diff --git a/esphome/components/http_request/http_request.h b/esphome/components/http_request/http_request.h
index c88360ca78..a427cc4a05 100644
--- a/esphome/components/http_request/http_request.h
+++ b/esphome/components/http_request/http_request.h
@@ -103,6 +103,42 @@ inline bool is_success(int const status) { return status >= HTTP_STATUS_OK && st
  *   - ESP-IDF: blocking reads, 0 only returned when all content read
  *   - Arduino: non-blocking, 0 means "no data yet" or "all content read"
  *
+ * Chunked responses that complete in a reasonable time work correctly on both
+ * platforms. The limitation below applies only to *streaming* chunked
+ * responses where data arrives slowly over a long period.
+ *
+ * Streaming chunked responses are NOT supported (all platforms):
+ *   The read helpers (http_read_loop_result, http_read_fully) block the main
+ *   event loop until all response data is received. For streaming responses
+ *   where data trickles in slowly (e.g., TTS streaming via ffmpeg proxy),
+ *   this starves the event loop on both ESP-IDF and Arduino. If data arrives
+ *   just often enough to avoid the caller's timeout, the loop runs
+ *   indefinitely. If data stops entirely, ESP-IDF fails with
+ *   -ESP_ERR_HTTP_EAGAIN (transport timeout) while Arduino spins with
+ *   delay(1) until the caller's timeout fires. Supporting streaming requires
+ *   a non-blocking incremental read pattern that yields back to the event
+ *   loop between chunks. Components that need streaming should use
+ *   esp_http_client directly on a separate FreeRTOS task with
+ *   esp_http_client_is_complete_data_received() for completion detection
+ *   (see audio_reader.cpp for an example).
+ *
+ * Chunked transfer encoding - platform differences:
+ *   - ESP-IDF HttpContainer:
+ *       HttpContainerIDF overrides is_read_complete() to call
+ *       esp_http_client_is_complete_data_received(), which is the
+ *       authoritative completion check for both chunked and non-chunked
+ *       transfers. When esp_http_client_read() returns 0 for a completed
+ *       chunked response, read() returns 0 and is_read_complete() returns
+ *       true, so callers get COMPLETE from http_read_loop_result().
+ *
+ *   - Arduino HttpContainer:
+ *       Chunked responses are decoded internally (see
+ *       HttpContainerArduino::read_chunked_()). When the final chunk arrives,
+ *       is_chunked_ is cleared and content_length is set to bytes_read_.
+ *       Completion is then detected via is_read_complete(), and a subsequent
+ *       read() returns 0 to indicate "all content read" (not
+ *       HTTP_ERROR_CONNECTION_CLOSED).
+ *
  * Use the helper functions below instead of checking return values directly:
  *   - http_read_loop_result(): for manual loops with per-chunk processing
  *   - http_read_fully(): for simple "read N bytes into buffer" operations
@@ -204,9 +240,13 @@ class HttpContainer : public Parented<HttpRequestComponent> {
 
   size_t get_bytes_read() const { return this->bytes_read_; }
 
-  /// Check if all expected content has been read
-  /// For chunked responses, returns false (completion detected via read() returning error/EOF)
-  bool is_read_complete() const {
+  /// Check if all expected content has been read.
+  /// Base implementation handles non-chunked responses and status-code-based no-body checks.
+  /// Platform implementations may override for chunked completion detection:
+  ///   - ESP-IDF: overrides to call esp_http_client_is_complete_data_received() for chunked.
+  ///   - Arduino: read_chunked_() clears is_chunked_ and sets content_length on the final
+  ///     chunk, after which the base implementation detects completion.
+  virtual bool is_read_complete() const {
     // Per RFC 9112, these responses have no body:
     // - 1xx (Informational), 204 No Content, 205 Reset Content, 304 Not Modified
     if ((this->status_code >= 100 && this->status_code < 200) || this->status_code == HTTP_STATUS_NO_CONTENT ||
diff --git a/esphome/components/http_request/http_request_idf.cpp b/esphome/components/http_request/http_request_idf.cpp
index bd12b7d123..486984a694 100644
--- a/esphome/components/http_request/http_request_idf.cpp
+++ b/esphome/components/http_request/http_request_idf.cpp
@@ -218,32 +218,50 @@ std::shared_ptr<HttpContainer> HttpRequestIDF::perform(const std::string &url, c
   return container;
 }
 
