diff --git a/CHANGELOG.md b/CHANGELOG.md
index 28273fd..1d23a6c 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -8,6 +8,43 @@ and the project adheres to [Semantic Versioning 2.0.0](https://semver.org/spec/v
 > Tags prior to **v0.4.0** were cut in the private repository and produced no
 > public artifacts; the first publicly released version is v0.4.0.
 
+## [1.4.1] - 2026-05-30
+
+### Changed
+- **`setup` no longer forces an API-key re-entry when one already exists.**
+  Re-running `commitbrief setup` for a provider that already has a key in the
+  target config now lets you leave the key prompt blank to keep the stored
+  key, so switching only the active provider/model doesn't require retyping
+  credentials. First-time configuration (no existing key) still requires a
+  non-empty key. For a non-interactive path, `providers use <name>` switches
+  the active provider and `config set providers.<name>.model <model>` changes
+  the model — both leave API keys untouched.
+
+### Added
+- **OpenAI: GPT-5 family support — `gpt-5.5`, `gpt-5.4-mini`, `gpt-5.5-pro`.**
+  All three are selectable in `commitbrief setup` and carry correct context
+  windows (1.05M / 400K / 1.05M input) and pricing. The new default OpenAI
+  model is **`gpt-5.4-mini`** (was `gpt-4o`); existing configs that pin a
+  model are unaffected. `gpt-5.5-pro` is served through OpenAI's Responses
+  API (it is not available on Chat Completions) and may take several minutes
+  per review. Reasoning models get a larger default output-token budget so a
+  findings JSON isn't truncated by reasoning tokens.
+- **Gemini: `gemini-3.5-flash` support.**
+
+### Changed
+- **Gemini model lineup refreshed to the 3.x family.** `gemini-2.5-pro` →
+  `gemini-3.1-pro-preview`, `gemini-2.5-flash` → `gemini-3.5-flash`,
+  `gemini-1.5-flash` → `gemini-3.1-flash-lite`. The new default Gemini model
+  is **`gemini-3.5-flash`** (was `gemini-2.5-pro`). Pricing and context
+  windows updated; `gemini-3.1-pro-preview` carries tiered pricing (the
+  ≤200K-token base tier is snapshotted). Configs pinning a removed 2.x/1.5
+  model should switch to a 3.x ID.
+- **Anthropic: `claude-opus-4-7` → `claude-opus-4-8`.** Opus 4.8 is now the
+  default Anthropic model. Its built-in pricing ($5 input / $25 output / $0.50
+  cache-read per 1M) and context window (1M tokens, up from the 200K the old
+  table recorded for 4.7) are updated to match Anthropic's current model docs.
+  Configs pinning `claude-opus-4-7` should switch to `claude-opus-4-8`.
+
 ## [1.4.0] - 2026-05-29
 
 ### Fixed
diff --git a/README.md b/README.md
index 3eb0104..9d3703c 100644
--- a/README.md
+++ b/README.md
@@ -260,9 +260,9 @@ Four API providers + two CLI-tool-backed providers ship in the box:
 
 | Provider | Models | Notes |
 |----------|--------|-------|
-| **Anthropic** | Claude Opus 4.7, Sonnet 4.6, Haiku 4.5 | Ephemeral prompt caching (5 m TTL) cuts repeated input cost ~10×. |
-| **OpenAI** | GPT-4o, GPT-4o-mini | Automatic prompt caching at ≥1024-token prefixes. |
-| **Google Gemini** | Gemini 2.5 Pro (2 M context!), 2.5 Flash, 1.5 Flash | Largest free-tier context windows. |
+| **Anthropic** | Claude Opus 4.8 (default), Sonnet 4.6, Haiku 4.5 | Ephemeral prompt caching (5 m TTL) cuts repeated input cost ~10×. Opus 4.8 advertises a 1 M-token context. |
+| **OpenAI** | GPT-5.4-mini (default), GPT-5.5, GPT-5.5-pro, GPT-4o, GPT-4o-mini | Automatic prompt caching at ≥1024-token prefixes. `gpt-5.5-pro` runs via the Responses API (not Chat Completions) and can take several minutes per review. |
+| **Google Gemini** | Gemini 3.5 Flash (default), 3.1 Pro, 3.1 Flash-Lite | ~1 M-token context windows. `gemini-3.1-pro-preview` is a preview model. |
 | **DeepSeek** | deepseek-chat, deepseek-reasoner | OpenAI-compatible API (`DEEPSEEK_API_KEY`); JSON is prompt-driven (degrades gracefully). |
 | **Mistral** | Mistral Large / Small, Codestral | OpenAI-compatible API (`MISTRAL_API_KEY`). |
 | **Cohere** | Command R+ / R, Command A | Cohere's OpenAI-compatibility endpoint (`COHERE_API_KEY`). |
@@ -361,7 +361,7 @@ Two-tier YAML config with field-level merge:
 
