Add the Meta Muse Spark 1.3 model card and metadata

app/src/modelMeta.ts

@@ -173,6 +173,7 @@ export const MODEL_RELEASE_DATES: Record<string, string> = {
   "qwen3.8-max": "2026-08-03",
   inkling: "2026-07-15",
   "muse-wonder": "2026-10-12",
+  "muse-spark-1.3": "2026-09-02",
 };
 
 export const PROVIDER_LABELS: Record<ProviderKey, string> = {

policybench/config.py

@@ -113,6 +113,13 @@
     # meta/muse-wonder-20261012; openrouter.ai/api/v1/models/meta/
     # muse-wonder/endpoints, read 2026-10-12).
     "muse-wonder": "openrouter/meta/muse-wonder",
+    # Meta Muse Spark 1.3 (2026-09-02). Like Muse Wonder it runs through
+    # OpenRouter, whose one endpoint for the id is Meta itself (provider
+    # "Meta", serving meta/muse-spark-1.3-20260902; openrouter.ai/api/v1/
+    # models/meta/muse-spark-1.3/endpoints, read 2026-10-13). The cheaper
+    # meta/muse-spark-1.3-contributor listing is Meta's contributor mode, a
+    # different endpoint with its own rate card; the board runs standard mode.
+    "muse-spark-1.3": "openrouter/meta/muse-spark-1.3",
 }
 
 # Configured per-1M-token USD list prices, keyed by the predictions.csv.gz model
@@ -173,6 +180,12 @@
     # the OpenRouter listing, both read 2026-10-12); litellm's price map
     # does not yet cover the id.
     "muse-wonder": {"input": 4.0, "output": 20.0, "cache_read": 0.20},
+    # muse-spark-1.3: $1.25 / $4.25 per 1M input/output tokens, $0.15 cached
+    # input, standard mode, one rate at every prompt length ("No long-context
+    # premium"; Meta's rate card, dev.meta.ai/docs/pricing-rate-limits, and
+    # the OpenRouter endpoint listing, both read 2026-10-13); litellm's price
+    # map is not relied on for the id.
+    "muse-spark-1.3": {"input": 1.25, "output": 4.25, "cache_read": 0.15},
     # gemini-3.6-flash: $1.50 / $7.50 per 1M input/output tokens (Google
     # launch pricing, 2026-07-21); litellm's price map does not yet cover it.
     "gemini-3.6-flash": {"input": 1.5, "output": 7.5},

policybench/model_cards.py

@@ -601,6 +601,41 @@ class ModelCard:
             "cache reads."
         ),
     ),
+    "openrouter/meta/muse-spark-1.3": ModelCard(
+        litellm_id="openrouter/meta/muse-spark-1.3",
+        answer_contract="json",
+        thinking_budget=True,
+        expected_cost_per_scenario_usd=0.02,
+        notes=(
+            "Onboarded 2026-10-13; Meta released it on 2026-09-02. OpenRouter "
+            "lists one endpoint, Meta's own (meta/muse-spark-1.3-20260902), so "
+            "OpenRouter passes requests to Meta's API. As for Muse Wonder, "
+            "that API accepts tool_choice 'auto' only, and Meta rejects forced "
+            "tool choices: a named-function or 'required' tool choice returned "
+            'a 400 on each of three tries (\'only "auto" is supported for '
+            "tool_choice'), so the row runs the JSON contract whole-scenario. "
+            "A request with no reasoning parameter reasons at the provider "
+            "default: on a one-question probe, 4,311, 7,684 and 8,174 of "
+            "4,690, 8,073 and 8,586 completion tokens with no tools, and "
+            "4,844, 5,481 and 6,022 of 5,055, 5,667 and 6,266 with "
+            "tool_choice auto. No reasoning-suppression sensitivity run is "
+            "needed: the row already reasons on the JSON contract. Gauntlet: "
+            "without the reasoning budget the 3-variable JSON probe stopped at "
+            "the token limit (384 tokens); with it, 3/3 and 16/16 parsed on "
+            "the JSON contract (8,163 and 4,258 completion tokens, 106s and "
+            "58s), about $0.020 per scenario at $1.25/$4.25 per 1M with $0.15 "
+            "cache reads. A four-household smoke on the run path (16, 20, 28 "
+            "and 32 outputs) parsed 96/96 in one request each. The board run "
+            "answered all 100 households in one request each (four replayed "
+            "from the smoke, below), the largest "
+            "(scenario_008, 56 outputs) included: 1,984/1,984 parsed, no "
+            "budget escalation, at most 96s and 9,871 completion tokens per "
+            "household, so the row keeps the thinking class's 300s timeout "
+            "and 16,384-token starting budget. Its four smoke households "
+            "replay the smoke's responses from the harness cache (identical "
+            "requests), as the smoke households of earlier rows do."
+        ),
+    ),
     "openrouter/z-ai/glm-5.2": ModelCard(
         litellm_id="openrouter/z-ai/glm-5.2",
         answer_contract="json",

