diff --git a/crates/tinyagents-registry/src/pricing/data.rs b/crates/tinyagents-registry/src/pricing/data.rs
index 975ca6f58..b7eece79c 100644
--- a/crates/tinyagents-registry/src/pricing/data.rs
+++ b/crates/tinyagents-registry/src/pricing/data.rs
@@ -1,17 +1,19 @@
-//! July 2026 published price and context-window snapshot.
+//! Published price and context-window snapshot, last checked 2026-11-11.
 
-use super::ModelPrice;
+use super::{LongContextTier, ModelPrice};
 
 /// Published list prices and context windows for the default models the product
 /// can route to.
 ///
-/// Sources (captured [`super::PRICING_AS_OF`]): vendor pricing/model pages. Anthropic
-/// price rows are authoritative (cached = 0.1× input, the documented cache-read
-/// rate); other providers' cached rates use the published discount where known
-/// and a conservative provider-typical fraction otherwise. Context windows are
-/// the published maximums.
+/// Each provider block names the page its rows were checked against and the
+/// date ([`super::PRICING_AS_OF`]). Rates are standard (not batch) list prices;
+/// the cached rate is the cache-read price. Models their provider retired are
+/// removed rather than kept at stale prices. Context
+/// windows are the published maximums. Rows with a long-context price step are
+/// the short-context rates; [`LONG_CONTEXT_TIERS`] holds the step.
 pub const KNOWN_MODEL_PRICING: &[ModelPrice] = &[
-    // ── Anthropic (authoritative prices; cache read = 0.1× input) ────────────
+    // ── Anthropic: platform.claude.com/docs/en/about-claude/pricing (2026-11-11).
+    // Cache read = 0.1× input; 4.6 and later bill the full 1M window flat.
     ModelPrice {
         provider: "anthropic",
         model_id: "claude-fable-5",
@@ -76,28 +78,29 @@ pub const KNOWN_MODEL_PRICING: &[ModelPrice] = &[
         output_per_mtok_usd: 5.00,
         context_window: 200_000,
     },
-    // ── OpenAI (cache read ≈ 0.25× input — published 75% off) ────────────────
+    // ── OpenAI: developers.openai.com/api/docs/pricing (2026-11-11).
+    // Cache read is 0.1× input on GPT-5.x and 0.25× on GPT-4.1 and o3.
     ModelPrice {
         provider: "openai",
         model_id: "gpt-5.5",
         input_per_mtok_usd: 5.00,
-        cached_input_per_mtok_usd: 1.25,
+        cached_input_per_mtok_usd: 0.50,
         output_per_mtok_usd: 30.00,
-        context_window: 1_000_000,
+        context_window: 1_050_000,
     },
     ModelPrice {
         provider: "openai",
         model_id: "gpt-5.4",
         input_per_mtok_usd: 2.50,
-        cached_input_per_mtok_usd: 0.625,
+        cached_input_per_mtok_usd: 0.25,
         output_per_mtok_usd: 15.00,
-        context_window: 1_000_000,
+        context_window: 1_050_000,
     },
     ModelPrice {
         provider: "openai",
         model_id: "gpt-5.4-mini",
         input_per_mtok_usd: 0.75,
-        cached_input_per_mtok_usd: 0.1875,
+        cached_input_per_mtok_usd: 0.075,
         output_per_mtok_usd: 4.50,
         context_window: 400_000,
     },
@@ -105,7 +108,7 @@ pub const KNOWN_MODEL_PRICING: &[ModelPrice] = &[
         provider: "openai",
         model_id: "gpt-5.4-nano",
         input_per_mtok_usd: 0.20,
-        cached_input_per_mtok_usd: 0.05,
+        cached_input_per_mtok_usd: 0.02,
         output_per_mtok_usd: 1.25,
