diff --git a/lib/ruby_llm/provider.rb b/lib/ruby_llm/provider.rb
index e8402fa2..22ac8e4b 100644
--- a/lib/ruby_llm/provider.rb
+++ b/lib/ruby_llm/provider.rb
@@ -14,7 +14,6 @@ def complete(messages, tools:, temperature:, model:, connection:, params: {}, sc
         normalized_temperature = maybe_normalize_temperature(temperature, model)
 
         payload = Utils.deep_merge(
-          params,
           render_payload(
             messages,
             tools: tools,
@@ -22,9 +21,12 @@ def complete(messages, tools:, temperature:, model:, connection:, params: {}, sc
             model: model,
             stream: block_given?,
             schema: schema
-          )
+          ),
+          params
         )
 
+        payload.compact!
+
         if block_given?
           stream_response connection, payload, &
         else
diff --git a/lib/ruby_llm/utils.rb b/lib/ruby_llm/utils.rb
index df281c92..b8ac18e0 100644
--- a/lib/ruby_llm/utils.rb
+++ b/lib/ruby_llm/utils.rb
@@ -24,12 +24,12 @@ def to_safe_array(item)
       end
     end
 
-    def deep_merge(params, payload)
-      params.merge(payload) do |_key, params_value, payload_value|
-        if params_value.is_a?(Hash) && payload_value.is_a?(Hash)
-          deep_merge(params_value, payload_value)
+    def deep_merge(hash1, hash2)
+      hash1.merge(hash2) do |_key, value1, value2|
+        if value1.is_a?(Hash) && value2.is_a?(Hash)
+          deep_merge(value1, value2)
         else
-          payload_value
+          value2
         end
       end
     end
diff --git a/spec/fixtures/vcr_cassettes/chat_with_params_anthropic_claude-3-5-haiku-20241022_can_override_max_tokens_param_with_a_custom_value.yml b/spec/fixtures/vcr_cassettes/chat_with_params_anthropic_claude-3-5-haiku-20241022_can_override_max_tokens_param_with_a_custom_value.yml
new file mode 100644
index 00000000..b40cdf91
--- /dev/null
+++ b/spec/fixtures/vcr_cassettes/chat_with_params_anthropic_claude-3-5-haiku-20241022_can_override_max_tokens_param_with_a_custom_value.yml
@@ -0,0 +1,81 @@
+---
+http_interactions:
+- request:
+    method: post
+    uri: https://api.anthropic.com/v1/messages
+    body:
+      encoding: UTF-8
+      string: '{"model":"claude-3-5-haiku-20241022","messages":[{"role":"user","content":[{"type":"text","text":"Always
+        answer with \"Once upon a time\""}]}],"temperature":0.7,"stream":false,"max_tokens":2}'
+    headers:
+      User-Agent:
+      - Faraday v2.13.4
+      X-Api-Key:
+      - "<ANTHROPIC_API_KEY>"
+      Anthropic-Version:
+      - '2023-06-01'
+      Content-Type:
+      - application/json
+      Accept-Encoding:
+      - gzip;q=1.0,deflate;q=0.6,identity;q=0.3
+      Accept:
+      - "*/*"
+  response:
+    status:
+      code: 200
+      message: OK
+    headers:
+      Date:
+      - Mon, 28 Jul 2025 16:36:30 GMT
+      Content-Type:
+      - application/json
+      Transfer-Encoding:
+      - chunked
+      Connection:
+      - keep-alive
+      Anthropic-Ratelimit-Input-Tokens-Limit:
+      - '100000'
+      Anthropic-Ratelimit-Input-Tokens-Remaining:
+      - '100000'
+      Anthropic-Ratelimit-Input-Tokens-Reset:
+      - '2025-07-28T16:36:30Z'
+      Anthropic-Ratelimit-Output-Tokens-Limit:
+      - '20000'
+      Anthropic-Ratelimit-Output-Tokens-Remaining:
+      - '20000'
+      Anthropic-Ratelimit-Output-Tokens-Reset:
+      - '2025-07-28T16:36:30Z'
+      Anthropic-Ratelimit-Requests-Limit:
+      - '1000'
+      Anthropic-Ratelimit-Requests-Remaining:
+      - '999'
+      Anthropic-Ratelimit-Requests-Reset:
+      - '2025-07-28T16:36:30Z'
+      Anthropic-Ratelimit-Tokens-Limit:
+      - '120000'
+      Anthropic-Ratelimit-Tokens-Remaining:
+      - '120000'
+      Anthropic-Ratelimit-Tokens-Reset:
+      - '2025-07-28T16:36:30Z'
+      Request-Id:
+      - "<REQUEST_ID>"
+      Strict-Transport-Security:
+      - max-age=31536000; includeSubDomains; preload
+      Anthropic-Organization-Id:
+      - "<ANTHROPIC_ORGANIZATION_ID>"
+      Via:
+      - 1.1 google
+      Cf-Cache-Status:
+      - DYNAMIC
+      X-Robots-Tag:
+      - none
+      Server:
+      - cloudflare
+      Cf-Ray:
+      - "<CF_RAY>"
+    body:
+      encoding: ASCII-8BIT
+      string: '{"id":"msg_01RsFPoB8wheUAgoR2SQ6HkY","type":"message","role":"assistant","model":"claude-3-5-haiku-20241022","content":[{"type":"text","text":"Once
+        upon"}],"stop_reason":"max_tokens","stop_sequence":null,"usage":{"input_tokens":16,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"output_tokens":2,"service_tier":"standard"}}'
+  recorded_at: Mon, 28 Jul 2025 16:36:30 GMT
+recorded_with: VCR 6.3.1
diff --git a/spec/ruby_llm/chat_request_options_spec.rb b/spec/ruby_llm/chat_request_options_spec.rb
index 07a3f87b..efa9e8f0 100644
--- a/spec/ruby_llm/chat_request_options_spec.rb
+++ b/spec/ruby_llm/chat_request_options_spec.rb
@@ -78,6 +78,21 @@
         json_response = JSON.parse('{' + response.content) # rubocop:disable Style/StringConcatenation
         expect(json_response).to eq({ 'result' => 8 })
       end
+
+      it "#{provider}/#{model} can override max_tokens param with a custom value" do # rubocop:disable RSpec/ExampleLength
+        chat = RubyLLM
+               .chat(model: model, provider: provider)
+               .with_params(max_tokens: 2)
+
+        chat.add_message(
+          role: :user,
+          content: 'Always answer with "Once upon a time"'
+        )
+
+        response = chat.complete
+
+        expect(response.content).to eq('Once upon') # Only 2 tokens
+      end
     end
 
     # Providers [:openrouter, :bedrock] supports a {top_k: ...} param to remove low-probability next tokens.