Replace GPT 5.5 with GPT 5.6 Sol and route OpenAI models through the Responses API

apps/cli/ai/providers.ts

@@ -168,10 +168,10 @@ const AI_PROVIDER_DEFINITIONS: Record< AiProviderId, AiProviderDefinition > = {
 			}
 			env.ANTHROPIC_CUSTOM_HEADERS = buildAnthropicCustomHeaders( anthropicHeaders );
 
-			// OpenAI completions path. The wpcom proxy accepts the same bearer token and
+			// OpenAI Responses path. The wpcom proxy accepts the same bearer token and
 			// dispatches to the right upstream based on the request path.
-			// @ai-sdk/openai expects baseURL to include /v1 (like the real
-			// OpenAI API), so the request path becomes /v1/chat/completions —
+			// The OpenAI SDK expects baseURL to include /v1 (like the real
+			// OpenAI API), so the request path becomes /v1/responses —
 			// mirroring the Anthropic path's /v1/messages.
 			env.OPENAI_BASE_URL = `${ gatewayBaseUrl.replace( /\/+$/, '' ) }/v1`;
 			env.OPENAI_API_KEY = accessToken;

apps/cli/ai/runtimes/pi/index.ts

@@ -59,7 +59,7 @@ import type { AskUserHandler, SiteInfo } from 'cli/ai/types';
 
 // eslint-disable-next-line @typescript-eslint/no-explicit-any
 type AgentToolAny = AgentTool< any >;
-type StudioModel = Model< 'openai-completions' > | Model< 'anthropic-messages' >;
+type StudioModel = Model< 'openai-responses' > | Model< 'anthropic-messages' >;
 type ProviderConfigInput = Parameters< ModelRegistry[ 'registerProvider' ] >[ 1 ];
 
 const STUDIO_WPCOM_ANTHROPIC_PROVIDER = 'studio-wpcom-anthropic';
@@ -363,13 +363,22 @@ function buildModel(
 	};
 
 	if ( family === 'openai' ) {
+		// GPT-5.6 models reject function tools on /v1/chat/completions unless
+		// reasoning is disabled; the Responses API supports tools + reasoning.
+		// GPT-5.6 Sol's real context window is 1.05M tokens, but we declare
+		// 272K — the threshold where OpenAI's 2x long-context pricing kicks
+		// in — so compaction keeps sessions below it. Understating the window
+		// is also load-bearing for correctness: pi clamps max output tokens to
+		// the declared window minus its (post-compaction, sometimes stale)
+		// context estimate, and a too-small window can clamp all the way down
+		// to 1, which the API rejects with a 400.
 		return {
 			...common,
-			api: 'openai-completions',
+			api: 'openai-responses',
 			provider: 'openai',
-			reasoning: false,
-			contextWindow: 200_000,
-			maxTokens: 16_384,
+			reasoning: true,
+			contextWindow: 272_000,
+			maxTokens: 32_000,
 		};
 	}
 	return {

apps/cli/ai/slash-commands.ts

@@ -335,8 +335,8 @@ export const AI_CHAT_SLASH_COMMANDS: SlashCommandDef[] = [
 			// Build options and a reverse lookup at the same time so we never
 			// have to recover the model id from the label. A startsWith-based
 			// match is buggy when one model's label is a prefix of another's
-			// (e.g. "GPT 5.5" prefixes "GPT 5.5 Pro" — picking Pro silently
-			// returns the non-pro id), so we keep the label → id mapping
+			// (e.g. "GPT 5.6" prefixes "GPT 5.6 Sol" — picking Sol silently
+			// returns the other id), so we keep the label → id mapping
 			// explicit here and look up by exact match below.
 			const labelToId = new Map< string, AiModelId >();
 			const modelOptions = availableModels.map( ( id ) => {

apps/cli/ai/tests/pi-runtime.test.ts

@@ -65,9 +65,9 @@ const DEFAULT_MOCK_EVENTS: AgentSessionEvent[] = [
 		message: {
 			role: 'assistant',
 			content: [ { type: 'text', text: 'mocked openai response' } ],
-			api: 'openai-completions',
+			api: 'openai-responses',
 			provider: 'openai',
-			model: 'gpt-5.5',
+			model: 'gpt-5.6-sol',
 			usage: {
 				input: 0,
 				output: 0,
@@ -85,9 +85,9 @@ const DEFAULT_MOCK_EVENTS: AgentSessionEvent[] = [
 		message: {
 			role: 'assistant',
 			content: [ { type: 'text', text: 'mocked openai response' } ],
-			api: 'openai-completions',
+			api: 'openai-responses',
 			provider: 'openai',
-			model: 'gpt-5.5',
+			model: 'gpt-5.6-sol',
 			usage: {
 				input: 0,
 				output: 0,
@@ -115,9 +115,9 @@ const assistantMessage = (
 		message: {
 			role: 'assistant',
 			content,
-			api: 'openai-completions',
+			api: 'openai-responses',
 			provider: 'openai',
-			model: 'gpt-5.5',
+			model: 'gpt-5.6-sol',
 			usage: {
 				input: 0,
 				output: 0,
@@ -224,7 +224,7 @@ describe( 'pi runtime', () => {
 		const events = await runRuntime( {
 			prompt: 'hello',
 			env: {},
-			model: 'gpt-5.5',
+			model: 'gpt-5.6-sol',
 			session: newSession(),
 		} );
 
