Charge xAI's real rate for the slugs it retired on 2026-05-15

CLAUDE.md

@@ -124,15 +124,31 @@ Server configuration:
 
 ### Supported Providers
 
-Model name prefixes determine routing:
-- **OpenAI**: gpt-6-astra, gpt-4.1, gpt-5, gpt-5-mini, gpt-5.2, gpt-5.6-sol/terra/luna, o4-mini; image generation: gpt-image-2.5-flare, gpt-image-2.5-sunburst, gpt-image-2
-- **Anthropic**: claude-sonnet-4-0/4-5/4-6, claude-sonnet-5, claude-haiku-4-5, claude-opus-4-5/4-6/4-7/4-8, claude-opus-5, claude-fable-5, claude-fable-5-1, claude-3-7-sonnet, claude-3-5-haiku
-- **Google**: gemini-3.8-flash, gemini-3.7-flash, gemini-3.6-flash, gemini-3.5-flash-lite, gemini-2.5-flash, gemini-2.5-flash-lite, gemini-2.5-pro, gemini-3-pro-preview, gemini-3-flash-preview, gemini-3.1-pro-preview, gemini-3.5-flash; image generation: gemini-2.5-flash-image, gemini-3.1-flash-image
-- **xAI**: grok-2, grok-3, grok-3-mini, grok-4, grok-4.3, grok-4.5, grok-4.6, grok-4-fast, grok-4-1-fast; image generation: grok-2-image, grok-imagine-image-2.0
+`model_registry.py` is the source of truth; this list mirrors it. Model name
+prefixes determine routing:
+- **OpenAI**: gpt-6-astra, gpt-4.1, gpt-4.1-mini, gpt-5, gpt-5-mini, gpt-5.2, gpt-5.4, gpt-5.4-mini, gpt-5.4-nano, gpt-5.5, gpt-5.6-sol/terra/luna, o3; image generation: gpt-image-2.5-flare, gpt-image-2.5-sunburst, gpt-image-2
+- **Anthropic**: claude-sonnet-4-5/4-6, claude-sonnet-5, claude-haiku-4-5, claude-opus-4-5/4-6/4-7/4-8, claude-opus-5, claude-fable-5, claude-fable-5-1
+- **Google**: gemini-3.8-flash, gemini-3.7-flash, gemini-3.6-flash, gemini-3.5-flash, gemini-3.5-flash-lite, gemini-2.5-flash, gemini-2.5-flash-lite, gemini-2.5-pro, gemini-3-flash-preview, gemini-3.1-pro-preview; image generation: gemini-2.5-flash-image, gemini-3.1-flash-image
+- **xAI**: grok-4.6, grok-4.5, grok-4.3, grok-4.20-reasoning, grok-4.20-non-reasoning; image generation: grok-2-image, grok-imagine-image-2.0
 - **ByteDance** (BytePlus ModelArk, OpenAI-compatible, ap-southeast): seed-1.6, seed-1.8, seed-2.0-lite, deepseek-v4-flash, deepseek-v4-pro, glm-5.2 (Z.ai's model served via a ModelArk deployment endpoint); image generation: seedream-4.0, seedream-5.0-lite, seedance-4.5, seedance-5.0
 - **OpenRouter** (OpenAI-compatible): hy4-preview, hermes-4-405b, hermes-4-70b, hy3
 - **Z.ai** (Model API, OpenAI-compatible): image generation: glm-image (glm-5.2 chat is routed through BytePlus ModelArk, see ByteDance above)
 
+Models kept registered but no longer offered to new clients — each still
+resolves so older SDK versions keep working, and each is priced at what the
+provider actually bills, not at its pre-retirement rate:
+- **Retired by xAI on 2026-05-15**: grok-4, grok-4-fast, grok-4-1-fast,
+  grok-4-1-fast-non-reasoning, grok-code-fast-1, grok-3, grok-3-mini. xAI still
+  accepts these slugs but silently redirects them (to grok-4.3, or
+  grok-build-0.1 for grok-code-fast-1) and bills every one at grok-4.3's rate.
+- **Shutting down at OpenAI on 2026-10-23**: gpt-4.1-nano, o4-mini. OpenAI
+  names those exact slugs, so they stop resolving that day; replacements are
+  gpt-5.6-luna and gpt-5.6-terra.
+- **Watch 2026-12-11**: o3, gpt-5 and gpt-5-mini are undated aliases of
+  snapshots retiring then. OpenAI does not document whether such an alias is
+  repointed or retired with its snapshot; if repointed, they become mispriced
+  the way the xAI slugs were.
+
 Image generation via OpenAI (gpt-image-2.5-flare, gpt-image-2.5-sunburst, gpt-image-2), xAI (grok-2-image, grok-imagine-image-2.0), ByteDance
 (seedream-4.0, seedream-5.0-lite, seedance-4.5, seedance-5.0), and Z.ai (glm-image) is served
 through a provider `/images/generations` endpoint rather than the chat path (see