         context_window: 400_000,
     },
@@ -133,12 +136,14 @@ pub const KNOWN_MODEL_PRICING: &[ModelPrice] = &[
         output_per_mtok_usd: 8.00,
         context_window: 200_000,
     },
-    // ── Google Gemini (cache read ≈ 0.25× input; 1M-token windows) ───────────
+    // ── Google: aistudio.google.com/docs/pricing (2026-11-11). Text cache read
+    // = 0.1× input. 2.5 Pro and Flash are deprecated and re-routed from
+    // 2026-12-01.
     ModelPrice {
         provider: "google",
         model_id: "gemini-2.5-pro",
         input_per_mtok_usd: 1.25,
-        cached_input_per_mtok_usd: 0.3125,
+        cached_input_per_mtok_usd: 0.125,
         output_per_mtok_usd: 10.00,
         context_window: 1_048_576,
     },
@@ -146,7 +151,7 @@ pub const KNOWN_MODEL_PRICING: &[ModelPrice] = &[
         provider: "google",
         model_id: "gemini-2.5-flash",
         input_per_mtok_usd: 0.30,
-        cached_input_per_mtok_usd: 0.075,
+        cached_input_per_mtok_usd: 0.03,
         output_per_mtok_usd: 2.50,
         context_window: 1_048_576,
     },
@@ -154,59 +159,28 @@ pub const KNOWN_MODEL_PRICING: &[ModelPrice] = &[
         provider: "google",
         model_id: "gemini-2.5-flash-lite",
         input_per_mtok_usd: 0.10,
-        cached_input_per_mtok_usd: 0.025,
+        cached_input_per_mtok_usd: 0.01,
         output_per_mtok_usd: 0.40,
         context_window: 1_048_576,
     },
-    // ── DeepSeek (cache hit = 0.1× input, published) ─────────────────────────
-    ModelPrice {
-        provider: "deepseek",
-        model_id: "deepseek-chat",
-        input_per_mtok_usd: 0.14,
-        cached_input_per_mtok_usd: 0.014,
-        output_per_mtok_usd: 0.28,
-        context_window: 128_000,
-    },
-    ModelPrice {
-        provider: "deepseek",
-        model_id: "deepseek-reasoner",
-        input_per_mtok_usd: 0.55,
-        cached_input_per_mtok_usd: 0.055,
-        output_per_mtok_usd: 2.19,
-        context_window: 128_000,
-    },
-    // ── Moonshot Kimi (cache hit published) ──────────────────────────────────
+    // ── Moonshot: platform.kimi.ai/docs/pricing/chat (2026-11-11).
     ModelPrice {
         provider: "moonshot",
         model_id: "kimi-k2.6",
         input_per_mtok_usd: 0.95,
         cached_input_per_mtok_usd: 0.16,
         output_per_mtok_usd: 4.00,
-        context_window: 256_000,
-    },
-    ModelPrice {
-        provider: "moonshot",
-        model_id: "kimi-k2.5",
-        input_per_mtok_usd: 0.60,
-        cached_input_per_mtok_usd: 0.10,
-        output_per_mtok_usd: 3.00,
-        context_window: 256_000,
-    },
-    // ── Qwen / Alibaba (cache read ≈ 0.1× input) ─────────────────────────────
-    ModelPrice {
-        provider: "qwen",
-        model_id: "qwen3-max",
-        input_per_mtok_usd: 1.20,
-        cached_input_per_mtok_usd: 0.12,
-        output_per_mtok_usd: 6.00,
-        context_window: 256_000,
+        context_window: 262_144,
     },
+    // ── Qwen: alibabacloud.com/help/en/model-studio/model-pricing, International
+    // (2026-11-11). Cache read is the explicit-cache rate (0.1× input);
+    // qwen-max has implicit cache only (0.2×).
     ModelPrice {
         provider: "qwen",
         model_id: "qwen-max",
-        input_per_mtok_usd: 1.20,
-        cached_input_per_mtok_usd: 0.12,
-        output_per_mtok_usd: 6.00,