@@ -249,7 +249,7 @@ describe( 'pi runtime', () => {
 				OPENAI_API_KEY: 'sk-test',
 				OPENAI_BASE_URL: 'https://proxy.example.com/v1',
 			},
-			model: 'gpt-5.5',
+			model: 'gpt-5.6-sol',
 			session: newSession(),
 		} );
 
@@ -265,7 +265,7 @@ describe( 'pi runtime', () => {
 				OPENAI_API_KEY: 'sk-test',
 				OPENAI_BASE_URL: 'https://proxy.example.com/v1',
 			},
-			model: 'gpt-5.5',
+			model: 'gpt-5.6-sol',
 			session: newSession(),
 		} );
 
@@ -279,7 +279,7 @@ describe( 'pi runtime', () => {
 				OPENAI_API_KEY: 'sk-test',
 				OPENAI_BASE_URL: 'https://proxy.example.com/v1',
 			},
-			model: 'gpt-5.5',
+			model: 'gpt-5.6-sol',
 			session: newSession(),
 		} );
 
@@ -336,7 +336,7 @@ describe( 'pi runtime', () => {
 				OPENAI_API_KEY: 'sk-test',
 				OPENAI_BASE_URL: 'https://proxy.example.com/v1',
 			},
-			model: 'gpt-5.5',
+			model: 'gpt-5.6-sol',
 			session: newSession(),
 			activeSite: {
 				name: 'Remote',
@@ -367,7 +367,7 @@ describe( 'pi runtime', () => {
 				OPENAI_API_KEY: 'sk-test',
 				OPENAI_BASE_URL: 'https://proxy.example.com/v1',
 			},
-			model: 'gpt-5.5',
+			model: 'gpt-5.6-sol',
 			session: newSession(),
 		} );
 
@@ -388,7 +388,7 @@ describe( 'pi runtime', () => {
 		const session = newSession();
 		const otherOpenAiModel = 'gpt-test-other' as AiModelId;
 
-		await runRuntime( { prompt: 'hi', env, model: 'gpt-5.5', session } );
+		await runRuntime( { prompt: 'hi', env, model: 'gpt-5.6-sol', session } );
 		await runRuntime( { prompt: 'follow-up', env, model: otherOpenAiModel, session } );
 		await runRuntime( {
 			prompt: 'still on the second model',
@@ -398,7 +398,7 @@ describe( 'pi runtime', () => {
 		} );
 
 		expect( mocks.createdSessions.map( ( s ) => s.state.model.id ) ).toEqual( [
-			'gpt-5.5',
+			'gpt-5.6-sol',
 			otherOpenAiModel,
 			otherOpenAiModel,
 		] );
@@ -469,7 +469,7 @@ describe( 'pi runtime', () => {
 					OPENAI_BASE_URL: 'https://proxy.example.com/v1',
 					STUDIO_OPENAI_DEFAULT_HEADERS: '{not json',
 				},
-				model: 'gpt-5.5',
+				model: 'gpt-5.6-sol',
 				session: newSession(),
 			} );
 

