← All changes | includes/class-mxchat-model-catalog.php +98 -8 3.2.20 → 3.2.23 View file →

		@@ -55,11 +55,15 @@
55	55	'label' => __('Google Gemini', 'mxchat'),
56	56	'key_option' => 'gemini_api_key',
57	57	'requires_key_to_load_models' => false,
58	58	'models' => array(
59		- 'gemini-3.5-flash' => array('label' => 'Gemini 3.5 Flash', 'description' => __('Stable — newest Flash generation, recommended default', 'mxchat')),
	59	+ 'gemini-3.8-flash' => array('label' => 'Gemini 3.8 Flash', 'description' => __('Stable — newest Flash generation, recommended default', 'mxchat')),
	60	+ 'gemini-3.7-flash' => array('label' => 'Gemini 3.7 Flash', 'description' => __('Stable — efficiency-first Flash', 'mxchat')),
	61	+ 'gemini-3.6-flash' => array('label' => 'Gemini 3.6 Flash', 'description' => __('Stable — previous Flash generation', 'mxchat')),
	62	+ 'gemini-3.5-flash' => array('label' => 'Gemini 3.5 Flash', 'description' => __('Stable — older Flash, priced above 3.8 Flash on every token', 'mxchat')),
60	63	'gemini-3.1-pro-preview' => array('label' => 'Gemini 3.1 Pro', 'description' => __('Preview — most intelligent, multimodal & agentic', 'mxchat')),
61	64	'gemini-3-flash-preview' => array('label' => 'Gemini 3 Flash', 'description' => __('Preview — balanced speed and scale', 'mxchat')),
	65	+ 'gemini-3.5-flash-lite' => array('label' => 'Gemini 3.5 Flash-Lite', 'description' => __('Stable — cheapest Gemini, no thinking', 'mxchat')),
62	66	'gemini-3.1-flash-lite' => array('label' => 'Gemini 3.1 Flash-Lite', 'description' => __('Stable — cost-efficient, high throughput', 'mxchat')),
63	67	'gemini-3.1-flash-lite-preview' => array('label' => 'Gemini 3.1 Flash-Lite (Preview)', 'description' => __('Preview — latest cost-efficient model', 'mxchat')),
64	68	'gemini-2.5-pro' => array('label' => 'Gemini 2.5 Pro', 'description' => __('Stable — advanced thinking, code, math & long context', 'mxchat')),
65	69	'gemini-2.5-flash' => array('label' => 'Gemini 2.5 Flash', 'description' => __('Stable — strong price-performance with thinking', 'mxchat')),
		@@ -79,9 +83,18 @@
79	83	// 'deprecated_on' (Y-m-d) + 'replacement' are optional per-model
80	84	// fields: pickers label a future date as "retires <date>" and hide
81	85	// the entry once the date passes (a saved value keeps rendering via
82	86	// settings_dropdown_groups($current_model)).
83		- 'gpt-5.6-sol' => array('label' => 'GPT-5.6 Sol', 'description' => __('Recommended — newest OpenAI flagship for reasoning, coding and chat', 'mxchat')),
	87	+ // GPT-6 (announced 2026-09-22; plan 45c22d). Probed live on
	88	+ // 2026-09-27: all three need max_completion_tokens, reject any
	89	+ // temperature but 1 and reject reasoning_effort 'minimal';
	90	+ // Astra also rejects 'none'. Every "modern OpenAI" gate goes
	91	+ // through is_modern_openai() so these ids are not sent legacy
	92	+ // parameters (which 400 on the first message).
