta/muse-spark-1.1", ["low", "medium", "high"])).toEqual([
+    "low",
+    "medium",
+    "high",
+  ]);
+});
+
+test("reads the lab ladder through the provider directory, not the model record", () => {
+  // `models/zhipuai/*.toml` has no reasoning_options; `providers/zai/models`
+  // does. A lookup on the wrong one returns undefined and looks like "no lab
+  // ladder", which is a different claim.
+  expect(labEffortValues("zhipuai/glm-5.3-flash")).toEqual(["low", "high", "max"]);
+  expect(labEffortValues("deepseek/deepseek-v4-pro")).toEqual(["high", "max"]);
+});
+
+test("a dated snapshot borrows the ladder of the model it snapshots", () => {
+  // There is no `deepseek-v4-pro-0813.toml` in the lab; without the borrow it
+  // fell through to the host enum and came out wider than its own base.
+  expect(labEffortValues("deepseek/deepseek-v4-pro-0813")).toEqual(labEffortValues("deepseek/deepseek-v4-pro"));
+});
+
+test("a lab entry with toggle only and no effort ladder yields undefined", () => {
+  expect(labEffortValues("zhipuai/glm-4.7-flash")).toBeUndefined();
+});
+
+const HOST = ["none", "low", "medium", "high"] as const;
+
+test("intersects the host enum with the lab ladder and drops the rest", () => {
+  // lab high/max, host none..high: `low` and `medium` go although the host
+  // accepts them — that is the doctrine, and this test pins its cost.
+  expect(resolveLadder("deepseek/deepseek-v4-pro", "deepseek/deepseek-v4-pro", HOST)).toEqual(["none", "high"]);
+});
+
+test("`none` survives only where measured as a real off switch", () => {
+  // Same lab, same host enum; only the measurement differs.
+  expect(resolveLadder("deepseek/deepseek-v4-pro", "deepseek/deepseek-v4-pro", HOST)).toContain("none");
+  expect(resolveLadder("deepseek/deepseek-v4-flash", "deepseek/deepseek-v4-flash", HOST)).toContain("none");
+});
+
+test("`none` is kept where the lab itself lists it", () => {
+  const host = ["none", "low", "medium", "high"];
+  expect(resolveLadder("openai/gpt-5.6-luna", "openai/gpt-5.6-luna", host)).toContain("none");
+});
+
+test("a measurement narrows but never widens past the lab", () => {
+  // gpt-5.4-pro: lab medium/high/xhigh, host low/medium/high, measured medium/high.
+  expect(resolveLadder("openai/gpt-5.4-pro", "openai/gpt-5.4-pro", ["low", "medium", "high"])).toEqual(["medium", "high"]);
+});
+
+// The fallthrough the review found: a toggle-only lab must never turn into a
+// host-schema L/M/H dump. Either the host field is measured live for that id,
+// or the model is unresolved.
+test("a toggle-only lab base never yields the host's effort dump", () => {
+  const host = ["none", "minimal", "low", "medium", "high"];
+  // glm-4.5v: lab toggle-only -> unresolved.
+  expect(resolveLadder("z-ai/glm-4.5v", "zhipuai/glm-4.5v", host)).toBeUndefined();
+  // qwen3.7-max: lab toggle+budget. The host's effort field measured live
+  // (its names select the budget tiers), and it is still unresolved: a
+  // toggle/budget baseline has no effort ladder to intersect with.
+  expect(resolveLadder("alibaba/qwen3.7-max", "alibaba/qwen3.7-max", host)).toBeUndefined();
+  expect(resolveLadder("google/gemini-2.5-flash-lite", "google/gemini-2.5-flash-lite", host)).toBeUndefined();
+});
+
+test("an empty intersection is unresolved, not the host enum", () => {
+  // lab high/max, host offers neither -> skip, do not republish low/medium.
+  expect(resolveLadder("deepseek/deepseek-v4-pro", "deepseek/deepseek-v4-pro", ["low", "medium"])).toBeUndefined();
+});
+
+test("a base with no lab entry falls back to the OpenRouter peer's ladder, intersected", () => {
+  // No `providers/tencent/models/hy3.toml`; OpenRouter lists none/low/high.
+  expect(labLadder("tencent/hy3")).toEqual({ kind: "effort", values: ["none", "low", "high"], source: "peer" });
+  expect(resolveLadder("tencent/hy3", "tencent/hy3", ["low", "medium", "high"])).toEqual(["low", "high"]);
+  // A peer-listed `none` is not taken on trust the way a lab-listed one is.
+  expect(resolveLadder("tencent/hy3", "tencent/hy3", ["none", "low", "medium", "high"])).toEqual(["low", "high"]);
+  // OpenRouter files Alibaba under `qwen/`.
+  expect(labLadder("alibaba/qwen3.8-27b")).toEqual({ kind: "effort", values: ["low", "medium", "xhigh"], source: "peer" });
+});
+
+test("a peer that is toggle-only counts as no effort ladder, not as a licence for the host enum", () => {
+  // No lab entry for Opus 4.1; OpenRouter's card is toggle-only.
+  expect(labLadder("anthropic/claude-opus-4.1")).toEqual({ kind: "no-effort", source: "peer" });
+  expect(resolveLadder("anthropic/claude-opus-4.1", "anthropic/claude-opus-4.1", ["low", "medium", "high"])).toBeUndefined();
+});
+
+test("a base with neither a lab nor a peer entry is unresolved, not the host enum", () => {
+  expect(labLadder("nolab/nothing-here")).toEqual({ kind: "no-entry" });
+  expect(resolveLadder("some/model", "nolab/nothing-here", ["low", "high"])).toBeUndefined();
+});
+
+test("a wire alias resolves through the measurements of the id it answers as", () => {
+  const host = ["none", "low", "medium", "high"];
+  // `deepseek-chat` answers `"model": "deepseek-flash"` exactly as
+  // `deepseek-v4-flash` does, so the probes run against the latter (`none`
+  // switches reasoning off; `medium` is inert) apply to the alias too.
+  expect(resolveLadder("deepseek/deepseek-chat", "deepseek/deepseek-v4-flash", host)).toEqual(
+    resolveLadder("deepseek/deepseek-v4-flash", "deepseek/deepseek-v4-flash", host),
+  );
+  expect(resolveLadder("deepseek/deepseek-chat", "deepseek/deepseek-v4-flash", host)).toEqual(["none", "low", "high"]);
+});