 Plus environment variables for credentials and runtime tweaks, and
 CLI flags for one-off overrides (`--provider gemini --model
-gemini-2.5-flash`).
+gemini-3.5-flash`).
 
 | Variable | Effect |
 |---|---|
@@ -379,13 +379,13 @@ version: 1
 provider: anthropic                # default provider
 providers:
   anthropic:
-    model: claude-opus-4-7
+    model: claude-opus-4-8
     pricing:                         # optional: override built-in $/1M rates
-      claude-opus-4-7:               # (cost preflight / verbose footer / cache)
-        input_per_1m: 15.0
-        output_per_1m: 75.0          # omitted fields keep the built-in value
+      claude-opus-4-8:               # (cost preflight / verbose footer / cache)
+        input_per_1m: 5.0
+        output_per_1m: 25.0          # omitted fields keep the built-in value
   openai:
-    model: gpt-4o
+    model: gpt-5.4-mini
   ollama:
     model: qwen2.5-coder:14b
     base_url: http://localhost:11434
diff --git a/internal/config/defaults.go b/internal/config/defaults.go
index 5c229b5..66be301 100644
--- a/internal/config/defaults.go
+++ b/internal/config/defaults.go
@@ -7,9 +7,9 @@ func Default() *Config {
 		Version:  CurrentSchemaVersion,
 		Provider: "anthropic",
 		Providers: map[string]ProviderConfig{
-			"anthropic": {Model: "claude-opus-4-7", BaseURL: "https://api.anthropic.com"},
-			"openai":    {Model: "gpt-4o", BaseURL: "https://api.openai.com/v1"},
-			"gemini":    {Model: "gemini-2.5-pro"},
+			"anthropic": {Model: "claude-opus-4-8", BaseURL: "https://api.anthropic.com"},
+			"openai":    {Model: "gpt-5.4-mini", BaseURL: "https://api.openai.com/v1"},
+			"gemini":    {Model: "gemini-3.5-flash"},
 			"ollama":    {Model: "qwen2.5-coder:14b", BaseURL: "http://localhost:11434"},
 		},
 		Output: OutputConfig{
diff --git a/internal/i18n/messages.en.yml b/internal/i18n/messages.en.yml
index 4ec3bfe..9dc128b 100644
--- a/internal/i18n/messages.en.yml
+++ b/internal/i18n/messages.en.yml
@@ -3,6 +3,8 @@ setup.provider.prompt: "Which provider would you like to configure?"
 setup.provider.help: "Pick the LLM provider you have an API key for."
 setup.api_key.prompt: "Enter your API key:"
 setup.api_key.empty: "API key cannot be empty."
+setup.api_key.prompt_keep: "Enter a new API key (leave blank to keep the existing one):"
+setup.api_key.help_keep: "A key is already configured for this provider. Leave blank to keep it, or enter a new one to replace it."
 setup.base_url.prompt: "Ollama base URL:"
 setup.model.prompt: "Pick a model:"
 setup.model.discover_failed: "Model name (could not discover from Ollama):"
diff --git a/internal/i18n/messages.tr.yml b/internal/i18n/messages.tr.yml
index 444a586..4428e4b 100644
--- a/internal/i18n/messages.tr.yml
+++ b/internal/i18n/messages.tr.yml
@@ -3,6 +3,8 @@ setup.provider.prompt: "Hangi sağlayıcıyı yapılandırmak istersiniz?"
 setup.provider.help: "API anahtarınız olan LLM sağlayıcısını seçin."
 setup.api_key.prompt: "API anahtarınızı girin:"
 setup.api_key.empty: "API anahtarı boş olamaz."
+setup.api_key.prompt_keep: "Yeni bir API anahtarı girin (mevcut anahtarı korumak için boş bırakın):"
+setup.api_key.help_keep: "Bu sağlayıcı için zaten bir anahtar yapılandırılmış. Korumak için boş bırakın ya da değiştirmek için yenisini girin."
 setup.base_url.prompt: "Ollama temel URL'si:"
 setup.model.prompt: "Bir model seçin:"
 setup.model.discover_failed: "Model adı (Ollama'dan keşfedilemedi):"
diff --git a/internal/provider/anthropic/anthropic_test.go b/internal/provider/anthropic/anthropic_test.go
index ff5a4b9..f587e0f 100644
--- a/internal/provider/anthropic/anthropic_test.go
+++ b/internal/provider/anthropic/anthropic_test.go
@@ -21,7 +21,7 @@ func TestModelsList(t *testing.T) {
 		t.Errorf("Models() length = %d, want 3", len(got))
 	}
 	want := map[string]bool{
-		ModelOpus47:   true,
+		ModelOpus48:   true,
 		ModelSonnet46: true,
 		ModelHaiku45:  true,
 	}
@@ -41,8 +41,8 @@ func TestModelsDefensiveCopy(t *testing.T) {
 }
 