	93	+ 'gpt-6-astra' => array('label' => 'GPT-6 Astra', 'description' => __('Recommended — newest OpenAI flagship; strongest reasoning, coding and agentic work', 'mxchat')),
	94	+ 'gpt-6-sol' => array('label' => 'GPT-6 Sol', 'description' => __('Strong reasoning on demanding tasks at mid-tier cost', 'mxchat')),
	95	+ 'gpt-6-luna' => array('label' => 'GPT-6 Luna', 'description' => __('Fastest and cheapest GPT-6 for high-volume, repeatable work', 'mxchat')),
	96	+ 'gpt-5.6-sol' => array('label' => 'GPT-5.6 Sol', 'description' => __('Previous OpenAI flagship for reasoning, coding and chat', 'mxchat')),
84	97	'gpt-5.6-terra' => array('label' => 'GPT-5.6 Terra', 'description' => __('Balanced GPT-5.6 — strong capability at lower cost', 'mxchat')),
85	98	'gpt-5.6-luna' => array('label' => 'GPT-5.6 Luna', 'description' => __('Fastest and cheapest GPT-5.6 for lightweight tasks', 'mxchat')),
86	99	'gpt-5.5' => array('label' => 'GPT-5.5', 'description' => __('Previous flagship — powerful reasoning and coding model', 'mxchat')),
87	100	'gpt-5.4' => array('label' => 'GPT-5.4', 'description' => __('Flagship reasoning and coding model', 'mxchat')),
		@@ -671,9 +684,9 @@
671	684	$map = array(
672	685	'openai' => 'gpt-5.6-sol',
673	686	'claude' => 'claude-haiku-4-5-20251001',
674	687	'xai' => 'grok-4.6',
675		- 'gemini' => 'gemini-3.5-flash',
	688	+ 'gemini' => 'gemini-3.8-flash',
676	689	);
677	690	$model = isset($map[$provider]) ? $map[$provider] : '';
678	691	return (string) apply_filters('mxchat_default_vision_model', $model, $provider);
679	692	}
		@@ -718,10 +731,29 @@
718	731	*/
719	732	public static function reasoning_effort_for($model, $context = 'chat') {
720	733	$model = (string) $model;
721	734
722		- // Only the gpt-5 family carries reasoning_effort. Everything else omits.
723		- if (strpos($model, 'gpt-5') !== 0) {
	735	+ // plan a9774b: every Gemini 3.x output price is "including thinking
	736	+ // tokens" and the model thinks at its default level unless told
	737	+ // otherwise — measured on this site's own support prompt, 71% of
	738	+ // billed output on 3.8 Flash was hidden thinking, with indistinguishable
	739	+ // replies at 'low' (0 hidden tokens). So the chat surface sends 'low':
	740	+ // as reasoning_effort on the OpenAI-compatible endpoint, and as
	741	+ // generationConfig.thinkingConfig.thinkingLevel on the native one (see
	742	+ // gemini_thinking_config()). Flash-Lite does not think and returns 400
	743	+ // on any reasoning parameter — it must send nothing.
	744	+ if (strpos($model, 'gemini-3') === 0) {
	745	+ if (strpos($model, 'flash-lite') !== false) {
	746	+ return null;
	747	+ }
	748	+ return ($context === 'chat') ? 'low' : null;
	749	+ }
	750	+
	751	+ // Only the modern OpenAI family (gpt-5*, gpt-6*) carries
	752	+ // reasoning_effort. Everything else omits. plan 45c22d widened this
	753	+ // from a literal gpt-5 prefix: a gpt-6 id used to fall through all
	754	+ // three maps below and be sent legacy parameters.
	755	+ if (!self::is_modern_openai($model)) {
724	756	return null;
725	757	}
726	758
727	759	switch ($context) {
		@@ -728,8 +760,11 @@
728	760	case 'websearch':
729	761	// integrator web-search (Responses API). Explicit allowlist, no
730	762	// "else" — anything not listed omits reasoning entirely.
731	763	$map = array(
	764	+ 'gpt-6-astra' => 'low', // plan 45c22d, probed 2026-09-27 on /v1/responses
	765	+ 'gpt-6-sol' => 'low',
	766	+ 'gpt-6-luna' => 'low',
732	767	'gpt-5.6-sol' => 'low',
733	768	'gpt-5.6-terra' => 'low',
734	769	'gpt-5.6-luna' => 'low',
735	770	'gpt-5.5' => 'low',
		@@ -748,8 +783,14 @@
748	783	// 400'd on every call); the original gpt-5/-mini/-nano still
749	784	// REQUIRE 'minimal' and reject 'none'. Do not "harmonize" these
750	785	// — the two generations genuinely differ.
751	786	$map = array(
	787	+ // plan 45c22d (probed 2026-09-27): gpt-6-astra accepts only
	788	+ // low|medium|high; gpt-6-sol/-luna accept none|low|medium|high.
	789	+ // 'low' for all three, as for their 5.6 siblings.
