feat(ai-gateway): remove gemma-4-26b-a4b-it:free, fall kilo-auto/small back to free rotation

apps/web/src/lib/ai-gateway/auto-model/index.ts

@@ -149,17 +149,23 @@ export const KILO_AUTO_BALANCED_MODEL: AutoModel = {
   opencode_settings: undefined,
 };
 
+// INVARIANT: kilo-auto/small metadata must stay the lowest common denominator
+// of every model it can resolve to (google/gemma-4-26b-a4b-it with balance,
+// otherwise the kilo-auto/free rotation). ling-3.0-flash and laguna-s-2.1 are
+// text-only, so supports_images must stay false; max_completion_tokens is
+// bounded by gemma's 16384. Re-check all resolution targets before raising
+// any of these values.
 export const KILO_AUTO_SMALL_MODEL: AutoModel = {
   id: 'kilo-auto/small',
   name: 'Auto Small',
   description: 'Automatically routes your request to a small model.',
   context_length: 262144,
-  max_completion_tokens: 32768,
+  max_completion_tokens: 16384,
   prompt_price: '0.00000005',
   completion_price: '0.0000004',
   input_cache_read_price: '0.000000005',
   input_cache_write_price: undefined,
-  supports_images: true,
+  supports_images: false,
   supports_pdf: false,
   opencode_settings: undefined,
 };

apps/web/src/lib/ai-gateway/auto-model/resolution.test.ts

@@ -13,8 +13,12 @@ import {
   BALANCED_QWEN_MODEL,
   FRONTIER_MODE_TO_MODEL,
   KILO_AUTO_EFFICIENT_MODEL,
+  KILO_AUTO_FREE_MODEL,
+  KILO_AUTO_SMALL_MODEL,
   ORG_AUTO_MODEL,
 } from '@/lib/ai-gateway/auto-model';
+import { GEMMA_4_26B_A4B_IT_ID } from '@/lib/ai-gateway/providers/google';
+import { stepfun_37_flash_free_model } from '@/lib/ai-gateway/providers/stepfun';
 import type { AutoRoutingDecision } from '@kilocode/auto-routing-contracts';
 
 const baseParams = {
@@ -386,6 +390,44 @@ describe('resolveAutoModel — kilo-auto/efficient branch', () => {
   });
 });
 
+describe('resolveAutoModel — kilo-auto/small branch', () => {
+  const smallParams = {
+    ...baseParams,
+    model: KILO_AUTO_SMALL_MODEL.id,
+    apiKind: 'chat_completions' as const,
+  };
+
+  it('resolves to the paid Gemma model when the user has balance', async () => {
+    const result = await resolveAutoModel(smallParams, nullUserPromise, Promise.resolve(100));
+
+    expect(result).toEqual({ kind: 'ok', resolved: { model: GEMMA_4_26B_A4B_IT_ID } });
+  });
+
+  it('falls back to the kilo-auto/free rotation when the user has no balance', async () => {
+    const params = { ...smallParams, sessionId: 'session-1' };
+    const smallResult = await resolveAutoModel(params, nullUserPromise, zeroBalancePromise);
+    const freeResult = await resolveAutoModel(
+      { ...params, model: KILO_AUTO_FREE_MODEL.id },
+      nullUserPromise,
+      zeroBalancePromise
+    );
+
+    expect(smallResult).toEqual(freeResult);
+    expect(smallResult.kind).toBe('ok');
+  });
+
+  it('resolves to a free candidate when the user has no balance', async () => {
+    // The Redis mock returns no OpenRouter models, so the only candidate is
+    // the public Kilo-exclusive free model.
+    const result = await resolveAutoModel(smallParams, nullUserPromise, zeroBalancePromise);
+
+    expect(result).toEqual({
+      kind: 'ok',
+      resolved: { model: stepfun_37_flash_free_model.public_id },
+    });
+  });
+});
+
 describe('resolveAutoModel — Organization Auto branch', () => {
   it('uses exact built-in alias routes before canonical fallback routes', async () => {
     const result = await resolveAutoModel(

apps/web/src/lib/ai-gateway/auto-model/resolution.ts

@@ -1,8 +1,5 @@
 import type { FeatureValue } from '@/lib/feature-detection';
-import {
-  gemma_4_26b_a4b_it_free_model,
-  GEMMA_4_26B_A4B_IT_ID,
-} from '@/lib/ai-gateway/providers/google';
+import { GEMMA_4_26B_A4B_IT_ID } from '@/lib/ai-gateway/providers/google';
 import type {
   GatewayRequest,
   OpenRouterChatCompletionRequest,
@@ -142,6 +139,23 @@ export type ResolveAutoModelResult =
   | { kind: 'no_free_models_available' }
   | { kind: 'organization_auto_configuration_error'; message: string };
 