+        input_per_mtok_usd: 1.60,
+        cached_input_per_mtok_usd: 0.32,
+        output_per_mtok_usd: 6.40,
         context_window: 256_000,
     },
     ModelPrice {
@@ -215,7 +189,7 @@ pub const KNOWN_MODEL_PRICING: &[ModelPrice] = &[
         input_per_mtok_usd: 0.40,
         cached_input_per_mtok_usd: 0.04,
         output_per_mtok_usd: 1.20,
-        context_window: 256_000,
+        context_window: 1_000_000,
     },
     ModelPrice {
         provider: "qwen",
@@ -223,32 +197,45 @@ pub const KNOWN_MODEL_PRICING: &[ModelPrice] = &[
         input_per_mtok_usd: 0.05,
         cached_input_per_mtok_usd: 0.005,
         output_per_mtok_usd: 0.40,
-        context_window: 256_000,
+        context_window: 1_000_000,
     },
-    // ── Mistral (cache read ≈ 0.1× input) ────────────────────────────────────
+    // ── Mistral: docs.mistral.ai model cards (2026-11-11). mistral-large is
+    // Large 4 (26.10, published cache price), also listed under its own ids
+    // mistral-large-4 and mistral-large-4-0 (a one-digit suffix is not a
+    // snapshot decoration, so lookup would not reach the alias row); mistral-small is Small 4 (26.03);
+    // ministral-8b is Ministral 3 8B (25.12); codestral is 25.08. Small, 8B and
+    // Codestral publish no cache price, so theirs is an estimate (0.1× input).
     ModelPrice {
         provider: "mistral",
         model_id: "mistral-large",
-        input_per_mtok_usd: 2.00,
-        cached_input_per_mtok_usd: 0.20,
-        output_per_mtok_usd: 6.00,
-        context_window: 128_000,
+        input_per_mtok_usd: 0.70,
+        cached_input_per_mtok_usd: 0.07,
+        output_per_mtok_usd: 2.10,
+        context_window: 1_000_000,
     },
     ModelPrice {
         provider: "mistral",
-        model_id: "mistral-medium",
-        input_per_mtok_usd: 0.40,
-        cached_input_per_mtok_usd: 0.04,
-        output_per_mtok_usd: 2.00,
-        context_window: 128_000,
+        model_id: "mistral-large-4",
+        input_per_mtok_usd: 0.70,
+        cached_input_per_mtok_usd: 0.07,
+        output_per_mtok_usd: 2.10,
+        context_window: 1_000_000,
+    },
+    ModelPrice {
+        provider: "mistral",
+        model_id: "mistral-large-4-0",
+        input_per_mtok_usd: 0.70,
+        cached_input_per_mtok_usd: 0.07,
+        output_per_mtok_usd: 2.10,
+        context_window: 1_000_000,
     },
     ModelPrice {
         provider: "mistral",
         model_id: "mistral-small",
-        input_per_mtok_usd: 0.20,
-        cached_input_per_mtok_usd: 0.02,
+        input_per_mtok_usd: 0.15,
+        cached_input_per_mtok_usd: 0.015,
         output_per_mtok_usd: 0.60,
-        context_window: 128_000,
+        context_window: 256_000,
     },
     ModelPrice {
         provider: "mistral",
@@ -256,38 +243,77 @@ pub const KNOWN_MODEL_PRICING: &[ModelPrice] = &[
         input_per_mtok_usd: 0.30,
         cached_input_per_mtok_usd: 0.03,
         output_per_mtok_usd: 0.90,
-        context_window: 256_000,
+        context_window: 128_000,
     },
     ModelPrice {
         provider: "mistral",
         model_id: "ministral-8b",
-        input_per_mtok_usd: 0.10,
-        cached_input_per_mtok_usd: 0.01,
-        output_per_mtok_usd: 0.10,
-        context_window: 128_000,
+        input_per_mtok_usd: 0.15,
+        cached_input_per_mtok_usd: 0.015,