	790	+ 'gpt-6-astra' => 'low',
	791	+ 'gpt-6-sol' => 'low',
	792	+ 'gpt-6-luna' => 'low',
752	793	'gpt-5.6-sol' => 'low',
753	794	'gpt-5.6-terra' => 'low',
754	795	'gpt-5.6-luna' => 'low',
755	796	'gpt-5.5' => 'low',
		@@ -783,8 +824,14 @@
783	824	}
784	825	if ($model === 'gpt-5.4') {
785	826	return 'none';
786	827	}
	828	+ // plan 45c22d (probed 2026-09-27): gpt-6-astra accepts only
	829	+ // low|medium|high; -sol/-luna accept none|low|medium|high.
	830	+ // 'low' for all three, as for their 5.6 siblings.
	831	+ if (in_array($model, array('gpt-6-astra', 'gpt-6-sol', 'gpt-6-luna'), true)) {
	832	+ return 'low';
	833	+ }
787	834	if (in_array($model, array('gpt-5.6-sol', 'gpt-5.6-terra', 'gpt-5.6-luna'), true)) {
788	835	return 'low';
789	836	}
790	837	// plan-25b972: the original gpt-5 generation REQUIRES 'minimal'
		@@ -815,13 +862,35 @@
815	862	*
816	863	* @param string $model Chat model id.
817	864	* @return bool true if a custom temperature may be sent.
818	865	*/
	866	+ /**
	867	+ * generationConfig.thinkingConfig for the native Gemini endpoint, or null
	868	+ * to send none (plan a9774b). Derived from reasoning_effort_for() so the
	869	+ * two Gemini paths — native generateContent and the OpenAI-compatible
	870	+ * endpoint — always agree on the thinking level. Overridable via
	871	+ * mxchat_gemini_thinking_level (return '' to send none).
	872	+ *
	873	+ * @param string $model Chat model id.
	874	+ * @return array|null e.g. array('thinkingLevel' => 'low').
	875	+ */
	876	+ public static function gemini_thinking_config($model) {
	877	+ $level = self::reasoning_effort_for($model, 'chat');
	878	+ if (strpos((string) $model, 'gemini-') !== 0) {
	879	+ return null;
	880	+ }
	881	+ $level = apply_filters('mxchat_gemini_thinking_level', $level, $model);
	882	+ if (!is_string($level) || $level === '' || !in_array($level, array('minimal', 'low', 'medium', 'high'), true)) {
	883	+ return null;
	884	+ }
	885	+ return array('thinkingLevel' => $level);
	886	+ }
	887	+
819	888	public static function supports_temperature($model) {
820	889	$model = (string) $model;
821	890
822		- // gpt-5 family accepts only the default temperature (1).
823		- if (strpos($model, 'gpt-5') === 0) {
	891	+ // gpt-5 / gpt-6 families accept only the default temperature (1).
	892	+ if (self::is_modern_openai($model)) {
824	893	return false;
825	894	}
826	895
827	896	// Anthropic flagships that reject the temperature param outright.
		@@ -850,7 +919,28 @@
850	919	* @param string $model Chat model id.
851	920	* @return string 'max_completion_tokens' | 'max_tokens'
852	921	*/
853	922	public static function openai_token_param($model) {
854		- return strpos((string) $model, 'gpt-5') === 0 ? 'max_completion_tokens' : 'max_tokens';
	923	+ return self::is_modern_openai($model) ? 'max_completion_tokens' : 'max_tokens';
	924	+ }
	925	+
	926	+ /**
	927	+ * plan 45c22d: ONE predicate for "a modern OpenAI chat model" — the
	928	+ * generations that take max_completion_tokens, only the default
	929	+ * temperature, and reasoning_effort. Every gate that used to test the
	930	+ * literal prefix gpt-5 routes through here, so the next generation is a
	931	+ * one-line change instead of nine edits. Before this, gpt-6-* ids fell
	932	+ * through all nine and were sent legacy parameters: a 400 on the first
	933	+ * message while looking fine in the picker.
	934	+ *
	935	+ * The frozen partial-upgrade fallbacks in the integrator, the content
	936	+ * generator and the editor assistant carry the same two-prefix test
	937	+ * inline (they exist for the window where this class is not loaded).
	938	+ *
	939	+ * @param string $model Chat model id.
	940	+ * @return bool
	941	+ */
	942	+ public static function is_modern_openai($model) {
	943	+ $model = (string) $model;
	944	+ return strpos($model, 'gpt-5') === 0 || strpos($model, 'gpt-6') === 0;
855	945	}
856	946	}