Add gptel-reasoning-effort and add it to the transient menu

gptel-openai-responses.el

@@ -228,6 +228,10 @@ Mutate state INFO with response metadata."
     ;; Temperature
     (when (and gptel-temperature (not o-model-p))
       (plist-put prompts-plist :temperature gptel-temperature))
+    ;; Reasoning effort
+    (when gptel-reasoning-effort
+      (plist-put prompts-plist :reasoning
+                 (list :effort (symbol-name gptel-reasoning-effort))))
     ;; Max tokens
     (when gptel-max-tokens
       (plist-put prompts-plist :max_output_tokens gptel-max-tokens))

gptel-openai.el

@@ -476,6 +476,7 @@ Media files, if present, are placed in `gptel-context'."
   '((gpt-5.4-mini
      :description "Faster, more cost-efficient version of GPT-5.4"
      :capabilities (media tool-use json url responses-api)
+     :reasoning-effort (member none low medium high xhigh)
      :mime-types ("image/jpeg" "image/png" "image/gif" "image/webp")
      :context-window 400
      :input-cost 0.75
@@ -484,6 +485,7 @@ Media files, if present, are placed in `gptel-context'."
     (gpt-5.4-nano
      :description "Fastest, cheapest version of GPT-5.4"
      :capabilities (media tool-use json url responses-api)
+     :reasoning-effort (member none low medium high xhigh)
      :mime-types ("image/jpeg" "image/png" "image/gif" "image/webp")
      :context-window 400
      :input-cost 0.20
@@ -492,6 +494,7 @@ Media files, if present, are placed in `gptel-context'."
     (gpt-5.4
      :description "The best model for coding and agentic tasks"
      :capabilities (media tool-use json url responses-api)
+     :reasoning-effort (member none low medium high xhigh)
      :mime-types ("image/jpeg" "image/png" "image/gif" "image/webp")
      :context-window 1050
      :input-cost 2.50
@@ -500,6 +503,7 @@ Media files, if present, are placed in `gptel-context'."
     (gpt-5.4-pro
      :description "Maximum performance model for reasoning tasks"
      :capabilities (media tool-use json url responses-api)
+     :reasoning-effort (member none low medium high xhigh)
      :mime-types ("image/jpeg" "image/png" "image/gif" "image/webp")
      :context-window 1050
      :input-cost 30
@@ -508,6 +512,7 @@ Media files, if present, are placed in `gptel-context'."
     (gpt-5.3-chat-latest
      :description "Answers right away"
      :capabilities (media tool-use json url responses-api)
+     :reasoning-effort (member medium)
      :mime-types ("image/jpeg" "image/png" "image/gif" "image/webp")
      :context-window 400
      :input-cost 1.75
@@ -516,6 +521,7 @@ Media files, if present, are placed in `gptel-context'."
     (gpt-5.2
      :description "The best model for coding and agentic tasks"
      :capabilities (media tool-use json url responses-api)
+     :reasoning-effort (member none low medium high)
      :mime-types ("image/jpeg" "image/png" "image/gif" "image/webp")
      :context-window 400
      :input-cost 1.75
@@ -524,6 +530,7 @@ Media files, if present, are placed in `gptel-context'."
     (gpt-5.1
      :description "The best model for coding and agentic tasks"
      :capabilities (media tool-use json url responses-api)
+     :reasoning-effort (member none low medium high)
      :mime-types ("image/jpeg" "image/png" "image/gif" "image/webp")
      :context-window 400
      :input-cost 1.25