README.md

@@ -33,13 +33,22 @@ The gateway solves this by running inside a hardware-isolated Nitro Enclave wher
 
 | Provider | Models |
 |----------|--------|
-| OpenAI | gpt-6-astra, gpt-4.1, gpt-5, gpt-5-mini, gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, o4-mini |
-| Anthropic | claude-fable-5-1, claude-sonnet-4-5, claude-sonnet-4-6, claude-haiku-4-5, claude-opus-4-5, claude-opus-4-6 |
-| Google | gemini-3.8-flash, gemini-3.7-flash, gemini-3.6-flash, gemini-3.5-flash-lite, gemini-2.5-flash, gemini-2.5-flash-lite, gemini-2.5-pro, gemini-3-pro-preview, gemini-3-flash-preview |
-| xAI | grok-4.6, grok-4.5, grok-4.3, grok-4, grok-4-fast, grok-4-1-fast, grok-4-1-fast-non-reasoning |
-| ByteDance | seed-1.6, seed-1.8, seed-2.0-lite, deepseek-v4-flash, deepseek-v4-pro |
+| OpenAI | gpt-6-astra, gpt-4.1, gpt-4.1-mini, gpt-5, gpt-5-mini, gpt-5.2, gpt-5.4, gpt-5.5, gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, o3 |
+| Anthropic | claude-fable-5-1, claude-fable-5, claude-opus-5, claude-opus-4-8, claude-opus-4-7, claude-opus-4-6, claude-opus-4-5, claude-sonnet-5, claude-sonnet-4-6, claude-sonnet-4-5, claude-haiku-4-5 |
+| Google | gemini-3.8-flash, gemini-3.7-flash, gemini-3.6-flash, gemini-3.5-flash, gemini-3.5-flash-lite, gemini-2.5-flash, gemini-2.5-flash-lite, gemini-2.5-pro, gemini-3-flash-preview, gemini-3.1-pro-preview |
+| xAI | grok-4.6, grok-4.5, grok-4.3, grok-4.20-reasoning, grok-4.20-non-reasoning |
+| ByteDance | seed-1.6, seed-1.8, seed-2.0-lite, dola-seed-2.0-mini, deepseek-v4-flash, deepseek-v4-pro, glm-5.2 |
 | OpenRouter | hy4-preview, hermes-4-405b, hermes-4-70b, hy3 |
 
+Image generation: gpt-image-2, gpt-image-2.5-flare, gpt-image-2.5-sunburst,
+gemini-2.5-flash-image, gemini-3.1-flash-image, grok-2-image,
+grok-imagine-image-2.0, seedream-4.0, seedream-5.0-lite, seedance-4.5,
+seedance-5.0, glm-image.
+
+Some retired models stay registered so older SDK versions keep working; see
+"Supported Providers" in CLAUDE.md. `tee_gateway/model_registry.py` is the
+source of truth for all of the above.
+
 ## Quick Start
 
 ### Local Development (without TEE)
@@ -89,7 +98,7 @@ curl -X POST http://127.0.0.1:8000/v1/chat/completions \
 curl -X POST http://127.0.0.1:8000/v1/completions \
   -H "Content-Type: application/json" \
   -d '{
-    "model": "claude-3.7-sonnet",
+    "model": "claude-sonnet-4-6",
     "prompt": "Explain quantum computing in one sentence"
   }'
 

tee_gateway/model_registry.py

@@ -169,6 +169,12 @@ def image_tier(self, resolution: Optional[str]) -> Optional[ImageResolutionTier]
 # settled (see WebSearchOutcome.billable).
 WEB_SEARCH_PRICE_USD: Decimal = Decimal("0.015")
 
