diff --git a/lib/ruby_llm/models.rb b/lib/ruby_llm/models.rb
index 85087addf..8ecf518e9 100644
--- a/lib/ruby_llm/models.rb
+++ b/lib/ruby_llm/models.rb
@@ -69,6 +69,30 @@ def schema_file
         File.expand_path('models_schema.json', __dir__)
       end
 
+      def overrides_file
+        File.expand_path('models_overrides.json', __dir__)
+      end
+
+      # Hand-verified prices keyed "provider:id", re-applied on top of every refresh so a
+      # models.dev regeneration cannot silently replace them (see models_overrides.json).
+      def read_overrides(file = overrides_file)
+        return {} unless File.exist?(file)
+
+        JSON.parse(File.read(file), symbolize_names: true).fetch(:overrides, {}).transform_keys(&:to_s)
+      end
+
+      # Replaces each overridden model's standard text pricing and metadata.cost with the
+      # override's cost. metadata.cost is replaced whole, so a tier models.dev adds for an
+      # overridden id is dropped too: the override is the authority for that id.
+      def apply_overrides(models, overrides = read_overrides)
+        return models if overrides.empty?
+
+        models.map do |model|
+          override = overrides["#{model.provider}:#{model.id}"]
+          override ? apply_override(model, override[:cost]) : model
+        end
+      end
+
       def load_models(file = RubyLLM.config.model_registry_file)
         source = RubyLLM.config.model_registry_source
         if source && file == RubyLLM.config.model_registry_file
@@ -109,6 +133,7 @@ def fetch_merged_models(remote_only: false)
           log_models_dev_fetch(models_dev_fetch)
 
           merged_models = merge_with_existing(existing_models, provider_fetch, models_dev_fetch)
+          merged_models = apply_overrides(merged_models)
           payload[:model_count] = merged_models.size
           merged_models
         end
@@ -218,7 +243,32 @@ def merge_with_existing(existing_models, provider_fetch, models_dev_fetch)
                               existing_models.select { |model| model.metadata[:source] == 'models.dev' }
                             end
 
-        merge_models(provider_models, models_dev_models)
+        keep_manual_models(merge_models(provider_models, models_dev_models), existing_models)
+      end
+
+      # A model with metadata source 'manual' was added by hand. Keep it when neither the
+      # provider listing nor models.dev returns it, instead of dropping it from the registry.
+      # A manual model therefore stays until someone deletes it from models.json by hand.
+      def keep_manual_models(merged_models, existing_models)
+        merged_keys = index_by_key(merged_models)
+        missing = existing_models.select do |model|
+          model.metadata[:source] == 'manual' && !merged_keys.key?("#{model.provider}:#{model.id}")
+        end
+        return merged_models if missing.empty?
+
+        (merged_models + missing).sort_by { |m| [m.provider, m.id] }
+      end
+
+      # text_tokens becomes { standard: ... } only, so no other text tier survives for an
+      # overridden id (a batch tier is dropped too). Model::Pricing keeps only standard and
+      # batch, so a long-context tier can only arrive through metadata.cost (tiers[] or
+      # context_over_200k, which Supernova's Llm::PricingRegistry reads); replacing
+      # metadata.cost whole is what removes it.
+      def apply_override(model, cost)
+        data = model.to_h
+        data[:pricing] = (data[:pricing] || {}).merge(text_tokens: models_dev_pricing(cost)[:text_tokens])
+        data[:metadata] = (data[:metadata] || {}).merge(cost: cost)
+        Model::Info.new(data)
       end
 
       def merge_models(provider_models, models_dev_models)
diff --git a/lib/ruby_llm/models_overrides.json b/lib/ruby_llm/models_overrides.json
new file mode 100644
index 000000000..66f654397
--- /dev/null
+++ b/lib/ruby_llm/models_overrides.json
@@ -0,0 +1,137 @@
+{
+  "description": [
+    "Prices verified by hand against the vendor. rake models:update re-applies every entry here after it merges the provider listings with models.dev (RubyLLM::Models.apply_overrides), so a regeneration cannot replace them.",
+    "Each entry sets the model's pricing.text_tokens to { standard: ... } only and replaces metadata.cost whole, so no long-context tier is kept for these ids in either place.",
+    "rake models:update also keeps every models.json entry with metadata source 'manual' that no provider listing or models.dev returns (RubyLLM::Models.keep_manual_models). Such an entry stays until it is deleted from models.json by hand.",
+    "No long-context tier exists for these models. Anthropic's pricing page (platform.claude.com/docs/en/about-claude/pricing, read 2026-10-14) says: 'Claude 4.6 and later models (except Claude Haiku 5.5) ... include the full 1M token context window at standard pricing.' The AWS Price List has long_ctx SKUs only for Claude Haiku 5.5 and Claude Sonnet 4.",
+    "Edit models.json and this file together. spec/ruby_llm/models_overrides_spec.rb fails when they disagree."