 func TestIsModelSupported(t *testing.T) {
-	if !IsModelSupported(ModelOpus47) {
-		t.Error("Opus 4.7 should be supported")
+	if !IsModelSupported(ModelOpus48) {
+		t.Error("Opus 4.8 should be supported")
 	}
 	if IsModelSupported("gpt-4o") {
 		t.Error("OpenAI model should not be reported as supported")
@@ -53,8 +53,8 @@ func TestIsModelSupported(t *testing.T) {
 }
 
 func TestContextWindow(t *testing.T) {
-	if contextWindowFor(ModelOpus47) != 200_000 {
-		t.Error("Opus 4.7 context window wrong")
+	if contextWindowFor(ModelOpus48) != 1_000_000 {
+		t.Error("Opus 4.8 should advertise the 1M window")
 	}
 	if contextWindowFor(ModelSonnet46) != 1_000_000 {
 		t.Error("Sonnet 4.6 should advertise the 1M window")
@@ -65,7 +65,7 @@ func TestContextWindow(t *testing.T) {
 }
 
 func TestPricingLookup(t *testing.T) {
-	p := pricingFor(ModelOpus47)
+	p := pricingFor(ModelOpus48)
 	if p.InputPer1M == 0 || p.OutputPer1M == 0 {
 		t.Errorf("Opus pricing missing: %+v", p)
 	}
@@ -223,7 +223,7 @@ func messagePayload(content []map[string]any, inputTokens, outputTokens int) map
 		"id":            "msg_test",
 		"type":          "message",
 		"role":          "assistant",
-		"model":         ModelOpus47,
+		"model":         ModelOpus48,
 		"content":       content,
 		"stop_reason":   "end_turn",
 		"stop_sequence": nil,
@@ -253,7 +253,7 @@ func TestReviewWithFakeServerDegradesToText(t *testing.T) {
 		t.Fatal(err)
 	}
 	resp, err := c.Review(context.Background(), provider.Request{
-		Model:        ModelOpus47,
+		Model:        ModelOpus48,
 		SystemPrompt: "rules",
 		UserPrompt:   "diff",
 		MaxTokens:    256,
@@ -291,7 +291,7 @@ func TestReviewWithToolUseFakeServer(t *testing.T) {
 		t.Fatal(err)
 	}
 	resp, err := c.Review(context.Background(), provider.Request{
-		Model:        ModelOpus47,
+		Model:        ModelOpus48,
 		SystemPrompt: "rules",
 		UserPrompt:   "diff",
 		MaxTokens:    256,
diff --git a/internal/provider/anthropic/context_window.go b/internal/provider/anthropic/context_window.go
index 5f7798f..c5b6155 100644
--- a/internal/provider/anthropic/context_window.go
+++ b/internal/provider/anthropic/context_window.go
@@ -5,7 +5,7 @@ package anthropic
 const defaultContextWindow = 200_000
 
 var contextWindows = map[string]int{
-	ModelOpus47:   200_000,
+	ModelOpus48:   1_000_000,
 	ModelSonnet46: 1_000_000,
 	ModelHaiku45:  200_000,
 }
diff --git a/internal/provider/anthropic/models.go b/internal/provider/anthropic/models.go
index cbf30f1..a7c6848 100644
--- a/internal/provider/anthropic/models.go
+++ b/internal/provider/anthropic/models.go
@@ -5,15 +5,15 @@ package anthropic
 const (
 	Name = "anthropic"
 
-	ModelOpus47   = "claude-opus-4-7"
+	ModelOpus48   = "claude-opus-4-8"
 	ModelSonnet46 = "claude-sonnet-4-6"
 	ModelHaiku45  = "claude-haiku-4-5-20251001"
 
-	DefaultModel = ModelOpus47
+	DefaultModel = ModelOpus48
 )
 
 var supportedModels = []string{
-	ModelOpus47,
+	ModelOpus48,
 	ModelSonnet46,
 	ModelHaiku45,
 }
diff --git a/internal/provider/anthropic/pricing.go b/internal/provider/anthropic/pricing.go
index 9b7cdcd..8802813 100644
--- a/internal/provider/anthropic/pricing.go
+++ b/internal/provider/anthropic/pricing.go
@@ -9,10 +9,10 @@ import "github.com/CommitBrief/commitbrief/internal/provider"
 // should be refreshed when a model price changes. Cached input is the
 // "prompt cache read" rate per Anthropic's ephemeral cache discount.
 var pricingTable = map[string]provider.Pricing{
-	ModelOpus47: {
-		InputPer1M:       15.00,
-		OutputPer1M:      75.00,
-		CachedInputPer1M: 1.50,
+	ModelOpus48: {
+		InputPer1M:       5.00,
+		OutputPer1M:      25.00,
+		CachedInputPer1M: 0.50,
 	},
 	ModelSonnet46: {
 		InputPer1M:       3.00,
diff --git a/internal/provider/gemini/context_window.go b/internal/provider/gemini/context_window.go
index e89a3b9..7c6e0f9 100644
--- a/internal/provider/gemini/context_window.go
+++ b/internal/provider/gemini/context_window.go
@@ -4,10 +4,14 @@ package gemini
 