+# What xAI charges for any slug it retired on 2026-05-15: every such request is
+# redirected to a live model and billed at grok-4.3's rate ($1.25 in / $2.50 out
+# per MTok), whatever name was sent.
+_XAI_RETIRED_INPUT_USD: Decimal = Decimal("0.00000125")
+_XAI_RETIRED_OUTPUT_USD: Decimal = Decimal("0.0000025")
+
 # ByteDance ModelArk image *deployment* endpoints (api_name "ep-…", e.g. Seedance
 # 4.5, Seedream 5.0 Lite) return the URL response format and require these extra
 # params. The gateway fetches the returned URL and inlines the bytes, so the
@@ -302,12 +308,21 @@ class SupportedModel(Enum):
         input_price_usd=Decimal("0.0000004"),
         output_price_usd=Decimal("0.0000016"),
     )
+    # Shuts down 2026-10-23 (OpenAI names this exact slug, not a dated
+    # snapshot, so it stops resolving that day). Replacement: gpt-5.6-luna.
+    # Dropped from the SDK already; kept here for older SDK versions until the
+    # date. Same for o4-mini below (replacement: gpt-5.6-terra).
     GPT_4_1_NANO = ModelConfig(
         provider="openai",
         api_name="gpt-4.1-nano",
         input_price_usd=Decimal("0.0000001"),
         output_price_usd=Decimal("0.0000004"),
     )
+    # o3, gpt-5 and gpt-5-mini are undated aliases of snapshots that shut down
+    # on 2026-12-11 (o3-2025-04-16, gpt-5-2025-08-07, gpt-5-mini-2025-08-07).
+    # OpenAI does not document whether such an alias is repointed to a
+    # successor or retired with its snapshot. If it is repointed, these become
+    # mispriced the way the retired xAI slugs were, so revisit before then.
     O3 = ModelConfig(
         provider="openai",
         api_name="o3",
@@ -558,9 +573,14 @@ class SupportedModel(Enum):
     )
 
     # ── Google Gemini ───────────────────────────────────────────────────
-    # Note: gemini-2.5-flash, gemini-2.5-pro, and gemini-2.5-flash-lite are scheduled
-    # for deprecation on June 17, 2026 (flash-lite: July 22, 2026). Use the Gemini 3
-    # replacements below for new integrations.
+    # Note: gemini-2.5-flash, gemini-2.5-pro and gemini-2.5-flash-lite are
+    # still served. A previous note here said they were "scheduled for
+    # deprecation on June 17, 2026 (flash-lite: July 22, 2026)"; those dates
+    # passed and Google's deprecation list does not carry them, so the note was
+    # wrong. The Gemini 3 models below are still the better default for new
+    # integrations. (Google did retire gemini-3.1-flash-image-preview on
+    # 2026-06-25 — the registry calls the GA gemini-3.1-flash-image, not that
+    # preview.)
     GEMINI_2_5_FLASH = ModelConfig(
         provider="google",
         api_name="gemini-2.5-flash",
@@ -685,29 +705,43 @@ class SupportedModel(Enum):
         input_price_usd=Decimal("0.00000125"),
         output_price_usd=Decimal("0.0000025"),
     )
+    # ── xAI models retired on 2026-05-15 ────────────────────────────────