+        output_per_mtok_usd: 0.15,
+        context_window: 256_000,
     },
+    // ── MiniMax: platform.minimax.io/docs/guides/pricing (2026-11-11). M3's row
+    // is the ≤512K rate.
     ModelPrice {
         provider: "minimax",
         model_id: "minimax-m3",
-        input_per_mtok_usd: 0.60,
-        cached_input_per_mtok_usd: 0.12,
-        output_per_mtok_usd: 2.40,
-        context_window: 1_000_000,
-    },
-    ModelPrice {
-        provider: "minimax",
-        model_id: "minimax-m2.7",
         input_per_mtok_usd: 0.30,
         cached_input_per_mtok_usd: 0.06,
         output_per_mtok_usd: 1.20,
-        context_window: 204_800,
+        context_window: 1_000_000,
     },
-    ModelPrice {
-        provider: "minimax",
-        model_id: "minimax-m2.7-highspeed",
+];
+
+/// Published long-context price steps, from the same pages as the rows. Above
+/// `above_input_tokens` input tokens (cached tokens included) the whole call
+/// bills at the step's rates.
+pub const LONG_CONTEXT_TIERS: &[LongContextTier] = &[
+    // OpenAI: "Long context: >272K input tokens".
+    LongContextTier {
+        model_id: "gpt-5.5",
+        above_input_tokens: 272_000,
+        input_per_mtok_usd: 10.00,
+        cached_input_per_mtok_usd: 1.00,
+        output_per_mtok_usd: 45.00,
+    },
+    LongContextTier {
+        model_id: "gpt-5.4",
+        above_input_tokens: 272_000,
+        input_per_mtok_usd: 5.00,
+        cached_input_per_mtok_usd: 0.50,
+        output_per_mtok_usd: 22.50,
+    },
+    // Google: prompts >200K.
+    LongContextTier {
+        model_id: "gemini-2.5-pro",
+        above_input_tokens: 200_000,
+        input_per_mtok_usd: 2.50,
+        cached_input_per_mtok_usd: 0.25,
+        output_per_mtok_usd: 15.00,
+    },
+    // Qwen: the 256K-1M input tier (non-thinking), billed per request.
+    LongContextTier {
+        model_id: "qwen-plus",
+        above_input_tokens: 256_000,
+        input_per_mtok_usd: 1.20,
+        cached_input_per_mtok_usd: 0.12,
+        output_per_mtok_usd: 3.60,
+    },
+    LongContextTier {
+        model_id: "qwen-flash",
+        above_input_tokens: 256_000,
+        input_per_mtok_usd: 0.25,
+        cached_input_per_mtok_usd: 0.025,
+        output_per_mtok_usd: 2.00,
+    },
+    // MiniMax: "> 512k input tokens". The page does not say whether cached
+    // tokens count toward the boundary; they are counted here.
+    LongContextTier {
+        model_id: "minimax-m3",
+        above_input_tokens: 512_000,
         input_per_mtok_usd: 0.60,
-        cached_input_per_mtok_usd: 0.06,
+        cached_input_per_mtok_usd: 0.12,
         output_per_mtok_usd: 2.40,
-        context_window: 204_800,
     },
 ];
diff --git a/crates/tinyagents-registry/src/pricing/mod.rs b/crates/tinyagents-registry/src/pricing/mod.rs
index 05951b347..45a38354a 100644
--- a/crates/tinyagents-registry/src/pricing/mod.rs
+++ b/crates/tinyagents-registry/src/pricing/mod.rs
@@ -1,17 +1,25 @@
 //! Offline published model prices, context windows and decorated-id lookup.
 //!
-//! Values preserve the July 2026 consumer snapshot; estimates remain best effort.
+//! Values were checked against provider pages on 2026-11-11; estimates remain
+//! best effort.
 //! This module performs no I/O. Consumers own billing authority, runtime capability
 //! overlays and product configuration.
 
