diff --git a/src/models.rs b/src/models.rs
index 7bd37dd..7e27267 100644
--- a/src/models.rs
+++ b/src/models.rs
@@ -1815,6 +1815,7 @@ fn populate_defaults(
     );
 
     // Google Models
+    // Source: https://ai.google.dev/gemini-api/docs/pricing
     add_model!(
         "gemini-3-flash-preview",
         PricingStructure::Flat {
@@ -1903,11 +1904,11 @@ fn populate_defaults(
             tiers: vec![
                 CachingTier {
                     max_tokens: Some(200_000),
-                    cached_input_per_1m: 0.31
+                    cached_input_per_1m: 0.125
                 },
                 CachingTier {
                     max_tokens: None,
-                    cached_input_per_1m: 0.625
+                    cached_input_per_1m: 0.25
                 },
             ],
             bracket_pricing: false,
@@ -1923,7 +1924,7 @@ fn populate_defaults(
         CachingSupport::Tiered(TieredCaching {
             tiers: vec![CachingTier {
                 max_tokens: None,
-                cached_input_per_1m: 0.075
+                cached_input_per_1m: 0.03
             }],
             bracket_pricing: false,
         }),
@@ -1938,7 +1939,7 @@ fn populate_defaults(
         CachingSupport::Tiered(TieredCaching {
             tiers: vec![CachingTier {
                 max_tokens: None,
-                cached_input_per_1m: 0.025
+                cached_input_per_1m: 0.01
             }],
             bracket_pricing: false,
         }),
@@ -2128,6 +2129,134 @@ fn populate_defaults(
 
     // xAI Models
     // Source: https://docs.x.ai/developers/pricing
+    add_model!(
+        "grok-4.3",
+        PricingStructure::Tiered(TieredPricing {
+            tiers: vec![
+                PricingTier {
+                    max_tokens: Some(200_000),
+                    input_per_1m: 1.25,
+                    output_per_1m: 2.50,
+                },
+                PricingTier {
+                    max_tokens: None,
+                    input_per_1m: 2.50,
+                    output_per_1m: 5.00,
+                },
+            ],
+            bracket_pricing: true,
+        }),
+        CachingSupport::Tiered(TieredCaching {
+            tiers: vec![
+                CachingTier {
+                    max_tokens: Some(200_000),
+                    cached_input_per_1m: 0.20,
+                },
+                CachingTier {
+                    max_tokens: None,
+                    cached_input_per_1m: 0.40,
+                },
+            ],
+            bracket_pricing: true,
+        }),
+        false
+    );
+    add_model!(
+        "grok-4.20-0309-reasoning",
+        PricingStructure::Tiered(TieredPricing {
+            tiers: vec![
+                PricingTier {
+                    max_tokens: Some(200_000),
+                    input_per_1m: 1.25,
+                    output_per_1m: 2.50,
+                },
+                PricingTier {
+                    max_tokens: None,
+                    input_per_1m: 2.50,
+                    output_per_1m: 5.00,
+                },
+            ],
+            bracket_pricing: true,
+        }),
+        CachingSupport::Tiered(TieredCaching {
+            tiers: vec![
+                CachingTier {
+                    max_tokens: Some(200_000),
+                    cached_input_per_1m: 0.20,
+                },
+                CachingTier {
+                    max_tokens: None,
+                    cached_input_per_1m: 0.40,
+                },
+            ],
+            bracket_pricing: true,
+        }),
+        false
+    );
+    add_model!(
+        "grok-4.20-0309-non-reasoning",
+        PricingStructure::Tiered(TieredPricing {
+            tiers: vec![
+                PricingTier {
+                    max_tokens: Some(200_000),
+                    input_per_1m: 1.25,
+                    output_per_1m: 2.50,
+                },
+                PricingTier {
+                    max_tokens: None,
+                    input_per_1m: 2.50,
+                    output_per_1m: 5.00,
+                },
+            ],
+            bracket_pricing: true,
+        }),
+        CachingSupport::Tiered(TieredCaching {
+            tiers: vec![
+                CachingTier {
+                    max_tokens: Some(200_000),
+                    cached_input_per_1m: 0.20,
+                },
+                CachingTier {
+                    max_tokens: None,
+                    cached_input_per_1m: 0.40,
+                },
+            ],
+            bracket_pricing: true,
+        }),
+        false
+    );
+    add_model!(
+        "grok-4.20-multi-agent-0309",
+        PricingStructure::Tiered(TieredPricing {
+            tiers: vec![
+                PricingTier {
+                    max_tokens: Some(200_000),
+                    input_per_1m: 1.25,
+                    output_per_1m: 2.50,
+                },
+                PricingTier {
+                    max_tokens: None,
+                    input_per_1m: 2.50,
+                    output_per_1m: 5.00,
+                },
+            ],
+            bracket_pricing: true,
+        }),
+        CachingSupport::Tiered(TieredCaching {
+            tiers: vec![
+                CachingTier {
+                    max_tokens: Some(200_000),
+                    cached_input_per_1m: 0.20,
+                },
+                CachingTier {
+                    max_tokens: None,
+                    cached_input_per_1m: 0.40,
+                },
+            ],
+            bracket_pricing: true,
+        }),
+        false
+    );
     add_model!(
         "grok-4.5",
         PricingStructure::Tiered(TieredPricing {
@@ -2160,6 +2289,38 @@ fn populate_defaults(
         }),
         false
     );
+    add_model!(
+        "grok-4.6",
+        PricingStructure::Tiered(TieredPricing {
+            tiers: vec![
+                PricingTier {
+                    max_tokens: Some(200_000),
+                    input_per_1m: 2.00,
+                    output_per_1m: 6.00,
+                },
+                PricingTier {
+                    max_tokens: None,
+                    input_per_1m: 4.00,
+                    output_per_1m: 12.00,
+                },
+            ],
+            bracket_pricing: true,
+        }),
+        CachingSupport::Tiered(TieredCaching {
+            tiers: vec![
+                CachingTier {
+                    max_tokens: Some(200_000),
+                    cached_input_per_1m: 0.50,
+                },
+                CachingTier {
+                    max_tokens: None,
+                    cached_input_per_1m: 1.00,
+                },
+            ],
+            bracket_pricing: true,
+        }),
+        false
+    );
     add_model!(
         "grok-build-0.1",
         PricingStructure::Tiered(TieredPricing {
@@ -2227,29 +2388,33 @@ fn populate_defaults(
     );
 