 const defaultContextWindow = 1_000_000
 
+// Gemini 3.x family baseline input window is 1M tokens. Exact published
+// limits for these preview models were not yet broken out in the model
+// docs at integration time; 1M is the documented floor and the safe value
+// for the over-context guard. Bump if Google publishes larger windows.
 var contextWindows = map[string]int{
-	ModelPro2_5:   2_000_000,
-	ModelFlash2_5: 1_000_000,
-	ModelFlash1_5: 1_000_000,
+	ModelPro31:       1_000_000,
+	ModelFlash35:     1_000_000,
+	ModelFlashLite31: 1_000_000,
 }
 
 func contextWindowFor(model string) int {
diff --git a/internal/provider/gemini/gemini_test.go b/internal/provider/gemini/gemini_test.go
index eac9379..9709356 100644
--- a/internal/provider/gemini/gemini_test.go
+++ b/internal/provider/gemini/gemini_test.go
@@ -32,8 +32,8 @@ func TestModelsDefensiveCopy(t *testing.T) {
 }
 
 func TestIsModelSupported(t *testing.T) {
-	if !IsModelSupported(ModelPro2_5) {
-		t.Error("gemini-2.5-pro should be supported")
+	if !IsModelSupported(ModelPro31) {
+		t.Error("gemini-3.1-pro-preview should be supported")
 	}
 	if IsModelSupported("gpt-4o") {
 		t.Error("OpenAI model should not be supported here")
@@ -41,8 +41,8 @@ func TestIsModelSupported(t *testing.T) {
 }
 
 func TestContextWindow(t *testing.T) {
-	if contextWindowFor(ModelPro2_5) != 2_000_000 {
-		t.Errorf("pro 2.5 context window wrong: %d", contextWindowFor(ModelPro2_5))
+	if contextWindowFor(ModelPro31) != 1_000_000 {
+		t.Errorf("pro 3.1 context window wrong: %d", contextWindowFor(ModelPro31))
 	}
 	if contextWindowFor("unknown") != defaultContextWindow {
 		t.Error("unknown model should fall back to defaultContextWindow")
@@ -50,7 +50,7 @@ func TestContextWindow(t *testing.T) {
 }
 
 func TestPricingLookup(t *testing.T) {
-	p := pricingFor(ModelPro2_5)
+	p := pricingFor(ModelPro31)
 	if p.InputPer1M == 0 || p.OutputPer1M == 0 {
 		t.Errorf("pro pricing missing: %+v", p)
 	}
@@ -138,7 +138,7 @@ func TestReviewWithFakeServer(t *testing.T) {
 		t.Fatal(err)
 	}
 	resp, err := c.Review(context.Background(), provider.Request{
-		Model:        ModelPro2_5,
+		Model:        ModelPro31,
 		SystemPrompt: "rules",
 		UserPrompt:   "diff",
 		MaxTokens:    256,
diff --git a/internal/provider/gemini/models.go b/internal/provider/gemini/models.go
index 199cf1c..3aa48f8 100644
--- a/internal/provider/gemini/models.go
+++ b/internal/provider/gemini/models.go
@@ -5,17 +5,20 @@ package gemini
 const (
 	Name = "gemini"
 
-	ModelPro2_5   = "gemini-2.5-pro"
-	ModelFlash2_5 = "gemini-2.5-flash"
-	ModelFlash1_5 = "gemini-1.5-flash"
+	// ModelPro31 is a preview model: the public pricing page lists the Pro
+	// tier only as gemini-3.1-pro-preview. Update to the stable ID once
+	// Google promotes it out of preview.
+	ModelPro31       = "gemini-3.1-pro-preview"
+	ModelFlash35     = "gemini-3.5-flash"
+	ModelFlashLite31 = "gemini-3.1-flash-lite"
 
-	DefaultModel = ModelPro2_5
+	DefaultModel = ModelFlash35
 )
 
 var supportedModels = []string{
-	ModelPro2_5,
-	ModelFlash2_5,
-	ModelFlash1_5,
+	ModelPro31,
+	ModelFlash35,
+	ModelFlashLite31,
 }
 
 func Models() []string {
diff --git a/internal/provider/gemini/pricing.go b/internal/provider/gemini/pricing.go
index 89427ac..6b2d048 100644
--- a/internal/provider/gemini/pricing.go
+++ b/internal/provider/gemini/pricing.go
@@ -4,26 +4,28 @@ package gemini
 
 import "github.com/CommitBrief/commitbrief/internal/provider"
 