providers/aimlapi/models/alibaba/qwen-max.toml

@@ -0,0 +1,9 @@
+base_model = "alibaba/qwen-max"
+
+[cost]
+input = 2.08
+output = 8.32
+cache_read = 0.416
+
+[limit]
+context = 32_000

providers/aimlapi/models/alibaba/qwen-plus.toml

@@ -0,0 +1,13 @@
+# No caller control: reasoning_effort is not read (an invalid value returns 200),
+# and toggle/budget fields are dropped by the gateway (unknown top-level keys -> 200).
+base_model = "alibaba/qwen-plus"
+reasoning_options = []
+
+[cost]
+input = 0.52
+output = 1.56
+cache_read = 0.104
+
+[limit]
+context = 131_000
+output = 16_384

providers/aimlapi/models/alibaba/qwen3-coder-480b-a35b-instruct.toml

@@ -0,0 +1,9 @@
+base_model = "alibaba/qwen3-coder-480b-a35b-instruct"
+name = "Qwen3 Coder"
+
+[cost]
+input = 1.95
+output = 9.75
+
+[limit]
+context = 262_000

providers/aimlapi/models/alibaba/qwen3-max.toml

@@ -0,0 +1,7 @@
+base_model = "alibaba/qwen3-max"
+name = "Qwen3 Max Instruct"
+
+[cost]
+input = 1.56
+output = 7.8
+cache_read = 0.312

providers/aimlapi/models/alibaba/qwen3-next-80b-a3b-instruct.toml

@@ -0,0 +1,10 @@
+base_model = "alibaba/qwen3-next-80b-a3b-instruct"
+name = "Qwen3 Next 80B A3B Instruct"
+
+[cost]
+input = 0.195
+output = 1.56
+
+[limit]
+context = 129_024
+output = 16_384