@@ -532,6 +539,7 @@ Media files, if present, are placed in `gptel-context'."
     (gpt-5-mini
      :description "Faster, more cost-efficient version of GPT-5"
      :capabilities (media tool-use json url responses-api)
+     :reasoning-effort (member minimal low medium high)
      :mime-types ("image/jpeg" "image/png" "image/gif" "image/webp")
      :context-window 400
      :input-cost 0.25
@@ -540,6 +548,7 @@ Media files, if present, are placed in `gptel-context'."
     (gpt-5-nano
      :description "Fastest, cheapest version of GPT-5"
      :capabilities (media tool-use json url responses-api)
+     :reasoning-effort (member minimal low medium high)
      :mime-types ("image/jpeg" "image/png" "image/gif" "image/webp")
      :context-window 400
      :input-cost 0.05
@@ -548,6 +557,7 @@ Media files, if present, are placed in `gptel-context'."
     (gpt-5
      :description "Flagship model for coding, reasoning, and agentic tasks across domains"
      :capabilities (media tool-use json url responses-api)
+     :reasoning-effort (member minimal low medium high)
      :mime-types ("image/jpeg" "image/png" "image/gif" "image/webp")
      :context-window 400
      :input-cost 1.25
@@ -619,13 +629,15 @@ Media files, if present, are placed in `gptel-context'."
     (o4-mini
      :description "Fast, effective reasoning with efficient performance in coding and visual tasks"
      :capabilities (reasoning media tool-use json url responses-api)
+     :reasoning-effort (member low medium high)
      :mime-types ("image/jpeg" "image/png" "image/gif" "image/webp")
      :context-window 200
      :input-cost 1.10
      :output-cost 4.40
      :cutoff-date "2024-05")
     (o3-mini
      :description "High intelligence at the same cost and latency targets of o1-mini"
+     :reasoning-effort (member low medium high)
      :context-window 200
      :input-cost 1.10
      :output-cost 4.40
@@ -634,6 +646,7 @@ Media files, if present, are placed in `gptel-context'."
     (o3
      :description "Well-rounded and powerful model across domains"
      :capabilities (reasoning media tool-use json url responses-api)
+     :reasoning-effort (member low medium high)
      :mime-types ("image/jpeg" "image/png" "image/gif" "image/webp")
      :context-window 200
      :input-cost 2
@@ -642,13 +655,15 @@ Media files, if present, are placed in `gptel-context'."
     (o3-pro
      :description "Maximum performance model for reasoning tasks"
      :capabilities (reasoning media tool-use json url responses-api)
+     :reasoning-effort (member low medium high)
      :mime-types ("image/jpeg" "image/png" "image/gif" "image/webp")
      :context-window 200
      :input-cost 20
      :output-cost 80
      :cutoff-date "2024-05")
     (o1-mini
      :description "Faster and cheaper reasoning model good at coding, math, and science"
+     :reasoning-effort (member low medium high)
      :context-window 128
      :input-cost 1.10
      :output-cost 4.40
@@ -657,6 +672,7 @@ Media files, if present, are placed in `gptel-context'."
     (o1
      :description "Reasoning model designed to solve hard problems across domains"
      :capabilities (media reasoning responses-api)
+     :reasoning-effort (member low medium high)
      :mime-types ("image/jpeg" "image/png" "image/gif" "image/webp")
      :context-window 200
      :input-cost 15