+async function resolveFreeRotation(
+  apiKind: GatewayRequest['kind'] | null,
+  sessionId: string | null,
+  clientIp: string | null,
+  userPromise: Promise<User | null>
+): Promise<ResolveAutoModelResult> {
+  const candidates = await getAutoFreeCandidates(apiKind);
+  if (candidates.length === 0) {
+    return { kind: 'no_free_models_available' };
+  }
+  const randomNumber = getRandomNumber(
+    'free_routing_' + (sessionId ?? (await userPromise)?.id ?? clientIp),
+    candidates.length
+  );
+  return { kind: 'ok', resolved: { model: candidates[randomNumber] } };
+}
+
 async function resolveOrganizationAutoModel(
   params: ResolveAutoModelParams,
   userPromise: Promise<User | null>,
@@ -282,26 +296,13 @@ export async function resolveAutoModel(
     return await resolveOrganizationAutoModel(params, userPromise, balancePromise);
   }
   if (model === KILO_AUTO_FREE_MODEL.id) {
-    const candidates = await getAutoFreeCandidates(apiKind);
-    if (candidates.length === 0) {
-      return { kind: 'no_free_models_available' };
-    }
-    const randomNumber = getRandomNumber(
-      'free_routing_' + (sessionId ?? (await userPromise)?.id ?? clientIp),
-      candidates.length
-    );
-    return { kind: 'ok', resolved: { model: candidates[randomNumber] } };
+    return resolveFreeRotation(apiKind, sessionId, clientIp, userPromise);
   }
   if (model === KILO_AUTO_SMALL_MODEL.id) {
-    return {
-      kind: 'ok',
-      resolved: {
-        model:
-          (await balancePromise) > 0
-            ? GEMMA_4_26B_A4B_IT_ID
-            : gemma_4_26b_a4b_it_free_model.public_id,
-      },
-    };
+    if ((await balancePromise) > 0) {
+      return { kind: 'ok', resolved: { model: GEMMA_4_26B_A4B_IT_ID } };
+    }
+    return resolveFreeRotation(apiKind, sessionId, clientIp, userPromise);
   }
   if (model === KILO_AUTO_EFFICIENT_MODEL.id) {
     const decision = params.efficientDecision ? await params.efficientDecision() : null;

apps/web/src/lib/ai-gateway/context-overflow.test.ts

@@ -4,7 +4,7 @@ import type {
   GatewayRequest,
   OpenRouterChatCompletionRequest,
 } from '@/lib/ai-gateway/providers/openrouter/types';
-import { gemma_4_26b_a4b_it_free_model } from '@/lib/ai-gateway/providers/google';
+import { stepfun_37_flash_free_model } from '@/lib/ai-gateway/providers/stepfun';
 import { ProxyErrorType } from '@/lib/proxy-error-types';
 
 function chatRequest(body: OpenRouterChatCompletionRequest): GatewayRequest {
@@ -158,16 +158,16 @@ describe('detectContextOverflow', () => {
   });
 
   it('triggers on a generic 500 when our estimate exceeds the window', async () => {
-    // Gemma 4 has context_length 262_144 and max_completion_tokens 32_768.
+    // Step 3.7 Flash has context_length 262_144.
     // This request estimates to more than 282_000 tokens, exceeding the context window.
     const hugeRequest = chatRequest({
-      model: gemma_4_26b_a4b_it_free_model.public_id,
+      model: stepfun_37_flash_free_model.public_id,
       messages: [{ role: 'user', content: 'x'.repeat(1_000_000) }],
       max_tokens: 32_768,
     });
 