+  ],
+  "overrides": {
+    "bedrock:us.anthropic.claude-fable-5-1": {
+      "source": "AWS Price List API, AmazonBedrockFoundationModels offer, us-west-2, publication 2026-10-12: on-demand regional ('Standard' / 'Regional') SKUs",
+      "cost": {
+        "input": 11,
+        "output": 55,
+        "cache_read": 0.275,
+        "cache_write": 13.75
+      }
+    },
+    "bedrock:us.anthropic.claude-fable-5": {
+      "source": "AWS Price List API, AmazonBedrockFoundationModels offer, us-west-2, publication 2026-10-12: on-demand regional ('Standard' / 'Regional') SKUs",
+      "cost": {
+        "input": 11,
+        "output": 55,
+        "cache_read": 1.1,
+        "cache_write": 13.75
+      }
+    },
+    "bedrock:us.anthropic.claude-sonnet-5-5": {
+      "source": "AWS Price List API, AmazonBedrockFoundationModels offer, us-west-2, publication 2026-10-12: on-demand regional ('Standard' / 'Regional') SKUs",
+      "cost": {
+        "input": 2.2,
+        "output": 11,
+        "cache_read": 0.11,
+        "cache_write": 2.75
+      }
+    },
+    "bedrock:us.anthropic.claude-sonnet-5": {
+      "source": "AWS Price List API, AmazonBedrockFoundationModels offer, us-west-2, publication 2026-10-12: on-demand regional ('Standard' / 'Regional') SKUs",
+      "cost": {
+        "input": 2.2,
+        "output": 11,
+        "cache_read": 0.22,
+        "cache_write": 2.75
+      }
+    },
+    "bedrock:us.anthropic.claude-opus-5-5": {
+      "source": "AWS Price List API, AmazonBedrockFoundationModels offer, us-west-2, publication 2026-10-12: on-demand regional ('Standard' / 'Regional') SKUs",
+      "cost": {
+        "input": 4.4,
+        "output": 22,
+        "cache_read": 0.22,
+        "cache_write": 5.5
+      }
+    },
+    "bedrock:us.anthropic.claude-opus-5": {
+      "source": "AWS Price List API, AmazonBedrockFoundationModels offer, us-west-2, publication 2026-10-12: on-demand regional ('Standard' / 'Regional') SKUs",
+      "cost": {
+        "input": 5.5,
+        "output": 27.5,
+        "cache_read": 0.55,
+        "cache_write": 6.875
+      }
+    },
+    "bedrock:us.anthropic.claude-opus-4-8": {
+      "source": "AWS Price List API, AmazonBedrockFoundationModels offer, us-west-2, publication 2026-10-12: on-demand regional ('Standard' / 'Regional') SKUs",
+      "cost": {
+        "input": 5.5,
+        "output": 27.5,
+        "cache_read": 0.55,
+        "cache_write": 6.875
+      }
+    },
+    "bedrock:us.anthropic.claude-opus-4-7": {
+      "source": "AWS Price List API, AmazonBedrockFoundationModels offer, us-west-2, publication 2026-10-12: on-demand regional ('Standard' / 'Regional') SKUs",
+      "cost": {
+        "input": 5.5,
+        "output": 27.5,
+        "cache_read": 0.55,
+        "cache_write": 6.875
+      }
+    },
+    "bedrock:us.anthropic.claude-opus-4-6-v1": {
+      "source": "AWS Price List API, AmazonBedrockFoundationModels offer, us-west-2, publication 2026-10-12: on-demand regional ('Standard' / 'Regional') SKUs",
+      "cost": {
+        "input": 5.5,
+        "output": 27.5,
+        "cache_read": 0.55,
+        "cache_write": 6.875
+      }
+    },
+    "bedrock:us.anthropic.claude-opus-4-5-20251101-v1:0": {
+      "source": "AWS Price List API, AmazonBedrockFoundationModels offer, us-west-2, publication 2026-10-12: on-demand regional ('Standard' / 'Regional') SKUs",
+      "cost": {
+        "input": 5.5,
+        "output": 27.5,
+        "cache_read": 0.55,
+        "cache_write": 6.875
+      }
+    },
+    "bedrock:us.anthropic.claude-sonnet-4-6": {