gptel-org.el

@@ -37,6 +37,7 @@
 (defvar gptel--system-message)
 (defvar gptel-model)
 (defvar gptel-temperature)
+(defvar gptel-reasoning-effort)
 (defvar gptel-max-tokens)
 (defvar gptel--link-type-cache)
 (defvar gptel--preset)
@@ -501,12 +502,13 @@ parameters.
 ARGS are the original function call arguments."
   (if (derived-mode-p 'org-mode)
       (pcase-let ((`( ,gptel--preset ,gptel--system-message ,gptel-backend
-                      ,gptel-model ,gptel-temperature ,gptel-max-tokens
-                      ,gptel--num-messages-to-send ,gptel-tools)
+                      ,gptel-model ,gptel-temperature ,gptel-reasoning-effort
+                      ,gptel-max-tokens ,gptel--num-messages-to-send ,gptel-tools)
                    (seq-mapn (lambda (a b) (or a b))
                              (gptel-org--entry-properties)
                              (list gptel--preset gptel--system-message gptel-backend
-                                   gptel-model gptel-temperature gptel-max-tokens
+                                   gptel-model gptel-temperature
+                                   gptel-reasoning-effort gptel-max-tokens
                                    gptel--num-messages-to-send gptel-tools))))
         (apply send-fun args))
     (apply send-fun args)))
@@ -528,8 +530,8 @@ ARGS are the original function call arguments."
          (mapcar
           (lambda (prop) (org-entry-get (or pt (point)) prop 'selective))
           '("GPTEL_PRESET" "GPTEL_SYSTEM" "GPTEL_BACKEND"
-            "GPTEL_MODEL" "GPTEL_TEMPERATURE" "GPTEL_MAX_TOKENS"
-            "GPTEL_NUM_MESSAGES_TO_SEND" "GPTEL_TOOLS"))))
+            "GPTEL_MODEL" "GPTEL_TEMPERATURE" "GPTEL_REASONING_EFFORT"
+            "GPTEL_MAX_TOKENS" "GPTEL_NUM_MESSAGES_TO_SEND" "GPTEL_TOOLS"))))
     (when preset (setq preset (gptel--intern preset)))
     (when system
       (setq system (string-replace "\\n" "\n" system)))
@@ -560,7 +562,7 @@ ARGS are the original function call arguments."
         (progn
           (when-let* ((bounds (org-entry-get (point-min) "GPTEL_BOUNDS")))
             (gptel--restore-props (read bounds)))
-          (pcase-let ((`(,preset ,system ,backend ,model ,temperature ,tokens ,num ,tools)
+          (pcase-let ((`(,preset ,system ,backend ,model ,temperature ,effort ,tokens ,num ,tools)
                        (gptel-org--entry-properties (point-min))))
             (when preset
               (if (gptel-get-preset preset)
@@ -582,6 +584,7 @@ ARGS are the original function call arguments."
                backend))
             (when model (setq-local gptel-model model))
             (when temperature (setq-local gptel-temperature temperature))
+            (when effort (setq-local gptel-reasoning-effort effort))
             (when tokens (setq-local gptel-max-tokens tokens))
             (when num (setq-local gptel--num-messages-to-send num))
             (when tools (setq-local gptel-tools tools))))
@@ -630,11 +633,17 @@ send in queries.  (See `gptel--num-messages-to-send' for the last one.)"
       (if (gptel--preset-mismatch-value preset-spec :tools tool-names)
           (org-entry-put pt "GPTEL_TOOLS" (string-join tool-names " "))
         (org-entry-delete pt "GPTEL_TOOLS")))
-    ;; Temperature, max tokens and cutoff
+    ;; Temperature, reasoning effort, max tokens and cutoff
     (if (and (gptel--preset-mismatch-value preset-spec :temperature gptel-temperature)
              (not (equal (default-value 'gptel-temperature) gptel-temperature)))
         (org-entry-put pt "GPTEL_TEMPERATURE" (number-to-string gptel-temperature))
       (org-entry-delete pt "GPTEL_TEMPERATURE"))
+    (if (and (gptel--preset-mismatch-value preset-spec :reasoning-effort
+                                           gptel-reasoning-effort)
+             (not (equal (default-value 'gptel-reasoning-effort)
+                         gptel-reasoning-effort)))
+        (org-entry-put pt "GPTEL_REASONING_EFFORT" (prin1-to-string gptel-reasoning-effort))
+      (org-entry-delete pt "GPTEL_REASONING_EFFORT"))
     (if (and (gptel--preset-mismatch-value preset-spec :max-tokens gptel-max-tokens)
              gptel-max-tokens)
         (org-entry-put pt "GPTEL_MAX_TOKENS" (number-to-string gptel-max-tokens))

gptel-request.el

@@ -288,6 +288,24 @@ To set the temperature for a chat session interactively call
   :type '(choice (number :tag "Temperature value")
                  (const :tag "Use default" nil)))
 
+(defcustom gptel-reasoning-effort nil
+  "Reasoning effort of the LLM response.
+
+This controls how hard the LLM will \"think\" before generating
+the final response. Not all models support reasoning effort. When
+this value is nil, the model's default reasoning effort will be
+used. The valid values vary depending on the model and the LLM
+provider.
+
+Symbols as well as non-negative integers are supported but may
+generate errors if the specified value isn't allowed."
+  :safe (lambda (v) (or (null v)
+                        (symbolp v)
+                        (and (integerp v) (>= v 0))))
+  :type '(choice (symbol :tag "Reasoning effort level")
+                 (integer :tag "Max thinking tokens")
+                 (const :tag "Use default" nil)))
+
 (defcustom gptel-cache nil
   "Whether the LLM should cache request content.
 