     // DeepSeek Models
-    // Source: https://api-docs.deepseek.com/quick_start/pricing/ (official pricing page;
-    // use the standard rates rather than the temporary promotional discount:
-    // cache hit $0.0145/M, cache miss $1.74/M, output $3.48/M)
+    // Source: https://api-docs.deepseek.com/quick_start/pricing/
+    // The page publishes peak and off-peak rates, where off-peak is half of
+    // peak. Splitrail prices usage without a time-of-day dimension, so the
+    // peak (standard) rates are used.
     add_model!(
         "deepseek-v4-pro",
         PricingStructure::Flat {
-            input_per_1m: 1.74,
-            output_per_1m: 3.48
+            input_per_1m: 1.32,
+            output_per_1m: 3.96
         },
         CachingSupport::OpenAI {
-            cached_input_per_1m: 0.0145
+            cached_input_per_1m: 0.044
         },
         false
     );
-    // Source: https://api-docs.deepseek.com/quick_start/pricing/
+    // DeepSeek now asks callers to use `deepseek-flash`; the legacy
+    // `deepseek-v4-flash` name is still accepted and billed at the same rate.
+    // From September 14, 2026 requests to `deepseek-v4-pro` are routed to
+    // DeepSeek-V4.1-Flash and billed at this price.
     add_model!(
         "deepseek-v4-flash",
         PricingStructure::Flat {
-            input_per_1m: 0.14,
-            output_per_1m: 0.28
+            input_per_1m: 0.30,
+            output_per_1m: 1.20
         },
         CachingSupport::OpenAI {
-            cached_input_per_1m: 0.0028
+            cached_input_per_1m: 0.006
         },
         false
     );
@@ -2338,14 +2503,17 @@ fn populate_defaults(
     );
 