+      "source": "AWS Price List API, AmazonBedrockFoundationModels offer, us-west-2, publication 2026-10-12: on-demand regional ('Standard' / 'Regional') SKUs",
+      "cost": {
+        "input": 3.3,
+        "output": 16.5,
+        "cache_read": 0.33,
+        "cache_write": 4.125
+      }
+    },
+    "bedrock:us.anthropic.claude-sonnet-4-5-20250929-v1:0": {
+      "source": "AWS Price List API, AmazonBedrockFoundationModels offer, us-west-2, publication 2026-10-12: on-demand regional ('Standard' / 'Regional') SKUs",
+      "cost": {
+        "input": 3.3,
+        "output": 16.5,
+        "cache_read": 0.33,
+        "cache_write": 4.125
+      }
+    },
+    "bedrock:us.anthropic.claude-haiku-4-5-20251001-v1:0": {
+      "source": "Global rate on purpose: the AWS bill for 2026-10-05 to 2026-10-12 shows us. Haiku 4.5 at 94% of its regional rate (reason unknown), and the global rate (AWS Price List API, AmazonBedrockFoundationModels offer, us-west-2, publication 2026-10-12: on-demand 'Standard, Global' SKUs) is closer to the bill.",
+      "cost": {
+        "input": 1,
+        "output": 5,
+        "cache_read": 0.1,
+        "cache_write": 1.25
+      }
+    },
+    "bedrock:global.anthropic.claude-fable-5-1": {
+      "source": "AWS Price List API, AmazonBedrockFoundationModels offer, us-west-2, publication 2026-10-12: on-demand 'Standard, Global' SKUs",
+      "cost": {
+        "input": 10,
+        "output": 50,
+        "cache_read": 0.25,
+        "cache_write": 12.5
+      }
+    }
+  }
+}
diff --git a/spec/ruby_llm/models_overrides_spec.rb b/spec/ruby_llm/models_overrides_spec.rb
new file mode 100644
index 000000000..23cea2159
--- /dev/null
+++ b/spec/ruby_llm/models_overrides_spec.rb
@@ -0,0 +1,136 @@
+# frozen_string_literal: true
+
+require 'spec_helper'
+
+RSpec.describe RubyLLM::Models do
+  include_context 'with configured RubyLLM'
+
+  overrides = described_class.read_overrides
+
+  def bedrock_entry(id)
+    described_class.read_from_json.find { |m| m.provider == 'bedrock' && m.id == id }
+  end
+
+  def standard_pricing(model)
+    model.pricing.text_tokens.standard.to_h.slice(
+      :input_per_million, :output_per_million,
+      :cache_read_input_per_million, :cache_write_input_per_million
+    )
+  end
+
+  def expected_standard(cost)
+    {
+      input_per_million: cost[:input],
+      output_per_million: cost[:output],
+      cache_read_input_per_million: cost[:cache_read],
+      cache_write_input_per_million: cost[:cache_write]
+    }
+  end
+
+  it 'has overrides to check' do
+    expect(overrides).not_to be_empty
+  end
+
+  # models.json must already hold each override, so the JSON file and a regenerated registry
+  # agree. A hand edit to one file without the other fails here.
+  overrides.each do |key, override|
+    provider, id = key.split(':', 2)
+
+    it "keeps #{key} in models.json at the override price" do
+      model = described_class.read_from_json.find { |m| m.provider == provider && m.id == id }
+
+      expect(model).not_to be_nil
+      expect(model.metadata[:cost]).to eq(override[:cost])
+      expect(standard_pricing(model)).to eq(expected_standard(override[:cost]))
+    end
+  end
+
+  # Anthropic's pricing page (platform.claude.com/docs/en/about-claude/pricing, read
+  # 2026-10-14), "Long context pricing": "Claude 4.6 and later models (except Claude Haiku 5.5)
+  # ... include the full 1M token context window at standard pricing." The AWS Price List
+  # (AmazonBedrockFoundationModels, us-west-2, publication 2026-10-12) has long_ctx SKUs only
+  # for Claude Haiku 5.5 and Claude Sonnet 4. A tier on any of these ids would overcharge every
+  # request over 200k tokens.
