| @@ -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')), |
| @@ -107,9 +120,8 @@ | ||
| 107 | 120 | 'claude-sonnet-5' => array('label' => 'Claude Sonnet 5', 'description' => __('Newest Sonnet - excellent balance of speed and capability', 'mxchat')), |
| 108 | 121 | 'claude-sonnet-4-6' => array('label' => 'Claude Sonnet 4.6', 'description' => __('Previous Sonnet - excellent balance of speed and capability', 'mxchat')), |
| 109 | 122 | 'claude-opus-4-5' => array('label' => 'Claude Opus 4.5', 'description' => __('Highly capable for complex tasks', 'mxchat')), |
| 110 | 123 | 'claude-sonnet-4-5-20250929' => array('label' => 'Claude Sonnet 4.5', 'description' => __('Best for complex agents and coding', 'mxchat')), |
| 111 | - 'claude-opus-4-1-20250805' => array('label' => 'Claude Opus 4.1', 'description' => __('Exceptional for specialized complex tasks', 'mxchat')), | |
| 112 | 124 | 'claude-haiku-4-5-20251001' => array('label' => 'Claude Haiku 4.5', 'description' => __('Fastest and most intelligent Haiku', 'mxchat')), |
| 113 | 125 | ), |
| 114 | 126 | ), |
| 115 | 127 | 'xai' => array( |
| @@ -115,16 +127,35 @@ | ||
| 115 | 127 | 'xai' => array( |
| 116 | 128 | 'label' => __('xAI Grok', 'mxchat'), |
| 117 | 129 | 'key_option' => 'xai_api_key', |
| 118 | 130 | 'requires_key_to_load_models' => false, |
| 131 | + // Refreshed 2026-08-14 (plan-mxchat-20260814-e1aa4a) against TWO | |
| 132 | + // primary sources that agree: a live pull of xAI's /v1/models with | |
| 133 | + // a real key, and docs.x.ai/docs/models. Context windows below are | |
| 134 | + // that listing's own context_length; image-input support for every | |
| 135 | + // current-generation id was confirmed by a 32x32 probe returning a | |
| 136 | + // correct answer (all HTTP 200), not inferred from docs. | |
| 137 | + // | |
| 138 | + // The legacy block below is INTENTIONALLY still here. xAI has | |
| 139 | + // published no retirement dates for these ids and they still answer | |
| 140 | + // HTTP 200 (grok-4-0709 re-probed 2026-08-14), so 'deprecated_on' | |
| 141 | + // would be an invented date and deleting them would strand every | |
| 142 | + // install saved on one. Current generation simply sorts above them. | |
| 119 | 143 | 'models' => array( |
| 120 | - 'grok-4-1-fast-reasoning' => array('label' => 'Grok 4.1 Fast (Reasoning)', 'description' => __('2M context window and reasoning', 'mxchat')), | |
| 121 | - 'grok-4-1-fast-non-reasoning' => array('label' => 'Grok 4.1 Fast (Non-Reasoning)', 'description' => __('2M context window and faster responses', 'mxchat')), | |
| 122 | - 'grok-4-0709' => array('label' => 'Grok 4', 'description' => __('Latest flagship model - unparalleled performance in natural language, math and reasoning', 'mxchat')), | |
| 123 | - 'grok-3-beta' => array('label' => 'Grok-3', 'description' => __('Powerful model with 131K context', 'mxchat')), | |
| 124 | - 'grok-3-fast-beta' => array('label' => 'Grok-3 Fast', 'description' => __('High performance with faster responses', 'mxchat')), | |
| 125 | - 'grok-3-mini-beta' => array('label' => 'Grok-3 Mini', 'description' => __('Affordable model with good performance', 'mxchat')), | |
| 126 | - 'grok-3-mini-fast-beta' => array('label' => 'Grok-3 Mini Fast', 'description' => __('Quick and cost-effective', 'mxchat')), | |
| 144 | + 'grok-4.6' => array('label' => 'Grok 4.6', 'description' => __('Newest xAI flagship — 500K context, accepts image input', 'mxchat')), | |