apps/cli/ai/tests/slash-commands.test.ts

@@ -351,12 +351,12 @@ function buildModelCtx(
 		// `as never` keeps this test framework-agnostic — we don't need a
 		// real AiChatUI, just the methods the /model handler touches.
 		ui: {
-			currentModel: overrides.currentModel ?? 'gpt-5.5',
+			currentModel: overrides.currentModel ?? 'gpt-5.6-sol',
 			askUser: vi.fn().mockResolvedValue( { 0: overrides.askUserResponse } ),
 			showInfo: vi.fn(),
 			showError: vi.fn(),
 		} as never,
-		currentModel: overrides.currentModel ?? 'gpt-5.5',
+		currentModel: overrides.currentModel ?? 'gpt-5.6-sol',
 		currentProvider: 'wpcom',
 		showCapabilitiesOnConnect: false,
 		switchProvider: vi.fn().mockResolvedValue( undefined ),
@@ -372,13 +372,13 @@ describe( '/model slash command', () => {
 	// Locks in the labelToId-map matcher introduced in slash-commands.ts.
 	// The previous implementation used `selectedLabel.startsWith( label )`,
 	// which silently picked the wrong id whenever one model's label was a
-	// prefix of another's (e.g. "GPT 5.5" prefix of "GPT 5.5 Pro"). We don't
+	// prefix of another's (e.g. "GPT 5.6" prefix of "GPT 5.6 Sol"). We don't
 	// currently expose any prefix-colliding labels, but the fix is general
 	// and we want the regression coverage to survive future additions.
 	it( 'resolves the picked model exactly by label, not by prefix', async () => {
 		expect( modelHandler ).toBeDefined();
 		const { ctx, persistMock } = buildModelCtx( {
-			currentModel: 'gpt-5.5',
+			currentModel: 'gpt-5.6-sol',
 			askUserResponse: 'Sonnet 5',
 		} );
 
@@ -390,14 +390,14 @@ describe( '/model slash command', () => {
 
 	it( 'still resolves the picked model when its label carries the "(current)" suffix', async () => {
 		const { ctx, persistMock } = buildModelCtx( {
-			currentModel: 'gpt-5.5',
-			askUserResponse: 'GPT 5.5 (current)',
+			currentModel: 'gpt-5.6-sol',
+			askUserResponse: 'GPT 5.6 Sol (current)',
 		} );
 
 		await modelHandler!( '/model', ctx );
 
 		// Same model picked → no swap, no persist.
-		expect( ctx.currentModel ).toBe( 'gpt-5.5' );
+		expect( ctx.currentModel ).toBe( 'gpt-5.6-sol' );
 		expect( persistMock ).not.toHaveBeenCalled();
 	} );
 } );

apps/ui/src/ui-classic/components/session-view/composer/index.test.tsx

@@ -283,13 +283,13 @@ describe( 'Composer menu', () => {
 		);
 
 		fireEvent.click( screen.getByRole( 'button', { name: 'Select model' } ) );
-		fireEvent.click( await screen.findByText( 'GPT 5.5' ) );
+		fireEvent.click( await screen.findByText( 'GPT 5.6 Sol' ) );
 		fireEvent.click( await screen.findByRole( 'button', { name: 'Start new conversation' } ) );
 
 		await waitFor( () => {
 			expect( onSwitchSession ).toHaveBeenCalledWith( 'fresh-session' );
 		} );
-		expect( connectorMocks.setSessionModel ).toHaveBeenCalledWith( 'fresh-session', 'gpt-5.5' );
+		expect( connectorMocks.setSessionModel ).toHaveBeenCalledWith( 'fresh-session', 'gpt-5.6-sol' );
 
 		const loadedSession = queryClient.getQueryData< LoadedAiSession >( [
 			...SESSIONS_QUERY_KEY,
@@ -299,7 +299,7 @@ describe( 'Composer menu', () => {
 		expect( loadedSession?.entries ).toEqual( [
 			expect.objectContaining( {
 				type: 'model_change',
-				modelId: 'gpt-5.5',
+				modelId: 'gpt-5.6-sol',
 			} ),
 		] );
 	} );

packages/common/ai/models.ts

@@ -13,15 +13,14 @@ export interface AiModel {
 }
 
 // Pro / o-series OpenAI variants (`gpt-*-pro`, `o[1-9]*`) are intentionally
-// excluded: they're reasoning-only models that, with our current Chat
-// Completions path through the wpcom AI proxy, return HTTP 400 ("only
-// supported in v1/responses"). Re-enable once we either route them through
-// `/v1/responses` server-side or extend the proxy/SDK timeout window for
-// reasoning turns.
+// excluded: their long reasoning turns can exceed the proxy/SDK timeout
+// window. (Routing is no longer a blocker — the OpenAI family now goes
+// through the proxy's `/v1/responses` path, which supports reasoning models
+// and function tools.)
 export const AI_MODELS = [
 	{ id: 'claude-sonnet-5', label: 'Sonnet 5', family: 'anthropic' },
 	{ id: 'claude-opus-4-8', label: 'Opus 4.8', family: 'anthropic' },
-	{ id: 'gpt-5.5', label: 'GPT 5.5', family: 'openai' },
+	{ id: 'gpt-5.6-sol', label: 'GPT 5.6 Sol', family: 'openai' },
 ] as const satisfies readonly AiModel[];
 
 export type AiModelId = ( typeof AI_MODELS )[ number ][ 'id' ];