     // MiniMax Models
-    // Source: https://platform.minimax.io/docs/api-reference/anthropic-api-compatible-cache
+    // Source: https://platform.minimax.io/docs/guides/pricing-paygo
     add_model!(
         "minimax-m2.1",
         PricingStructure::Flat {
             input_per_1m: 0.30,
             output_per_1m: 1.20
         },
-        CachingSupport::None,
+        CachingSupport::Anthropic {
+            cache_write_per_1m: 0.375,
+            cache_read_per_1m: 0.03
+        },
         false
     );
     add_model!(
@@ -2364,21 +2532,46 @@ fn populate_defaults(
         "minimax-m2.5",
         PricingStructure::Flat {
             input_per_1m: 0.30,
-            output_per_1m: 1.10
+            output_per_1m: 1.20
+        },
+        CachingSupport::Anthropic {
+            cache_write_per_1m: 0.375,
+            cache_read_per_1m: 0.03
         },
-        CachingSupport::None,
         false
     );
-    // Source: https://platform.minimax.io/docs/api-reference/anthropic-api-compatible-cache
+    // MiniMax-M3 is bracketed on input size: at or below 512k tokens the
+    // permanent 50% discount applies, above it the undiscounted rates apply.
     add_model!(
         "minimax-m3",
-        PricingStructure::Flat {
-            input_per_1m: 0.60,
-            output_per_1m: 2.40
-        },
-        CachingSupport::OpenAI {
-            cached_input_per_1m: 0.12
-        },
+        PricingStructure::Tiered(TieredPricing {
+            tiers: vec![
+                PricingTier {
+                    max_tokens: Some(512_000),
+                    input_per_1m: 0.30,
+                    output_per_1m: 1.20,
+                },
+                PricingTier {
+                    max_tokens: None,
+                    input_per_1m: 0.60,
+                    output_per_1m: 2.40,
+                },
+            ],
+            bracket_pricing: true,
+        }),
+        CachingSupport::Tiered(TieredCaching {
+            tiers: vec![
+                CachingTier {
+                    max_tokens: Some(512_000),
+                    cached_input_per_1m: 0.06,
+                },
+                CachingTier {
+                    max_tokens: None,
+                    cached_input_per_1m: 0.12,
+                },
+            ],
+            bracket_pricing: true,
+        }),
         false
     );
 
@@ -2772,6 +2965,9 @@ fn populate_defaults(
     add_alias!("minimax-m2.7", "minimax-m2.7");
     add_alias!("minimax-m3", "minimax-m3");
 
+    // DeepSeek aliases
+    add_alias!("deepseek-flash", "deepseek-v4-flash");
+
     // Moonshot / ByteDance / Qwen / Xiaomi / Meituan aliases
     add_alias!("doubao-seed-code", "doubao-seed-2.0-code");
     add_alias!("kimi-k2.6", "kimi-k2.6");
@@ -3840,6 +4036,35 @@ mod tests {
         approx_eq(cost, 0.375);
     }
 
+    #[test]
+    fn gemini_2_5_cache_reads_match_published_rates() {
+        // ai.google.dev lists context caching at 10% of input: $0.125/$0.25 for
+        // Pro, $0.03 for Flash, and $0.01 for Flash-Lite.
+        approx_eq(calculate_cache_cost("gemini-2.5-pro", 0, 100_000), 0.0125);
+        // Pro applies the tiers progressively rather than bracketing the whole
+        // request: 200k at $0.125 plus 50k at $0.25.
+        approx_eq(calculate_cache_cost("gemini-2.5-pro", 0, 250_000), 0.0375);
+        approx_eq(calculate_cache_cost("gemini-2.5-flash", 0, 1_000_000), 0.03);
+        approx_eq(
+            calculate_cache_cost("gemini-2.5-flash-lite", 0, 1_000_000),
+            0.01,
+        );
+    }
+
+    #[test]
+    fn minimax_m2_5_pricing_matches_published_rates() {
+        let model_info = get_model_info("minimax-m2.5").expect("model should exist");
+        assert!(!model_info.is_estimated);
+
+        approx_eq(calculate_input_cost("minimax-m2.5", 1_000_000), 0.30);
+        approx_eq(calculate_output_cost("minimax-m2.5", 1_000_000), 1.20);
+        // Pay-as-you-go lists cache reads at $0.03 and cache writes at $0.375
+        // for the legacy M2 era, matching the M2.7 entry.
+        approx_eq(calculate_cache_cost("minimax-m2.5", 0, 1_000_000), 0.03);
+        approx_eq(calculate_cache_cost("minimax-m2.5", 1_000_000, 0), 0.375);
+        approx_eq(calculate_cache_cost("minimax-m2.1", 0, 1_000_000), 0.03);
+    }
+
     #[test]
     fn gpt_6_astra_aliases_map_to_official_standard_pricing() {
         assert!(get_model_info("gpt-6-astra-2026-09-03").is_none());
@@ -4517,13 +4742,54 @@ mod tests {
                 .expect("Grok 4.5 should exist")
                 .is_estimated
         );
+        assert!(
+            !get_model_info("grok-4.6")
+                .expect("Grok 4.6 should exist")
+                .is_estimated
+        );
+        for model in [
+            "grok-4.3",
+            "grok-4.20-0309-reasoning",
+            "grok-4.20-0309-non-reasoning",
+            "grok-4.20-multi-agent-0309",
+        ] {
+            assert!(
+                !get_model_info(model)
+                    .unwrap_or_else(|| panic!("{model} should exist"))
+                    .is_estimated
+            );
+        }
 