| 145 | + 'grok-4.5' => array('label' => 'Grok 4.5', 'description' => __('Previous flagship — 500K context, accepts image input', 'mxchat')), | |
| 146 | + 'grok-4.3' => array('label' => 'Grok 4.3', 'description' => __('Largest context in the Grok lineup — 1M tokens, accepts image input', 'mxchat')), | |
| 147 | + 'grok-4.20-0309-reasoning' => array('label' => 'Grok 4.20 (Reasoning)', 'description' => __('1M context with reasoning — dated release, pinned behaviour', 'mxchat')), | |
| 148 | + 'grok-4.20-0309-non-reasoning'=> array('label' => 'Grok 4.20 (Non-Reasoning)', 'description' => __('1M context without reasoning — faster replies at the same price', 'mxchat')), | |
| 149 | + // --- Previous generations. Still served by xAI, no announced | |
| 150 | + // retirement date, kept so saved selections keep working. --- | |
| 151 | + 'grok-4-1-fast-reasoning' => array('label' => 'Grok 4.1 Fast (Reasoning)', 'description' => __('Previous generation — no longer listed by xAI, still served', 'mxchat')), | |
| 152 | + 'grok-4-1-fast-non-reasoning' => array('label' => 'Grok 4.1 Fast (Non-Reasoning)', 'description' => __('Previous generation — no longer listed by xAI, still served', 'mxchat')), | |
| 153 | + 'grok-4-0709' => array('label' => 'Grok 4', 'description' => __('Previous generation — no longer listed by xAI, still served', 'mxchat')), | |
| 154 | + 'grok-3-beta' => array('label' => 'Grok-3', 'description' => __('Older generation — no longer listed by xAI, still served', 'mxchat')), | |
| 155 | + 'grok-3-fast-beta' => array('label' => 'Grok-3 Fast', 'description' => __('Older generation — no longer listed by xAI, still served', 'mxchat')), | |
| 156 | + 'grok-3-mini-beta' => array('label' => 'Grok-3 Mini', 'description' => __('Older generation — no longer listed by xAI, still served', 'mxchat')), | |
| 157 | + 'grok-3-mini-fast-beta' => array('label' => 'Grok-3 Mini Fast', 'description' => __('Older generation — no longer listed by xAI, still served', 'mxchat')), | |
| 127 | 158 | ), |
| 128 | 159 | ), |
| 129 | 160 | 'deepseek' => array( |
| 130 | 161 | 'label' => __('DeepSeek', 'mxchat'), |
| @@ -184,8 +215,64 @@ | ||
| 184 | 215 | ); |
| 185 | 216 | } |
| 186 | 217 | |
| 187 | 218 | /** |
| 219 | + * Canonical retired-model map: every chat model id a provider has retired | |
| 220 | + * (or hard-deprecated) that stored settings may still carry, mapped to its | |
| 221 | + * current replacement (plan-mxchat-20260822-202df5). | |
| 222 | + * | |
| 223 | + * THE ONLY PLACE THIS MAPPING LIVES. mxchat_migrate_deprecated_models() | |
| 224 | + * consumes it for the main chat + content models, and add-ons that store | |
| 225 | + * their own model ids (mxchat-multi-bot's per-bot overrides) consume it for | |
| 226 | + * theirs — a second hardcoded list in any consumer is how a stranded-model | |
| 227 | + * bug recurs, so route every new retirement through this map. | |
| 228 | + * | |
| 229 | + * Claude ids map BY TIER so a rescued site lands on the current generation, | |
| 230 | + * not an older intermediate (plan dc91bd). Never map onto a model that is | |
| 231 | + * itself on a deprecation list (the e46b8f rule). | |
| 232 | + * | |
| 233 | + * Entry shape: retired_id => array( | |
| 234 | + * 'to' => replacement model id (must be current, never itself retired), | |