-// Gemini per-1M-token pricing snapshot (paid tier).
+// Gemini per-1M-token pricing snapshot (paid "Standard" tier).
 // Source: https://ai.google.dev/gemini-api/docs/pricing
-// CachedInputPer1M reflects Gemini's context-caching discount (separate
-// API; we track the per-token rate so cost reporting is accurate when a
-// cache is wired up in a future phase).
+// gemini-3.1-pro-preview has tiered pricing (≤200K vs >200K input tokens:
+// $2/$12 vs $4/$18); we snapshot the ≤200K base, so the verbose cost footer
+// may under-report on very large inputs. CachedInputPer1M reflects Gemini's
+// implicit-cache read discount (~0.25× input); context caching is not wired
+// into the client yet, so the rate is informational for now.
 var pricingTable = map[string]provider.Pricing{
-	ModelPro2_5: {
-		InputPer1M:       1.25,
-		OutputPer1M:      10.00,
-		CachedInputPer1M: 0.31,
+	ModelPro31: {
+		InputPer1M:       2.00,
+		OutputPer1M:      12.00,
+		CachedInputPer1M: 0.50,
 	},
-	ModelFlash2_5: {
-		InputPer1M:       0.30,
-		OutputPer1M:      2.50,
-		CachedInputPer1M: 0.075,
+	ModelFlash35: {
+		InputPer1M:       1.50,
+		OutputPer1M:      9.00,
+		CachedInputPer1M: 0.375,
 	},
-	ModelFlash1_5: {
-		InputPer1M:       0.075,
-		OutputPer1M:      0.30,
-		CachedInputPer1M: 0.01875,
+	ModelFlashLite31: {
+		InputPer1M:       0.25,
+		OutputPer1M:      1.50,
+		CachedInputPer1M: 0.0625,
 	},
 }
 
diff --git a/internal/provider/openai/client.go b/internal/provider/openai/client.go
index 33b0a17..b312a7e 100644
--- a/internal/provider/openai/client.go
+++ b/internal/provider/openai/client.go
@@ -9,6 +9,7 @@ import (
 
 	sdk "github.com/openai/openai-go"
 	"github.com/openai/openai-go/option"
+	"github.com/openai/openai-go/responses"
 	"github.com/openai/openai-go/shared"
 
 	"github.com/CommitBrief/commitbrief/internal/config"
@@ -18,8 +19,12 @@ import (
 
 const (
 	defaultMaxTokens = 4096
-	testPingPrompt   = "ping"
-	testPingMaxTok   = 8
+	// defaultReasoningMaxTokens is the output ceiling for GPT-5 reasoning
+	// models when the caller specifies none — larger than defaultMaxTokens
+	// because reasoning tokens are billed out of the same budget.
+	defaultReasoningMaxTokens = 16384
+	testPingPrompt            = "ping"
+	testPingMaxTok            = 8
 )
 
 type Client struct {
@@ -69,6 +74,13 @@ func (c *Client) Pricing(model string) provider.Pricing {
 }
 
 func (c *Client) Review(ctx context.Context, req provider.Request) (provider.Response, error) {
+	model := req.Model
+	if model == "" {
+		model = c.DefaultModel()
+	}
+	if usesResponsesAPI(model) {
+		return c.reviewViaResponses(ctx, req, model)
+	}
 	params := c.buildParams(req)
 	completion, err := c.sdk.Chat.Completions.New(ctx, params)
 	if err != nil {
@@ -81,9 +93,37 @@ func (c *Client) Review(ctx context.Context, req provider.Request) (provider.Res
 	}, nil
 }
 
+// reviewViaResponses drives a Responses-API-only model (gpt-5.5-pro). The
+// call is synchronous and may take several minutes for the pro model; it
+// honours the caller's context deadline (the SDK sets no client timeout of
+// its own).
+func (c *Client) reviewViaResponses(ctx context.Context, req provider.Request, model string) (provider.Response, error) {
+	resp, err := c.sdk.Responses.New(ctx, c.buildResponsesParams(req, model))
+	if err != nil {
+		return provider.Response{}, mapError(err)
+	}
+	return provider.Response{
+		Content: resp.OutputText(),
+		Model:   resp.Model,
+		Usage:   mapResponsesUsage(resp.Usage),
+	}, nil
+}
+
 func (c *Client) TestConnection(ctx context.Context) error {
+	model := c.DefaultModel()
+	if usesResponsesAPI(model) {
+		params := responses.ResponseNewParams{
+			Model:           shared.ResponsesModel(model),
+			MaxOutputTokens: sdk.Int(testPingMaxTok),
+			Input:           responses.ResponseNewParamsInputUnion{OfString: sdk.String(testPingPrompt)},
+		}
+		if _, err := c.sdk.Responses.New(ctx, params); err != nil {
+			return mapError(err)
+		}
+		return nil
+	}
 	params := sdk.ChatCompletionNewParams{
-		Model:               shared.ChatModel(c.DefaultModel()),
+		Model:               shared.ChatModel(model),
 		MaxCompletionTokens: sdk.Int(testPingMaxTok),
 		Messages: []sdk.ChatCompletionMessageParamUnion{
 			sdk.UserMessage(testPingPrompt),
@@ -102,7 +142,7 @@ func (c *Client) buildParams(req provider.Request) sdk.ChatCompletionNewParams {
 	}
 	maxTokens := int64(req.MaxTokens)
 	if maxTokens <= 0 {
-		maxTokens = defaultMaxTokens
+		maxTokens = defaultMaxTokensFor(model)
 	}
 
 	messages := make([]sdk.ChatCompletionMessageParamUnion, 0, 2)
@@ -124,6 +164,28 @@ func (c *Client) buildParams(req provider.Request) sdk.ChatCompletionNewParams {
 	return params
 }
 