+    # xAI retired these slugs (docs.x.ai/developers/migration/may-15-retirement)
+    # but still ACCEPTS them: each is silently redirected to a live model, and
+    # every redirected request is billed at grok-4.3's rate regardless of the
+    # name sent. They are kept registered so older SDK versions keep working
+    # (the current SDK no longer offers them), but they must be priced at what
+    # xAI actually charges — $1.25/$2.50 per MTok — or the gateway eats the
+    # difference. They were previously priced at their pre-retirement rates,
+    # which undercharged grok-4-fast by 6.25x on input and 5x on output.
+    # Remove once no client sends them.
+    # -> grok-4.3 (low reasoning effort)
     GROK_4 = ModelConfig(
         provider="x-ai",
         api_name="grok-4",
-        input_price_usd=Decimal("0.000003"),
-        output_price_usd=Decimal("0.000015"),
+        input_price_usd=_XAI_RETIRED_INPUT_USD,
+        output_price_usd=_XAI_RETIRED_OUTPUT_USD,
     )
+    # -> grok-4.3 (low reasoning effort)
     GROK_4_FAST = ModelConfig(
         provider="x-ai",
         api_name="grok-4-fast",
-        input_price_usd=Decimal("0.0000002"),
-        output_price_usd=Decimal("0.0000005"),
+        input_price_usd=_XAI_RETIRED_INPUT_USD,
+        output_price_usd=_XAI_RETIRED_OUTPUT_USD,
     )
+    # -> grok-4.3 (low reasoning effort)
     GROK_4_1_FAST = ModelConfig(
         provider="x-ai",
         api_name="grok-4-1-fast",
-        input_price_usd=Decimal("0.0000002"),
-        output_price_usd=Decimal("0.0000005"),
+        input_price_usd=_XAI_RETIRED_INPUT_USD,
+        output_price_usd=_XAI_RETIRED_OUTPUT_USD,
     )
+    # -> grok-4.3 (reasoning effort "none")
     GROK_4_1_FAST_NON_REASONING = ModelConfig(
         provider="x-ai",
         api_name="grok-4-1-fast-non-reasoning",
-        input_price_usd=Decimal("0.0000002"),
-        output_price_usd=Decimal("0.0000005"),
+        input_price_usd=_XAI_RETIRED_INPUT_USD,
+        output_price_usd=_XAI_RETIRED_OUTPUT_USD,
     )
     GROK_4_20_REASONING = ModelConfig(
         provider="x-ai",
@@ -721,11 +755,13 @@ class SupportedModel(Enum):
         input_price_usd=Decimal("0.000002"),
         output_price_usd=Decimal("0.000006"),
     )
+    # Retired 2026-05-15 -> grok-build-0.1, billed at grok-4.3's rate. See the
+    # retired-slug note above.
     GROK_CODE_FAST_1 = ModelConfig(
         provider="x-ai",
         api_name="grok-code-fast-1",
-        input_price_usd=Decimal("0.0000002"),
-        output_price_usd=Decimal("0.0000015"),
+        input_price_usd=_XAI_RETIRED_INPUT_USD,
+        output_price_usd=_XAI_RETIRED_OUTPUT_USD,
     )
     # Image generation via xAI's OpenAI-compatible /images/generations endpoint.
     # grok-2-image-1212 was retired in February 2026; grok-imagine-image is its
@@ -959,17 +995,22 @@ class SupportedModel(Enum):
     )
 