+bool HttpContainerIDF::is_read_complete() const {
+  // Base class handles no-body status codes and non-chunked content_length completion
+  if (HttpContainer::is_read_complete()) {
+    return true;
+  }
+  // For chunked responses, use the authoritative ESP-IDF completion check
+  return this->is_chunked_ && esp_http_client_is_complete_data_received(this->client_);
+}
+
 // ESP-IDF HTTP read implementation (blocking mode)
 //
 // WARNING: Return values differ from BSD sockets! See http_request.h for full documentation.
 //
 // esp_http_client_read() in blocking mode returns:
 //   > 0: bytes read
-//   0: connection closed (end of stream)
+//   0: all chunked data received (is_chunk_complete true) or connection closed
+//   -ESP_ERR_HTTP_EAGAIN: transport timeout, no data available yet
 //   < 0: error
 //
 // We normalize to HttpContainer::read() contract:
 //   > 0: bytes read
-//   0: all content read (only returned when content_length is known and fully read)
+//   0: all content read (for both content_length-based and chunked completion)
 //   < 0: error/connection closed
 //
 // Note on chunked transfer encoding:
 //   esp_http_client_fetch_headers() returns 0 for chunked responses (no Content-Length header).
-//   We handle this by skipping the content_length check when content_length is 0,
-//   allowing esp_http_client_read() to handle chunked decoding internally and signal EOF
-//   by returning 0.
+//   When esp_http_client_read() returns 0 for a chunked response, is_read_complete() calls
+//   esp_http_client_is_complete_data_received() to distinguish successful completion from
+//   connection errors. Callers use http_read_loop_result() which checks is_read_complete()
+//   to return COMPLETE for successful chunked EOF.
+//
+// Streaming chunked responses are not supported (see http_request.h for details).
+// When data stops arriving, esp_http_client_read() returns -ESP_ERR_HTTP_EAGAIN
+// after its internal transport timeout (configured via timeout_ms) expires.
+// This is passed through as a negative return value, which callers treat as an error.
 int HttpContainerIDF::read(uint8_t *buf, size_t max_len) {
   const uint32_t start = millis();
   watchdog::WatchdogManager wdm(this->parent_->get_watchdog_timeout());
 
-  // Check if we've already read all expected content (non-chunked only)
-  // For chunked responses (content_length == 0), esp_http_client_read() handles EOF
-  if (this->is_read_complete()) {
+  // Check if we've already read all expected content (non-chunked and no-body only).
+  // Use the base class check here, NOT the override: esp_http_client_is_complete_data_received()
+  // returns true as soon as all data arrives from the network, but data may still be in
+  // the client's internal buffer waiting to be consumed by esp_http_client_read().
+  if (HttpContainer::is_read_complete()) {
     return 0;  // All content read successfully
   }
 
@@ -258,15 +276,18 @@ int HttpContainerIDF::read(uint8_t *buf, size_t max_len) {
     return read_len_or_error;
   }
 
-  // esp_http_client_read() returns 0 in two cases:
-  // 1. Known content_length: connection closed before all data received (error)
-  // 2. Chunked encoding (content_length == 0): end of stream reached (EOF)
-  // For case 1, returning HTTP_ERROR_CONNECTION_CLOSED is correct.
-  // For case 2, 0 indicates that all chunked data has already been delivered
-  // in previous successful read() calls, so treating this as a closed
-  // connection does not cause any loss of response data.
+  // esp_http_client_read() returns 0 when:
+  // - Known content_length: connection closed before all data received (error)
+  // - Chunked encoding: all chunks received (is_chunk_complete true, genuine EOF)
+  //
+  // Return 0 in both cases. Callers use http_read_loop_result() which calls
+  // is_read_complete() to distinguish these:
+  // - Chunked complete: is_read_complete() returns true (via
+  //   esp_http_client_is_complete_data_received()), caller gets COMPLETE
+  // - Non-chunked incomplete: is_read_complete() returns false, caller
+  //   eventually gets TIMEOUT (since no more data arrives)
   if (read_len_or_error == 0) {
-    return HTTP_ERROR_CONNECTION_CLOSED;
+    return 0;
   }
 
   // Negative value - error, return the actual error code for debugging
diff --git a/esphome/components/http_request/http_request_idf.h b/esphome/components/http_request/http_request_idf.h
index ad11811a8f..2a130eae58 100644
--- a/esphome/components/http_request/http_request_idf.h
+++ b/esphome/components/http_request/http_request_idf.h
@@ -16,6 +16,7 @@ class HttpContainerIDF : public HttpContainer {
   HttpContainerIDF(esp_http_client_handle_t client) : client_(client) {}
   int read(uint8_t *buf, size_t max_len) override;
   void end() override;
+  bool is_read_complete() const override;
 
   /// @brief Feeds the watchdog timer if the executing task has one attached
   void feed_wdt();