+func (c *Client) buildResponsesParams(req provider.Request, model string) responses.ResponseNewParams {
+	maxTokens := int64(req.MaxTokens)
+	if maxTokens <= 0 {
+		maxTokens = defaultMaxTokensFor(model)
+	}
+
+	params := responses.ResponseNewParams{
+		Model:           shared.ResponsesModel(model),
+		MaxOutputTokens: sdk.Int(maxTokens),
+		Input:           responses.ResponseNewParamsInputUnion{OfString: sdk.String(req.UserPrompt)},
+	}
+	if req.SystemPrompt != "" {
+		params.Instructions = sdk.String(req.SystemPrompt)
+	}
+	// Structured-findings JSON contract (ADR-0014). Omitted for FreeForm
+	// (ADR-0015) so the model returns a plain-text completion.
+	if !req.FreeForm {
+		params.Text = buildResponsesTextFormat()
+	}
+	return params
+}
+
 func extractText(c *sdk.ChatCompletion) string {
 	if c == nil || len(c.Choices) == 0 {
 		return ""
@@ -139,6 +201,14 @@ func mapUsage(u sdk.CompletionUsage) provider.Usage {
 	}
 }
 
+func mapResponsesUsage(u responses.ResponseUsage) provider.Usage {
+	return provider.Usage{
+		InputTokens:       int(u.InputTokens),
+		OutputTokens:      int(u.OutputTokens),
+		CachedInputTokens: int(u.InputTokensDetails.CachedTokens),
+	}
+}
+
 func mapError(err error) error {
 	if err == nil {
 		return nil
diff --git a/internal/provider/openai/context_window.go b/internal/provider/openai/context_window.go
index e3781c4..41861c9 100644
--- a/internal/provider/openai/context_window.go
+++ b/internal/provider/openai/context_window.go
@@ -7,6 +7,9 @@ const defaultContextWindow = 128_000
 var contextWindows = map[string]int{
 	ModelGPT4o:     128_000,
 	ModelGPT4oMini: 128_000,
+	ModelGPT55:     1_050_000,
+	ModelGPT54Mini: 400_000,
+	ModelGPT55Pro:  1_050_000,
 }
 
 func contextWindowFor(model string) int {
diff --git a/internal/provider/openai/models.go b/internal/provider/openai/models.go
index d9509f9..569fc6c 100644
--- a/internal/provider/openai/models.go
+++ b/internal/provider/openai/models.go
@@ -7,15 +7,53 @@ const (
 
 	ModelGPT4o     = "gpt-4o"
 	ModelGPT4oMini = "gpt-4o-mini"
+	ModelGPT55     = "gpt-5.5"
+	ModelGPT54Mini = "gpt-5.4-mini"
+	ModelGPT55Pro  = "gpt-5.5-pro"
 
-	DefaultModel = ModelGPT4o
+	DefaultModel = ModelGPT54Mini
 )
 
 var supportedModels = []string{
+	ModelGPT54Mini,
+	ModelGPT55,
+	ModelGPT55Pro,
 	ModelGPT4o,
 	ModelGPT4oMini,
 }
 
+// responsesAPIModels are served only through the Responses API rather than
+// Chat Completions (gpt-5.5-pro operates exclusively there per OpenAI's
+// model docs). Review/TestConnection route these calls differently.
+var responsesAPIModels = map[string]bool{
+	ModelGPT55Pro: true,
+}
+
+// usesResponsesAPI reports whether the model must be driven through the
+// Responses API instead of Chat Completions.
+func usesResponsesAPI(model string) bool {
+	return responsesAPIModels[model]
+}
+
+// reasoningModels are the GPT-5 family models that spend reasoning tokens
+// out of the output-token budget. They need a larger default ceiling than
+// the gpt-4o-era 4096 so a findings JSON isn't truncated by reasoning.
+var reasoningModels = map[string]bool{
+	ModelGPT55:     true,
+	ModelGPT54Mini: true,
+	ModelGPT55Pro:  true,
+}
+
+// defaultMaxTokensFor returns the output-token ceiling to use when the
+// caller did not specify one. Reasoning models get a higher budget;
+// everything else keeps the historical gpt-4o default.
+func defaultMaxTokensFor(model string) int64 {
+	if reasoningModels[model] {
+		return defaultReasoningMaxTokens
+	}
+	return defaultMaxTokens
+}
+
 func Models() []string {
 	out := make([]string, len(supportedModels))
 	copy(out, supportedModels)
diff --git a/internal/provider/openai/openai_test.go b/internal/provider/openai/openai_test.go
index 7bff9cd..0cda7c5 100644
--- a/internal/provider/openai/openai_test.go
+++ b/internal/provider/openai/openai_test.go
@@ -17,10 +17,16 @@ import (
 