+  %w[
+    us.anthropic.claude-opus-4-8 us.anthropic.claude-opus-5 us.anthropic.claude-opus-5-5
+    us.anthropic.claude-sonnet-5 us.anthropic.claude-sonnet-5-5
+    us.anthropic.claude-fable-5 us.anthropic.claude-fable-5-1
+  ].each do |id|
+    it "prices #{id} with no long-context tier" do
+      model = bedrock_entry(id)
+
+      expect(model.context_window).to eq(1_000_000)
+      expect(model.metadata[:cost].keys).to contain_exactly(:input, :output, :cache_read, :cache_write)
+      expect(model.pricing.to_h[:text_tokens].keys).to eq([:standard])
+    end
+  end
+
+  describe '.apply_overrides' do
+    let(:cost) { { input: 4.4, output: 22, cache_read: 0.22, cache_write: 5.5 } }
+    let(:tiered_model) do
+      RubyLLM::Model::Info.new(
+        id: 'us.anthropic.claude-opus-5-5', name: 'Claude Opus 5.5 (US)', provider: 'bedrock',
+        pricing: {
+          text_tokens: {
+            standard: { input_per_million: 4, output_per_million: 20 },
+            context_over_200k: { input_per_million: 8, output_per_million: 30 }
+          }
+        },
+        metadata: {
+          cost: { input: 4, output: 20, context_over_200k: { input: 8, output: 30 },
+                  tiers: [{ input: 8, output: 30, tier: { type: 'context', size: 200_000 } }] }
+        }
+      )
+    end
+
+    # Model::Pricing already drops any text tier but standard and batch, so the tier that
+    # matters is the one in metadata.cost; this pins both places.
+    it 'drops a long-context tier from both pricing and metadata.cost' do
+      model = described_class.apply_overrides([tiered_model],
+                                              { 'bedrock:us.anthropic.claude-opus-5-5' => { cost: cost } })
+                             .first
+
+      expect(model.pricing.to_h[:text_tokens]).to eq(standard: expected_standard(cost))
+      expect(model.metadata[:cost]).to eq(cost)
+    end
+  end
+
+  describe 'refresh' do
+    let(:overridden_id) { 'us.anthropic.claude-opus-5-5' }
+    let(:models_dev_model) do
+      RubyLLM::Model::Info.new(
+        id: overridden_id, name: 'Claude Opus 5.5 (US)', provider: 'bedrock',
+        pricing: { text_tokens: { standard: { input_per_million: 4, output_per_million: 20 } } },
+        metadata: {
+          source: 'models.dev',
+          cost: { input: 4, output: 20, tiers: [{ input: 8, output: 30, tier: { type: 'context', size: 200_000 } }] }
+        }
+      )
+    end
+    let(:manual_model) do
+      RubyLLM::Model::Info.new(id: 'us.anthropic.claude-hand-added', name: 'Hand added', provider: 'bedrock',
+                               metadata: { source: 'manual', cost: { input: 1, output: 2 } })
+    end
+
+    before do
+      allow(described_class).to receive(:read_existing_models).and_return([manual_model])
+      allow(described_class).to receive_messages(
+        fetch_provider_models: { models: [], fetched_providers: ['bedrock'], configured_names: [], failed: [] },
+        fetch_models_dev_models: { models: [models_dev_model], fetched: true }
+      )
+    end
+
+    it 'applies the override over the models.dev price, tiers included' do
+      model = described_class.fetch_merged_models.find { |m| m.id == overridden_id }
+
+      expect(model.metadata[:cost]).to eq(overrides["bedrock:#{overridden_id}"][:cost])
+      expect(standard_pricing(model)).to eq(expected_standard(overrides["bedrock:#{overridden_id}"][:cost]))
+    end
+
+    it 'keeps a manual model that no source returns' do
+      ids = described_class.fetch_merged_models.map(&:id)
+
+      expect(ids).to include('us.anthropic.claude-hand-added')
+    end
+  end
+end