     # ── Legacy models (not in current SDK — retained for older SDK versions) ──
+    # grok-3 was retired on 2026-05-15 alongside the slugs above and is
+    # redirected to grok-4.3 (reasoning effort "none"), billed at grok-4.3's
+    # rate. grok-3-mini is not named in that notice but is no longer offered on
+    # xAI's models page, so it is priced the same way rather than at its old
+    # rate. See the retired-slug note above.
     GROK_3_MINI = ModelConfig(
         provider="x-ai",
         api_name="grok-3-mini",
-        input_price_usd=Decimal("0.0000003"),
-        output_price_usd=Decimal("0.0000005"),
+        input_price_usd=_XAI_RETIRED_INPUT_USD,
+        output_price_usd=_XAI_RETIRED_OUTPUT_USD,
     )
     GROK_3 = ModelConfig(
         provider="x-ai",
         api_name="grok-3-latest",
-        input_price_usd=Decimal("0.000003"),
-        output_price_usd=Decimal("0.000015"),
+        input_price_usd=_XAI_RETIRED_INPUT_USD,
+        output_price_usd=_XAI_RETIRED_OUTPUT_USD,
     )
 
 

tests/test_pricing.py

@@ -413,8 +413,33 @@ def test_grok_4_20_non_reasoning_resolves(self):
     def test_grok_code_fast_1_resolves(self):
         cfg = get_model_config("grok-code-fast-1")
         self.assertEqual(cfg.provider, "x-ai")
-        self.assertEqual(cfg.input_price_usd, Decimal("0.0000002"))
-        self.assertEqual(cfg.output_price_usd, Decimal("0.0000015"))
+        # Retired 2026-05-15 and redirected to grok-build-0.1; xAI bills it at
+        # grok-4.3's rate, so that is what we charge.
+        self.assertEqual(cfg.input_price_usd, Decimal("0.00000125"))
+        self.assertEqual(cfg.output_price_usd, Decimal("0.0000025"))
+
+    def test_retired_xai_slugs_priced_at_grok_4_3_rate(self):
+        """Every slug xAI retired on 2026-05-15 bills at grok-4.3's rate.
+
+        xAI redirects each of these to a live model and bills the request at
+        grok-4.3's price whatever name was sent, so pricing them at their old
+        pre-retirement rates made the gateway eat the difference (grok-4-fast
+        was 6.25x under on input, 5x on output).
+        """
+        live = get_model_config("grok-4.3")
+        for name in (
+            "grok-4",
+            "grok-4-fast",
+            "grok-4-1-fast",
+            "grok-4-1-fast-non-reasoning",
+            "grok-code-fast-1",
+            "grok-3",
+            "grok-3-mini",
+        ):
+            with self.subTest(model=name):
+                cfg = get_model_config(name)
+                self.assertEqual(cfg.input_price_usd, live.input_price_usd)
+                self.assertEqual(cfg.output_price_usd, live.output_price_usd)
 
     def test_grok_imagine_image_2_0_resolves(self):
         cfg = get_model_config("grok-imagine-image-2.0")
@@ -873,15 +898,17 @@ def test_grok_4_cost(self):
         cost = self._calc("grok-4", 1000, 500)
         expected = _expected_cost_opg("grok-4", 1000, 500)
         self.assertEqual(cost, expected)
-        # Same pricing tier as claude-sonnet-4-5
-        self.assertEqual(cost, 10_500_000_000_000_000)
+        # Retired slug, billed by xAI at grok-4.3's rate:
+        # 1000*0.00000125 + 500*0.0000025 = 0.00125 + 0.00125 = 0.0025 USD
+        self.assertEqual(cost, 2_500_000_000_000_000)
 
     def test_grok_4_fast_cost(self):
         cost = self._calc("grok-4-fast", 1000, 500)
         expected = _expected_cost_opg("grok-4-fast", 1000, 500)
         self.assertEqual(cost, expected)
-        # 1000*0.0000002 + 500*0.0000005 = 0.0002 + 0.00025 = 0.00045 USD
-        self.assertEqual(cost, 450_000_000_000_000)
+        # Retired slug, billed by xAI at grok-4.3's rate:
+        # 1000*0.00000125 + 500*0.0000025 = 0.00125 + 0.00125 = 0.0025 USD
+        self.assertEqual(cost, 2_500_000_000_000_000)
 
     def test_grok_4_1_fast_cost(self):
         cost = self._calc("grok-4-1-fast", 1000, 500)
@@ -902,8 +929,9 @@ def test_grok_code_fast_1_cost(self):
         cost = self._calc("grok-code-fast-1", 1000, 500)
         expected = _expected_cost_opg("grok-code-fast-1", 1000, 500)
         self.assertEqual(cost, expected)
-        # 1000*0.0000002 + 500*0.0000015 = 0.0002 + 0.00075 = 0.00095 USD = 9.5e14 wei
-        self.assertEqual(cost, 950_000_000_000_000)
+        # Retired slug, billed by xAI at grok-4.3's rate:
+        # 1000*0.00000125 + 500*0.0000025 = 0.0025 USD = 2.5e15 wei
+        self.assertEqual(cost, 2_500_000_000_000_000)
 
     def test_grok_3_mini_cost(self):
         cost = self._calc("grok-3-mini", 1000, 500)
@@ -981,10 +1009,13 @@ def test_gemini_flash_lite_cheaper_than_flash(self):
         flash = self._calc("gemini-2.5-flash", 1000, 1000)
         self.assertLess(lite, flash)
 
-    def test_grok_4_fast_cheaper_than_grok_4(self):
-        fast = self._calc("grok-4-fast", 1000, 1000)
-        full = self._calc("grok-4", 1000, 1000)
-        self.assertLess(fast, full)
+    def test_grok_4_3_cheaper_than_grok_4_6(self):
+        # grok-4 / grok-4-fast were retired on 2026-05-15 and both now bill at
+        # grok-4.3's rate, so the old fast-vs-full comparison is vacuous. The
+        # live tiers still hold this ordering.
+        cheap = self._calc("grok-4.3", 1000, 1000)
+        flagship = self._calc("grok-4.6", 1000, 1000)
+        self.assertLess(cheap, flagship)
 
 
 class TestCalculateSessionCostEdgeCases(unittest.TestCase):