 /// Month the published values below were last verified. Bump when refreshing.
-pub const PRICING_AS_OF: &str = "2026-07";
+pub const PRICING_AS_OF: &str = "2026-11-11";
 
 mod data;
 mod types;
 
-pub use data::KNOWN_MODEL_PRICING;
-pub use types::ModelPrice;
+pub use data::{KNOWN_MODEL_PRICING, LONG_CONTEXT_TIERS};
+pub use types::{LongContextTier, ModelPrice};
+
+/// The long-context price step for a catalogued model id, if it has one.
+pub fn long_context_tier(model_id: &str) -> Option<&'static LongContextTier> {
+    LONG_CONTEXT_TIERS
+        .iter()
+        .find(|tier| tier.model_id == model_id)
+}
 
 /// Normalise a model string for matching: lower-case, trim, drop a trailing
 /// `:tag` / `@date` decoration and a `[...]` suffix.
@@ -109,15 +117,23 @@ pub fn estimate_cost_usd(
     };
     let cached = cached_input_tokens.min(input_tokens);
     let standard_input = input_tokens.saturating_sub(cached);
-    if p.model_id == "minimax-m3" && input_tokens <= 512_000 {
-        return (standard_input as f64) * 0.30 / 1_000_000.0
-            + (cached as f64) * 0.06 / 1_000_000.0
-            + (output_tokens as f64) * 1.20 / 1_000_000.0;
-    }
+    let (input_rate, cached_rate, output_rate) =
+        match long_context_tier(p.model_id).filter(|tier| input_tokens > tier.above_input_tokens) {
+            Some(tier) => (
+                tier.input_per_mtok_usd,
+                tier.cached_input_per_mtok_usd,
+                tier.output_per_mtok_usd,
+            ),
+            None => (
+                p.input_per_mtok_usd,
+                p.cached_input_per_mtok_usd,
+                p.output_per_mtok_usd,
+            ),
+        };
     let per_tok = |mtok_rate: f64| mtok_rate / 1_000_000.0;
-    (standard_input as f64) * per_tok(p.input_per_mtok_usd)
-        + (cached as f64) * per_tok(p.cached_input_per_mtok_usd)
-        + (output_tokens as f64) * per_tok(p.output_per_mtok_usd)
+    (standard_input as f64) * per_tok(input_rate)
+        + (cached as f64) * per_tok(cached_rate)
+        + (output_tokens as f64) * per_tok(output_rate)
 }
 
 /// Resolve a model's flat list-price projection, including explicit `:free`
@@ -166,17 +182,11 @@ const PROMPT_CACHING_MODELS: &[&str] = &[
     "gemini-2.5-pro",
     "gemini-2.5-flash",
     "gemini-2.5-flash-lite",
-    "deepseek-chat",
-    "deepseek-reasoner",
     "kimi-k2.6",
-    "kimi-k2.5",
-    "qwen3-max",
     "qwen-max",
     "qwen-plus",
     "qwen-flash",
     "minimax-m3",
-    "minimax-m2.7",
-    "minimax-m2.7-highspeed",
 ];
 
 /// Whether the published vision allowlist includes this decorated model id.
@@ -184,17 +194,16 @@ pub fn model_accepts_image_input(model_id: &str) -> bool {
     lookup(model_id).is_some_and(|price| VISION_INPUT_MODELS.contains(&price.model_id))
 }
 
-/// Project one published catalog row as flat metadata (not a tier-aware estimate) into a TinyAgents model-catalog entry.
+/// Project one published catalog row into a TinyAgents model-catalog entry.
 ///
 /// This is intentionally a pricing/window projection only. Consumers still
 /// derives live runtime capability flags (tools, vision, streaming) from the
 /// provider adapter at construction time, because the static cost catalog does
 /// not yet encode those fields authoritatively for every provider/model.
 ///
-/// This preserves the migrated flat metadata projection. For MiniMax M3,
-/// `estimate_cost_usd` applies a lower rate through 512K input tokens, while
-/// this projection uses the row rates and intentionally contains no tiers.
-/// Consumers must not treat these two paths as equivalent call estimates.
+/// A model with a [`LongContextTier`] gets two price tiers (the row's rates up
+/// to the step, the step's above it), so a catalogue estimate and
+/// [`estimate_cost_usd`] agree on every call size.
 pub fn flat_catalog_entry(price: &ModelPrice, source: &str) -> crate::ModelCatalogEntry {
     crate::ModelCatalogEntry {
         provider: price.provider.to_string(),
@@ -212,8 +221,27 @@ pub fn flat_catalog_entry(price: &ModelPrice, source: &str) -> crate::ModelCatal
             cache_creation_input_per_token: None,
             input_audio_per_token: None,
             output_reasoning_per_token: None,
-            // The local price table is flat; context-size tiers are not modeled.
-            tiers: Vec::new(),
+            // The same long-context step `estimate_cost_usd` applies.
+            tiers: long_context_tier(price.model_id)
+                .map(|tier| {
+                    vec![
+                        tinyagents_harness::cost::PriceTier {
+                            up_to_tokens: Some(tier.above_input_tokens),
+                            input: per_token(price.input_per_mtok_usd),
+                            output: per_token(price.output_per_mtok_usd),
+                            cache_read: per_token(price.cached_input_per_mtok_usd),
+                            cache_write: None,
+                        },
+                        tinyagents_harness::cost::PriceTier {
+                            up_to_tokens: None,
+                            input: per_token(tier.input_per_mtok_usd),
+                            output: per_token(tier.output_per_mtok_usd),
+                            cache_read: per_token(tier.cached_input_per_mtok_usd),
+                            cache_write: None,
+                        },
+                    ]
+                })
+                .unwrap_or_default(),
         },
         capabilities: crate::ModelCapabilities {
             prompt_caching: PROMPT_CACHING_MODELS.contains(&price.model_id),
diff --git a/crates/tinyagents-registry/src/pricing/pricing_tests.rs b/crates/tinyagents-registry/src/pricing/pricing_tests.rs
index 91eb70ed2..2c358905e 100644
--- a/crates/tinyagents-registry/src/pricing/pricing_tests.rs
+++ b/crates/tinyagents-registry/src/pricing/pricing_tests.rs
@@ -21,11 +21,8 @@ fn lookup_strips_vendor_prefix_openrouter_style() {
         lookup("anthropic/claude-sonnet-4-6").unwrap().model_id,
         "claude-sonnet-4-6"
     );
-    assert_eq!(
-        lookup("deepseek/deepseek-chat").unwrap().model_id,
-        "deepseek-chat"
-    );
-    assert_eq!(lookup("qwen/qwen3-max").unwrap().model_id, "qwen3-max");
+    assert_eq!(lookup("moonshot/kimi-k2.6").unwrap().model_id, "kimi-k2.6");
+    assert_eq!(lookup("qwen/qwen-plus").unwrap().model_id, "qwen-plus");
 }
 