+        approx_eq(
+            calculate_total_cost_for_context_at(
+                "grok-4.3", 1_000_000, 1_000_000, 0, 1_000_000, 199_999, None,
+            ),
+            3.95,
+        );
+        approx_eq(
+            calculate_total_cost_for_context_at(
+                "grok-4.3", 1_000_000, 1_000_000, 0, 1_000_000, 200_001, None,
+            ),
+            7.9,
+        );
         approx_eq(
             calculate_total_cost_for_context_at(
                 "grok-4.5", 1_000_000, 1_000_000, 0, 1_000_000, 199_999, None,
             ),
             8.3,
         );
+        approx_eq(
+            calculate_total_cost_for_context_at(
+                "grok-4.6", 1_000_000, 1_000_000, 0, 1_000_000, 199_999, None,
+            ),
+            8.5,
+        );
+        approx_eq(
+            calculate_total_cost_for_context_at(
+                "grok-4.6", 1_000_000, 1_000_000, 0, 1_000_000, 200_001, None,
+            ),
+            17.0,
+        );
         approx_eq(
             calculate_total_cost_for_context_at(
                 "grok-code-fast-1",
@@ -4544,6 +4810,14 @@ mod tests {
                 "grok-4.5", 1_000_000, 1_000_000, 0, 1_000_000, 2_000_000, None,
             ),
         );
+        approx_eq(
+            calculate_total_cost_for_context_at(
+                "grok-4.6", 1_000_000, 1_000_000, 999_999, 1_000_000, 2_000_000, None,
+            ),
+            calculate_total_cost_for_context_at(
+                "grok-4.6", 1_000_000, 1_000_000, 0, 1_000_000, 2_000_000, None,
+            ),
+        );
     }
 
     #[test]
@@ -4609,9 +4883,20 @@ mod tests {
         let output_cost = calculate_output_cost("deepseek-v4-pro", 1_000_000);
         let cache_cost = calculate_cache_cost("deepseek-v4-pro", 0, 1_000_000);
 
-        approx_eq(input_cost, 1.74);
-        approx_eq(output_cost, 3.48);
-        approx_eq(cache_cost, 0.0145);
+        approx_eq(input_cost, 1.32);
+        approx_eq(output_cost, 3.96);
+        approx_eq(cache_cost, 0.044);
+    }
+
+    #[test]
+    fn deepseek_flash_alias_matches_legacy_v4_flash_pricing() {
+        // The page asks callers to use `deepseek-flash`; `deepseek-v4-flash`
+        // is the legacy name for the same DeepSeek-V4.1-Flash model.
+        for model in ["deepseek-flash", "deepseek-v4-flash"] {
+            approx_eq(calculate_input_cost(model, 1_000_000), 0.30);
+            approx_eq(calculate_output_cost(model, 1_000_000), 1.20);
+            approx_eq(calculate_cache_cost(model, 0, 1_000_000), 0.006);
+        }
     }
 
     #[test]
@@ -4696,6 +4981,10 @@ mod tests {
         approx_eq(calculate_input_cost("minimax-m3", 1_000_000), 0.60);
         approx_eq(calculate_output_cost("minimax-m3", 1_000_000), 2.40);
         approx_eq(calculate_cache_cost("minimax-m3", 0, 1_000_000), 0.12);
+        // At or below 512k input tokens the permanent 50% discount applies.
+        approx_eq(calculate_input_cost("minimax-m3", 512_000), 0.30 * 0.512);
+        approx_eq(calculate_output_cost("minimax-m3", 512_000), 1.20 * 0.512);
+        approx_eq(calculate_cache_cost("minimax-m3", 0, 512_000), 0.06 * 0.512);
 
         approx_eq(calculate_input_cost("glm-5.1", 1_000_000), 1.40);
         approx_eq(calculate_output_cost("glm-5.1", 1_000_000), 4.40);
@@ -4715,11 +5004,11 @@ mod tests {
         approx_eq(calculate_input_cost("deepseek.v3.2", 1_000_000), 0.62);
         approx_eq(calculate_output_cost("deepseek.v3.2", 1_000_000), 1.85);
 
-        approx_eq(calculate_input_cost("deepseek-v4-flash", 1_000_000), 0.14);
-        approx_eq(calculate_output_cost("deepseek-v4-flash", 1_000_000), 0.28);
+        approx_eq(calculate_input_cost("deepseek-v4-flash", 1_000_000), 0.30);
+        approx_eq(calculate_output_cost("deepseek-v4-flash", 1_000_000), 1.20);
         approx_eq(
             calculate_cache_cost("deepseek-v4-flash", 0, 1_000_000),
-            0.0028,
+            0.006,
         );
 
         approx_eq(