 func TestModelsList(t *testing.T) {
 	got := Models()
-	if len(got) != 2 {
-		t.Errorf("Models() length = %d, want 2", len(got))
+	want := map[string]bool{
+		ModelGPT4o:     true,
+		ModelGPT4oMini: true,
+		ModelGPT55:     true,
+		ModelGPT54Mini: true,
+		ModelGPT55Pro:  true,
+	}
+	if len(got) != len(want) {
+		t.Errorf("Models() length = %d, want %d", len(got), len(want))
 	}
-	want := map[string]bool{ModelGPT4o: true, ModelGPT4oMini: true}
 	for _, m := range got {
 		if !want[m] {
 			t.Errorf("unexpected model %q", m)
diff --git a/internal/provider/openai/pricing.go b/internal/provider/openai/pricing.go
index eac4817..52873b3 100644
--- a/internal/provider/openai/pricing.go
+++ b/internal/provider/openai/pricing.go
@@ -20,6 +20,26 @@ var pricingTable = map[string]provider.Pricing{
 		OutputPer1M:      0.60,
 		CachedInputPer1M: 0.075,
 	},
+	// gpt-5.5 has tiered pricing: prompts above 272K input tokens incur 2x
+	// input / 1.5x output for the whole session. We snapshot the base
+	// (<=272K) tier; review diffs rarely exceed it, so the verbose cost
+	// footer may under-report on very large inputs.
+	ModelGPT55: {
+		InputPer1M:       5.00,
+		OutputPer1M:      30.00,
+		CachedInputPer1M: 0.50,
+	},
+	ModelGPT54Mini: {
+		InputPer1M:       0.75,
+		OutputPer1M:      4.50,
+		CachedInputPer1M: 0.075,
+	},
+	// gpt-5.5-pro does not offer a cached-input discount.
+	ModelGPT55Pro: {
+		InputPer1M:       30.00,
+		OutputPer1M:      180.00,
+		CachedInputPer1M: 0,
+	},
 }
 
 func pricingFor(model string) provider.Pricing {
diff --git a/internal/provider/openai/structured.go b/internal/provider/openai/structured.go
index 47ac9ac..87a22e1 100644
--- a/internal/provider/openai/structured.go
+++ b/internal/provider/openai/structured.go
@@ -4,6 +4,7 @@ package openai
 
 import (
 	sdk "github.com/openai/openai-go"
+	"github.com/openai/openai-go/responses"
 	"github.com/openai/openai-go/shared"
 )
 
@@ -81,3 +82,20 @@ func buildResponseFormat() sdk.ChatCompletionNewParamsResponseFormatUnion {
 		},
 	}
 }
+
+// buildResponsesTextFormat returns the Responses API equivalent of
+// buildResponseFormat — the same strict findings schema expressed as a
+// `text.format` json_schema config. Used by Responses-API-only models
+// (gpt-5.5-pro).
+func buildResponsesTextFormat() responses.ResponseTextConfigParam {
+	return responses.ResponseTextConfigParam{
+		Format: responses.ResponseFormatTextConfigUnionParam{
+			OfJSONSchema: &responses.ResponseFormatTextJSONSchemaConfigParam{
+				Name:        schemaName,
+				Description: sdk.String("Structured findings for a code review."),
+				Strict:      sdk.Bool(true),
+				Schema:      responseSchema,
+			},
+		},
+	}
+}
diff --git a/internal/setup/setup_test.go b/internal/setup/setup_test.go
index 61175cf..26b7c72 100644
--- a/internal/setup/setup_test.go
+++ b/internal/setup/setup_test.go
@@ -122,6 +122,37 @@ func TestApplyPreservesExistingKeys(t *testing.T) {
 	}
 }
 