     const result = await detectContextOverflow({
-      requestedModel: gemma_4_26b_a4b_it_free_model.public_id,
+      requestedModel: stepfun_37_flash_free_model.public_id,
       request: hugeRequest,
       response: new Response('Internal Server Error', { status: 500 }),
     });
@@ -180,12 +180,12 @@ describe('detectContextOverflow', () => {
 
   it('does not trigger on a 500 when the estimate fits the window', async () => {
     const smallRequest = chatRequest({
-      model: gemma_4_26b_a4b_it_free_model.public_id,
+      model: stepfun_37_flash_free_model.public_id,
       messages: [{ role: 'user', content: 'hi' }],
     });
 
     const result = await detectContextOverflow({
-      requestedModel: gemma_4_26b_a4b_it_free_model.public_id,
+      requestedModel: stepfun_37_flash_free_model.public_id,
       request: smallRequest,
       response: new Response('Internal Server Error', { status: 500 }),
     });

apps/web/src/lib/ai-gateway/forbidden-free-models.ts

@@ -19,7 +19,7 @@ const forbiddenFreeModelIds: ReadonlySet<string> = new Set([
   'google/gemma-3-4b-it:free',
   'google/gemma-3n-e2b-it:free',
   'google/gemma-3n-e4b-it:free',
-  'google/gemma-4-26b-a4b-it:free', // usable through kilo-auto
+  'google/gemma-4-26b-a4b-it:free',
   'google/gemma-4-31b-it:free',
   'kilo/auto-free', // discontinued variant of kilo-auto/free
   'kwaipilot/kat-coder-pro-v2.5:free',

apps/web/src/lib/ai-gateway/models.ts

@@ -21,7 +21,7 @@ import type { KiloExclusiveModel } from '@/lib/ai-gateway/providers/kilo-exclusi
 import { isMuseModel } from '@/lib/ai-gateway/providers/meta';
 import { MINIMAX_CURRENT_MODEL_ID } from '@/lib/ai-gateway/providers/minimax';
 import { KIMI_CURRENT_MODEL_ID } from '@/lib/ai-gateway/providers/moonshotai';
-import { gemma_4_26b_a4b_it_free_model, isGeminiModel } from '@/lib/ai-gateway/providers/google';
+import { isGeminiModel } from '@/lib/ai-gateway/providers/google';
 import { QWEN37_PLUS_MODEL_ID, qwen36_plus_stealth_model } from '@/lib/ai-gateway/providers/qwen';
 import { stepfun_37_flash_free_model } from '@/lib/ai-gateway/providers/stepfun';
 import { isGrokModel } from '@/lib/ai-gateway/providers/xai';
@@ -85,7 +85,6 @@ export function isKiloExclusiveModel(model: string): boolean {
 }
 
 export const kiloExclusiveModels = [
-  gemma_4_26b_a4b_it_free_model,
   ...deepseekDiscountedModels,
   qwen36_plus_stealth_model,
   gpt_5_6_sol_stealth_model,

apps/web/src/lib/ai-gateway/providers/google.ts

@@ -1,26 +1,9 @@
-import type { KiloExclusiveModel } from '@/lib/ai-gateway/providers/kilo-exclusive-model';
-
 export function isGemmaModel(model: string) {
   return model.includes('gemma');
 }
 
 export const GEMMA_4_26B_A4B_IT_ID = 'google/gemma-4-26b-a4b-it';
 
-export const gemma_4_26b_a4b_it_free_model: KiloExclusiveModel = {
-  public_id: 'google/gemma-4-26b-a4b-it:free',
-  display_name: 'Google: Gemma 4 26B A4B (free)',
-  description:
-    'Gemma 4 26B A4B IT is an instruction-tuned Mixture-of-Experts (MoE) model from Google DeepMind. Despite 25.2B total parameters, only 3.8B activate per token during inference — delivering near-31B quality at a fraction of the compute cost.',
-  context_length: 262144,
-  max_completion_tokens: 32768,
-  status: 'hidden', // usable through kilo-auto
-  flags: ['vision', 'vercel-routing'],
-  gateway: 'openrouter',
-  internal_id: GEMMA_4_26B_A4B_IT_ID,
-  pricing: null,
-  inference_provider_restriction: [],
-};
-
 export function isGeminiModel(model: string) {
   return model.includes('gemini');
 }