providers/aimlapi/models/alibaba/qwen3.6-max-preview.toml

@@ -0,0 +1,11 @@
+# No caller control: reasoning_effort is validated (invalid value -> 400) but low/high
+# do not order reasoning tokens; toggle/budget fields are dropped by the gateway.
+base_model = "alibaba/qwen3.6-max-preview"
+reasoning_options = []
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.65
+output = 3.9

providers/aimlapi/models/alibaba/qwen3.8-2.4t-a95b.toml

@@ -0,0 +1,16 @@
+base_model = "alibaba/qwen3.8-2.4t-a95b"
+
+[interleaved]
+field = "reasoning_details"
+
+[[reasoning_options]]
+type = "effort"
+values = ["low", "medium"]
+
+[cost]
+input = 3.4385
+output = 8.59625
+cache_read = 0.6877
+
+[limit]
+output = 52_429

providers/aimlapi/models/alibaba/qwen3.8-27b.toml

@@ -0,0 +1,15 @@
+base_model = "alibaba/qwen3.8-27b"
+
+[interleaved]
+field = "reasoning_details"
+
+[[reasoning_options]]
+type = "effort"
+values = ["low", "medium"]
+
+[cost]
+input = 0.61893
+output = 4.40128
+
+[limit]
+output = 131_072

providers/aimlapi/models/alibaba/qwen3.8-flash.toml

@@ -0,0 +1,13 @@
+base_model = "alibaba/qwen3.8-flash"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["low", "medium"]
+
+[cost]
+input = 0.208
+output = 0.611
+cache_read = 0.0208

providers/aimlapi/models/alibaba/qwen3.8-max.toml

@@ -0,0 +1,13 @@
+base_model = "alibaba/qwen3.8-max"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["low", "medium"]
+
+[cost]
+input = 2.6
+output = 7.8
+cache_read = 0.325

providers/aimlapi/models/anthropic/claude-fable-5.toml

@@ -0,0 +1,13 @@
+base_model = "anthropic/claude-fable-5"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["low", "medium", "high"]
+
+[cost]
+input = 13
+output = 65
+cache_read = 1.3

providers/aimlapi/models/anthropic/claude-haiku-4.5.toml

@@ -0,0 +1,10 @@
+# No caller control: reasoning_effort is validated (invalid value -> 400) but low/high
+# do not order reasoning tokens; toggle/budget fields are dropped by the gateway.
+base_model = "anthropic/claude-haiku-4-5"
+name = "Claude 4.5 Haiku"
+reasoning_options = []
+
+[cost]
+input = 1.3754
+output = 6.877
+cache_read = 0.13754

providers/aimlapi/models/anthropic/claude-opus-4.5.toml

@@ -0,0 +1,11 @@
+base_model = "anthropic/claude-opus-4-5"
+name = "Claude 4.5 Opus"
+
+[[reasoning_options]]
+type = "effort"
+values = ["low", "medium", "high"]
+
+[cost]
+input = 6.877
+output = 34.385
+cache_read = 0.6877

providers/aimlapi/models/anthropic/claude-opus-4.7.toml

@@ -0,0 +1,12 @@
+# Off: reasoning_effort = none (0 reasoning tokens); no toggle field on this surface.
+base_model = "anthropic/claude-opus-4-7"
+name = "Claude 4.7 Opus"
+
+[[reasoning_options]]
+type = "effort"
+values = ["none", "low", "medium", "high", "xhigh", "max"]
+
+[cost]
+input = 6.877
+output = 34.385
+cache_read = 0.6877

providers/aimlapi/models/anthropic/claude-opus-4.8.toml

@@ -0,0 +1,12 @@
+# Off: reasoning_effort = none (0 reasoning tokens); no toggle field on this surface.
+base_model = "anthropic/claude-opus-4-8"
+name = "Claude 4.8 Opus"
+
+[[reasoning_options]]
+type = "effort"
+values = ["none", "low", "medium", "high", "xhigh", "max"]
+
+[cost]
+input = 6.877
+output = 34.385
+cache_read = 0.6877