+// TestApplyKeepsKeyWhenReconfiguringSameProvider guards the wizard's
+// "leave blank to keep the existing key" behavior: re-running setup for a
+// provider that already has a key, submitting an empty key but a new model,
+// must preserve the stored key and update only the model. This is the
+// contract the promptAPIKey existing-key branch relies on.
+func TestApplyKeepsKeyWhenReconfiguringSameProvider(t *testing.T) {
+	base := config.Default()
+	base.Providers["anthropic"] = config.ProviderConfig{
+		APIKey:  "sk-ant-existing",
+		Model:   "claude-opus-4-8",
+		BaseURL: "https://api.anthropic.com",
+	}
+	base.Provider = "anthropic"
+
+	// Empty APIKey (user left the prompt blank), new model only.
+	cfg := Apply(base, Choices{
+		Provider: "anthropic",
+		Model:    "claude-sonnet-4-6",
+	})
+
+	if got := cfg.Providers["anthropic"].APIKey; got != "sk-ant-existing" {
+		t.Errorf("APIKey not preserved on empty input: got %q, want sk-ant-existing", got)
+	}
+	if got := cfg.Providers["anthropic"].Model; got != "claude-sonnet-4-6" {
+		t.Errorf("Model not updated: got %q, want claude-sonnet-4-6", got)
+	}
+	if got := cfg.Providers["anthropic"].BaseURL; got != "https://api.anthropic.com" {
+		t.Errorf("BaseURL lost: got %q", got)
+	}
+}
+
 func TestApplyFirstRunStartsFromDefault(t *testing.T) {
 	// Passing nil base mimics first-time setup: result must be a clean
 	// Default config with the choices layered on, no leftover state.
diff --git a/internal/setup/wizard.go b/internal/setup/wizard.go
index 065e9b6..7ee280f 100644
--- a/internal/setup/wizard.go
+++ b/internal/setup/wizard.go
@@ -44,21 +44,21 @@ var DefaultSpecs = []ProviderSpec{
 		Name:       "anthropic",
 		Label:      "Anthropic (Claude)",
 		NeedsKey:   true,
-		Models:     []string{"claude-opus-4-7", "claude-sonnet-4-6", "claude-haiku-4-5-20251001"},
+		Models:     []string{"claude-opus-4-8", "claude-sonnet-4-6", "claude-haiku-4-5-20251001"},
 		APIKeyHelp: "Get an API key from https://console.anthropic.com/",
 	},
 	{
 		Name:       "openai",
 		Label:      "OpenAI (GPT)",
 		NeedsKey:   true,
-		Models:     []string{"gpt-4o", "gpt-4o-mini"},
+		Models:     []string{"gpt-5.4-mini", "gpt-5.5", "gpt-5.5-pro", "gpt-4o", "gpt-4o-mini"},
 		APIKeyHelp: "Get an API key from https://platform.openai.com/",
 	},
 	{
 		Name:       "gemini",
 		Label:      "Google Gemini",
 		NeedsKey:   true,
-		Models:     []string{"gemini-2.5-pro", "gemini-2.5-flash", "gemini-1.5-flash"},
+		Models:     []string{"gemini-3.5-flash", "gemini-3.1-pro-preview", "gemini-3.1-flash-lite"},
 		APIKeyHelp: "Get an API key from https://aistudio.google.com/",
 	},
 	{
@@ -198,7 +198,12 @@ func Run(ctx context.Context, opts RunOptions) (*config.Config, error) {
 	}
 
 	if spec.NeedsKey {
-		if err := promptAPIKey(ctx, spec, &choices, opts.Catalog); err != nil {
+		// If the target config already holds a key for the chosen provider,
+		// let the user leave the prompt blank to keep it — switching the
+		// active provider/model shouldn't force a key re-entry. Apply()
+		// preserves the existing key on empty input.
+		hasExistingKey := base != nil && base.Providers[choices.Provider].APIKey != ""
+		if err := promptAPIKey(ctx, spec, &choices, opts.Catalog, hasExistingKey); err != nil {
 			return nil, err
 		}
 	}
@@ -261,15 +266,23 @@ func selectProvider(ctx context.Context, specs []ProviderSpec, choices *Choices,
 	return form.RunWithContext(ctx)
 }
 
-func promptAPIKey(ctx context.Context, spec *ProviderSpec, choices *Choices, cat *i18n.Catalog) error {
-	form := huh.NewForm(huh.NewGroup(
-		huh.NewInput().
+func promptAPIKey(ctx context.Context, spec *ProviderSpec, choices *Choices, cat *i18n.Catalog, hasExistingKey bool) error {
+	input := huh.NewInput().
+		EchoMode(huh.EchoModePassword).
+		Value(&choices.APIKey)
+	if hasExistingKey {
+		// A key already exists: allow an empty submission (Apply keeps the
+		// stored key) so the user can change only the provider/model.
+		input = input.
+			Title(tr(cat, "setup.api_key.prompt_keep", "Enter a new API key (leave blank to keep the existing one):")).
+			Description(tr(cat, "setup.api_key.help_keep", "A key is already configured for this provider. Leave blank to keep it, or enter a new one to replace it."))
+	} else {
+		input = input.
 			Title(tr(cat, "setup.api_key.prompt", "Enter your API key:")).
 			Description(spec.APIKeyHelp).
-			EchoMode(huh.EchoModePassword).
-			Validate(notEmptyFor(cat)).
-			Value(&choices.APIKey),
-	))
+			Validate(notEmptyFor(cat))
+	}
+	form := huh.NewForm(huh.NewGroup(input))
 	return form.RunWithContext(ctx)
 }