policybench/paper_results.py

@@ -3345,6 +3345,11 @@ def last_engine_upgrade(self) -> EngineUpgrade:
     # released alongside it, is limited to Muse Code subscriptions and has
     # no API, so the board cannot run it.
     "muse-wonder": "2026-10-12",
+    # research.meta.ai/blog/introducing-muse-spark-1-3 (2026-09-02: "available
+    # today in Muse Code and in Meta Model API"); datanorth.ai 2026-09-03;
+    # OpenRouter lists meta/muse-spark-1.3 from 2026-09-02. Proprietary: no
+    # weights.
+    "muse-spark-1.3": "2026-09-02",
     # openrouter.ai/stealth/ox-alpha ("released on August 20, 2026" as a
     # stealth preview; the 2026-09-01 table carried 08-21 from the listing's
     # first observed day). After the row's run Z.ai identified the preview as

tests/test_eval_no_tools.py

@@ -3087,6 +3087,47 @@ def test_local_claude_registration_works_without_the_remote_cost_map():
     assert all(out["models"].values()), out["models"]
 
 
+def test_muse_spark_1_3_routes_through_openrouter_at_metas_standard_rate():
+    """Muse Spark 1.3 runs through OpenRouter, whose one endpoint for the id
+    is Meta's own, at Meta's standard-mode rate: $1.25 / $4.25 per 1M
+    input/output tokens and $0.15 cached input, at every prompt length. The
+    contributor-mode listing is a different endpoint and is not the board's."""
+    assert MODELS["muse-spark-1.3"] == "openrouter/meta/muse-spark-1.3"
+    assert PRICE_OVERRIDES_PER_1M["muse-spark-1.3"] == {
+        "input": 1.25,
+        "output": 4.25,
+        "cache_read": 0.15,
+    }
+
+
+@pytest.mark.parametrize(
+    ("prompt_tokens", "completion_tokens", "cached_prompt_tokens", "expected"),
+    [
+        (1_000_000, 0, 0, 1.25),
+        (0, 1_000_000, 0, 4.25),
+        (1_000_000, 1_000_000, 0, 5.50),
+        (1_000_000, 0, 1_000_000, 0.15),
+        (1_000_000, 0, 250_000, 0.975),
+        # The smoke's largest household: 2,080 prompt and 6,966 completion
+        # tokens, which Meta billed at $0.0322.
+        (2_080, 6_966, 0, 0.0322055),
+    ],
+)
+def test_muse_spark_1_3_override_reconstructs_token_costs(
+    prompt_tokens, completion_tokens, cached_prompt_tokens, expected
+):
+    reconstructed = _reconstruct_token_cost(
+        model_name="muse-spark-1.3",
+        model_id="openrouter/meta/muse-spark-1.3",
+        prompt_tokens=prompt_tokens,
+        completion_tokens=completion_tokens,
+        cached_prompt_tokens=cached_prompt_tokens,
+    )
+
+    assert reconstructed.usd == pytest.approx(expected)
+    assert reconstructed.is_estimated is False
+
+
 def test_muse_wonder_routes_through_openrouter_at_metas_list_price():
     """Muse Wonder runs through OpenRouter, whose one endpoint for the id is
     Meta's own, at Meta's list price: $4 / $20 per 1M input/output tokens

tests/test_model_cards.py

@@ -59,6 +59,7 @@
     "openrouter/z-ai/glm-5.2": ("json", 3, 300, 16_384),
     "openrouter/minimax/minimax-m3": ("tool", None, 300, 16_384),
     "openrouter/qwen/qwen3.7-max": ("json", 3, 600, 16_384),
+    "openrouter/meta/muse-spark-1.3": ("json", None, 300, 16_384),
     "gemini/gemini-3.8-flash": ("tool", None, 300, 16_384),
     "gemini/gemini-3.5-flash-lite": ("tool", None, 300, 16_384),
     "gpt-6-astra": ("tool", None, 300, 16_384),

tests/test_paper_results.py

@@ -75,6 +75,13 @@ def test_claude_haiku_5_5_release_date_is_recorded():
     assert MODEL_RELEASE_DATES["claude-haiku-5.5"] == "2026-10-07"
 
 
+def test_muse_spark_1_3_release_date_is_recorded():
+    """Meta's post "Introducing Muse Spark 1.3" is dated 2026-09-02 and says
+    the model is available that day in Muse Code and the Meta Model API;
+    OpenRouter lists meta/muse-spark-1.3 from the same day."""
+    assert MODEL_RELEASE_DATES["muse-spark-1.3"] == "2026-09-02"
+
+
 def test_muse_wonder_release_date_is_recorded():
     """Meta released Muse Wonder on 2026-10-12 (TechCrunch and VentureBeat,
     both that day), on its Model API and OpenRouter from launch."""