@@ -1020,7 +1038,8 @@ For BUF, START, END and BODY-THUNK see `gptel--with-buffer-copy'."
                       gptel-use-tools gptel-tools gptel-use-curl gptel--schema
                       gptel-use-context gptel-context gptel--num-messages-to-send
                       gptel-stream gptel-include-reasoning gptel--request-params
-                      gptel-temperature gptel-max-tokens gptel-cache))
+                      gptel-temperature gptel-reasoning-effort gptel-max-tokens
+                      gptel-cache))
         (set (make-local-variable sym) (buffer-local-value sym buf)))
       (when (and start end) (insert-buffer-substring buf start end))
       (setq major-mode (buffer-local-value 'major-mode buf))

gptel-transient.el

@@ -850,6 +850,7 @@ Also format the value of OBJ in the transient menu."
      :if (lambda () (and gptel-expert-commands
                     (or gptel-mode gptel-track-response))))
     (gptel--infix-temperature :if (lambda () gptel-expert-commands))
+    (gptel--infix-reasoning-effort :if (lambda () gptel-expert-commands))
     (gptel--infix-use-context)
     (gptel--infix-include-reasoning)
     (gptel--infix-use-tools)
@@ -1333,6 +1334,60 @@ responses."
   :prompt "Temperature controls the response randomness (0.0-2.0, leave empty for API default): "
   :reader 'gptel--transient-read-number)
 
+(defun gptel--transient-read-reasoning-effort (prompt _initial-input history)
+  "Read the reasoning effort from the minibuffer.
+
+PROMPT, _INITIAL-INPUT and HISTORY are as in the transient reader
+documention.  Return nil if user does not provide a number, for default."
+  ;; Workaround for buggy transient behaviour when dealing with
+  ;; non-string values.  See: https://github.com/magit/transient/issues/172
+  (when-let* ((history-symbol (or (car-safe history) history))
+              (val (and (symbolp history-symbol) (symbol-value history-symbol))))
+    (unless (stringp (car val))
+      (setcar val (prin1-to-string (car val)))))
+  (if-let* ((effort-type (get gptel-model :reasoning-effort)))
+      (cond
+       ((eq (car effort-type) 'member)
+        (let* ((table (let ((effort-choices (cons 'default (cdr effort-type))))
+                        ;; Display the completion candidates in the order listed
+                        ;; instead of allowing the completion framework to sort
+                        ;; them. This is cleaner since they are listed in
+                        ;; increasing order of reasoning effort.
+                        (lambda (string predicate action)
+                          (if (eq action 'metadata)
+                              (let ((current-metadata (cdr (completion-metadata
+                                                            (minibuffer-contents)
+                                                            effort-choices
+                                                            minibuffer-completion-predicate))))
+                                `(metadata
+                                  ,@(map-merge 'alist
+                                               current-metadata
+                                               '((display-sort-function . identity)
+                                                 (cycle-sort-function . identity)))))
+                            (complete-with-action action effort-choices string predicate)))))
+               (effort (completing-read prompt table nil t)))
+          ;; Allow the user to restore the value to nil.
+          (unless (string= effort "default")
+            (intern effort))))
+       ((eq (car effort-type) 'integer)
+        (let* ((minibuffer-default-prompt-format "")
+               (num (read-number prompt -1 history)))
+          (if (= num -1) nil num)))
+       (t
+        (user-error "Unknown reasoning effort type: %S" effort-type)))
+    (user-error "Reasoning effort is not supported for this model")))
+
+(transient-define-infix gptel--infix-reasoning-effort ()
+  "Reasoning effort of request."
+  :description "Reasoning effort"
+  :display-nil "default"
+  :class 'gptel-lisp-variable
+  :variable 'gptel-reasoning-effort
+  :set-value #'gptel--set-with-scope
+  :key "-r"
+  :prompt "Reasoning effort controls how hard the LLM \"thinks\" (leave empty for API default): "
+  :reader 'gptel--transient-read-reasoning-effort)
+
 (transient-define-infix gptel--infix-track-response ()
   "Distinguish between user messages and LLM responses.