apps/web/src/lib/ai-gateway/providers/openrouter/index.test.ts

@@ -8,7 +8,6 @@ import {
 import { createMockResponse, mockOpenRouterModels } from '@/tests/helpers/openrouter-models.helper';
 import type { OpenRouterModel } from '@/lib/organizations/organization-types';
 import { qwen36_plus_stealth_model } from '@/lib/ai-gateway/providers/qwen';
-import { gemma_4_26b_a4b_it_free_model } from '@/lib/ai-gateway/providers/google';
 import {
   findKiloExclusiveModel,
   isDeadFreeModel,
@@ -42,6 +41,15 @@ const disabledFreeModel = {
   pricing: null,
 } satisfies KiloExclusiveModel;
 
+const hiddenFreeModel = {
+  ...qwen36_plus_stealth_model,
+  public_id: 'vendor/hidden-free-model',
+  internal_id: 'vendor/hidden-free-model-internal',
+  display_name: 'Hidden Free Kilo Model',
+  status: 'hidden',
+  pricing: null,
+} satisfies KiloExclusiveModel;
+
 function buildModel(overrides: Partial<OpenRouterModel> = {}): OpenRouterModel {
   return {
     id: 'vendor/model',
@@ -192,8 +200,8 @@ describe('shouldSuppressOpenRouterModel', () => {
   });
 
   it('suppresses hidden Kilo-exclusive models from OpenRouter', () => {
-    expect(gemma_4_26b_a4b_it_free_model.status).toBe('hidden');
-    expect(shouldSuppressOpenRouterModel(gemma_4_26b_a4b_it_free_model)).toBe(true);
+    expect(hiddenFreeModel.status).toBe('hidden');
+    expect(shouldSuppressOpenRouterModel(hiddenFreeModel)).toBe(true);
   });
 });
 

apps/web/src/lib/ai-gateway/providers/vercel/mapModelIdToVercel.test.ts

@@ -104,11 +104,9 @@ describe('mapModelIdToVercel', () => {
 
   describe('kilo-exclusive models', () => {
     it('maps an exclusive flagged with vercel-routing to its internal id', () => {
-      // google/gemma-4-26b-a4b-it:free is registered in kiloExclusiveModels
-      // with the 'vercel-routing' flag and internal_id 'google/gemma-4-26b-a4b-it'.
-      expect(mapModelIdToVercel('google/gemma-4-26b-a4b-it:free')).toBe(
-        'google/gemma-4-26b-a4b-it'
-      );
+      // stepfun/step-3.7-flash:free is registered in kiloExclusiveModels
+      // with the 'vercel-routing' flag and internal_id 'stepfun/step-3.7-flash'.
+      expect(mapModelIdToVercel('stepfun/step-3.7-flash:free')).toBe('stepfun/step-3.7-flash');
     });
 
     it('does not use internal_id for exclusives that are not vercel-routed', () => {

apps/web/src/lib/rewriteModelResponse.test.ts

@@ -8,6 +8,7 @@ import {
 } from './rewriteModelResponse';
 import { isDynamicallyOptedIntoRequestLogging } from '@/lib/ai-gateway/request-logging-opt-ins';
 import { QWEN37_PLUS_MODEL_ID } from '@/lib/ai-gateway/custom-pricing';
+import { stepfun_37_flash_free_model } from '@/lib/ai-gateway/providers/stepfun';
 import { KILO_ORGANIZATION_ID } from '@/lib/organizations/constants';
 import { logExceptInTest } from '@/lib/utils.server';
 
@@ -815,18 +816,18 @@ describe('rewriteModelResponse', () => {
   test('continues stripping cost for free models outside the Kilo organization', async () => {
     const result = await rewriteModelResponse(
       jsonResponse({
-        model: 'google/gemma-4-26b-a4b-it:free',
+        model: stepfun_37_flash_free_model.public_id,
         usage: { cost: 0, is_byok: false },
       }),
-      'google/gemma-4-26b-a4b-it:free',
+      stepfun_37_flash_free_model.public_id,
       'openrouter',
       'chat_completions',
       makeLogging()
     );
 
     expect(result).not.toBeNull();
     expect(await result?.json()).toEqual({
-      model: 'google/gemma-4-26b-a4b-it:free',
+      model: stepfun_37_flash_free_model.public_id,
       usage: {},
     });
   });