 #[test]
@@ -182,12 +179,49 @@ fn short_ids_do_not_match_inside_unrelated_tokens() {
 }
 
 #[test]
-fn flat_metadata_projection_is_distinct_from_tier_aware_call_estimates() {
-    let row = lookup("minimax-m3").unwrap();
-    let entry = flat_catalog_entry(row, "consumer");
-    assert_eq!(entry.pricing.input_per_token, Some(0.60 / 1_000_000.0));
+fn catalogue_tiers_and_call_estimates_agree_at_every_step() {
+    use tinyagents_harness::cost::{ModelPricing, estimate_cost};
+    use tinyinference_llm::usage::Usage;
+    for tier in LONG_CONTEXT_TIERS {
+        let row = lookup(tier.model_id).unwrap();
+        let pricing: ModelPricing = flat_catalog_entry(row, "consumer").pricing;
+        assert_eq!(pricing.tiers.len(), 2, "{}", tier.model_id);
+        for input in [1_000, tier.above_input_tokens, tier.above_input_tokens + 1] {
+            let usage = Usage {
+                cache_read_tokens: input / 2,
+                ..Usage::new(input, 1_000)
+            };
+            approx(
+                estimate_cost(&pricing, &usage).total_cost,
+                estimate_cost_usd(tier.model_id, input, 1_000, input / 2),
+            );
+        }
+    }
+}
+
+#[test]
+fn long_context_steps_bill_the_whole_call_at_the_higher_rate() {
+    approx(estimate_cost_usd("gpt-5.5", 272_000, 0, 0), 1.36);
+    approx(estimate_cost_usd("gpt-5.5", 272_001, 0, 0), 2.72001);
+    approx(
+        estimate_cost_usd("gemini-2.5-pro", 200_001, 1_000, 0),
+        0.5150025,
+    );
+    approx(estimate_cost_usd("gpt-5.4-mini", 400_000, 0, 0), 0.30);
+}
+
+#[test]
+fn rows_without_a_step_have_no_catalogue_tiers() {
+    let entry = flat_catalog_entry(lookup("claude-opus-4-8").unwrap(), "consumer");
     assert!(entry.pricing.tiers.is_empty());
-    approx(estimate_cost_usd("minimax-m3", 100_000, 0, 0), 0.03);
+    assert!(long_context_tier("claude-opus-4-8").is_none());
+    for tier in LONG_CONTEXT_TIERS {
+        assert!(
+            lookup(tier.model_id).is_some(),
+            "{} has no row",
+            tier.model_id
+        );
+    }
 }
 
 #[test]
@@ -225,6 +259,74 @@ fn free_model_pricing_is_zero_and_does_not_change_reported_charges() {
 
 #[test]
 fn approximate_cached_rates_do_not_advertise_prompt_caching() {
-    let mistral = flat_catalog_entry(lookup("mistral-large").unwrap(), "consumer");
+    let mistral = flat_catalog_entry(lookup("mistral-small").unwrap(), "consumer");
     assert!(!mistral.capabilities.prompt_caching);
 }
+
+#[test]
+fn gpt5_and_gemini_cache_reads_are_a_tenth_of_input() {
+    // #393: these were filled in as 0.25× input; the published rate is 0.1×.
+    for id in [
+        "gpt-5.5",
+        "gpt-5.4",
+        "gpt-5.4-mini",
+        "gpt-5.4-nano",
+        "gemini-2.5-pro",
+        "gemini-2.5-flash",
+        "gemini-2.5-flash-lite",
+    ] {
+        let row = lookup(id).unwrap();
+        approx(row.cached_input_per_mtok_usd, row.input_per_mtok_usd / 10.0);
+    }
+}
+
+#[test]
+fn retired_models_are_gone_and_renamed_mistral_ids_carry_their_new_models() {
+    for id in [
+        "deepseek-chat",
+        "deepseek-reasoner",
+        "kimi-k2.5",
+        "qwen3-max",
+        "minimax-m2.7",
+        "minimax-m2.7-highspeed",
+        "mistral-medium",
+    ] {
+        assert!(lookup(id).is_none(), "{id} is retired");
+        approx(estimate_cost_usd(id, 1_000, 1_000, 0), 0.0);
+    }
+    for (id, input, cached, output, window) in [
+        ("mistral-large", 0.70, 0.07, 2.10, 1_000_000),
+        ("mistral-small", 0.15, 0.015, 0.60, 256_000),
+        ("ministral-8b", 0.15, 0.015, 0.15, 256_000),
+    ] {
+        let row = lookup(id).unwrap();
+        approx(row.input_per_mtok_usd, input);
+        approx(row.cached_input_per_mtok_usd, cached);
+        approx(row.output_per_mtok_usd, output);
+        assert_eq!(row.context_window, window, "{id}");
+    }
+    assert_eq!(
+        lookup("ministral-8b-2512").unwrap().model_id,
+        "ministral-8b"
+    );
+}
+
+#[test]
+fn mistral_large_4_ids_price_like_mistral_large() {
+    let alias = lookup("mistral-large").unwrap();
+    for id in [
+        "mistral-large-4",
+        "mistral-large-4-0",
+        "mistral/mistral-large-4-0",
+    ] {
+        let row = lookup(id).unwrap_or_else(|| panic!("{id} has no row"));
+        approx(row.input_per_mtok_usd, alias.input_per_mtok_usd);
+        approx(
+            row.cached_input_per_mtok_usd,
+            alias.cached_input_per_mtok_usd,
+        );
+        approx(row.output_per_mtok_usd, alias.output_per_mtok_usd);
+        assert_eq!(row.context_window, alias.context_window, "{id}");
+        assert!(estimate_cost_usd(id, 1_000_000, 0, 0) > 0.0, "{id}");
+    }
+}
diff --git a/crates/tinyagents-registry/src/pricing/types.rs b/crates/tinyagents-registry/src/pricing/types.rs
index b5454ae71..847244ccd 100644
--- a/crates/tinyagents-registry/src/pricing/types.rs
+++ b/crates/tinyagents-registry/src/pricing/types.rs
@@ -21,3 +21,20 @@ pub struct ModelPrice {
     /// widely (128K–1M+); callers budget prompts / trigger compaction off this.
     pub context_window: u32,
 }
+
+/// A model's long-context price step: above `above_input_tokens` input tokens
+/// the whole call bills at these rates instead of its [`ModelPrice`] row's.
+#[derive(Debug, Clone, Copy)]
+pub struct LongContextTier {
+    /// The [`ModelPrice::model_id`] this step belongs to.
+    pub model_id: &'static str,
+    /// Largest input-token count (cached included) still billed at the row's
+    /// rates.
+    pub above_input_tokens: u64,
+    /// USD per million standard input tokens above the step.
+    pub input_per_mtok_usd: f64,
+    /// USD per million cached input tokens above the step.
+    pub cached_input_per_mtok_usd: f64,
+    /// USD per million output tokens above the step.
+    pub output_per_mtok_usd: f64,
+}