| 235 | + * 'label' => replacement display name for user-facing notices, | |
| 236 | + * 'reason' => notice clause explaining the retirement; deliberately NOT | |
| 237 | + * translated — migration messages are stored in an option and | |
| 238 | + * rendered later, matching the existing migration notices. | |
| 239 | + * ) | |
| 240 | + */ | |
| 241 | + public static function retired_model_map() { | |
| 242 | + return array( | |
| 243 | + // Anthropic Claude — Opus tier. | |
| 244 | + 'claude-3-opus-20240229' => array('to' => 'claude-opus-5', 'label' => 'Claude Opus 5', 'reason' => 'because Anthropic retired the older Claude model'), // Retired Jan 5, 2026 | |
| 245 | + 'claude-opus-4-20250514' => array('to' => 'claude-opus-5', 'label' => 'Claude Opus 5', 'reason' => 'because Anthropic retired the older Claude model'), // Deprecated | |
| 246 | + 'claude-opus-4-1-20250805' => array('to' => 'claude-opus-5', 'label' => 'Claude Opus 5', 'reason' => 'because Anthropic retired the older Claude model'), // Retired Aug 5, 2026 (first-party API) | |
| 247 | + // Anthropic Claude — Sonnet tier. | |
| 248 | + 'claude-3-5-sonnet-20240620' => array('to' => 'claude-sonnet-5', 'label' => 'Claude Sonnet 5', 'reason' => 'because Anthropic retired the older Claude model'), // Retired Oct 28, 2025 | |
| 249 | + 'claude-3-5-sonnet-20241022' => array('to' => 'claude-sonnet-5', 'label' => 'Claude Sonnet 5', 'reason' => 'because Anthropic retired the older Claude model'), // Retired Oct 28, 2025 | |
| 250 | + 'claude-3-7-sonnet-20250219' => array('to' => 'claude-sonnet-5', 'label' => 'Claude Sonnet 5', 'reason' => 'because Anthropic retired the older Claude model'), // Retired Feb 19, 2026 | |
| 251 | + 'claude-3-sonnet-20240229' => array('to' => 'claude-sonnet-5', 'label' => 'Claude Sonnet 5', 'reason' => 'because Anthropic retired the older Claude model'), // Legacy | |
| 252 | + 'claude-sonnet-4-20250514' => array('to' => 'claude-sonnet-5', 'label' => 'Claude Sonnet 5', 'reason' => 'because Anthropic retired the older Claude model'), // Deprecated | |
| 253 | + // Anthropic Claude — Haiku tier. | |
| 254 | + 'claude-3-haiku-20240307' => array('to' => 'claude-haiku-4-5-20251001', 'label' => 'Claude Haiku 4.5', 'reason' => 'because Anthropic retired the older Claude model'), // Legacy | |
| 255 | + 'claude-3-5-haiku-20241022' => array('to' => 'claude-haiku-4-5-20251001', 'label' => 'Claude Haiku 4.5', 'reason' => 'because Anthropic retired the older Claude model'), // Deprecated | |
| 256 | + // OpenAI — GPT-4 series + GPT-3.5 Turbo (deprecated 2026-02-17). | |
| 257 | + 'gpt-4o' => array('to' => 'gpt-5.6-sol', 'label' => 'GPT-5.6 Sol', 'reason' => 'due to OpenAI deprecating older models'), | |
| 258 | + 'gpt-4.1-2025-04-14' => array('to' => 'gpt-5.6-sol', 'label' => 'GPT-5.6 Sol', 'reason' => 'due to OpenAI deprecating older models'), | |
| 259 | + 'gpt-4-turbo' => array('to' => 'gpt-5.6-sol', 'label' => 'GPT-5.6 Sol', 'reason' => 'due to OpenAI deprecating older models'), | |
| 260 | + 'gpt-4' => array('to' => 'gpt-5.6-sol', 'label' => 'GPT-5.6 Sol', 'reason' => 'due to OpenAI deprecating older models'), | |
| 261 | + 'gpt-3.5-turbo' => array('to' => 'gpt-5.6-sol', 'label' => 'GPT-5.6 Sol', 'reason' => 'due to OpenAI deprecating older models'), | |
| 262 | + // OpenAI — the gpt-5.x-chat-latest aliases retired 2026-08-10 (e46b8f). | |
| 263 | + 'gpt-5.1-chat-latest' => array('to' => 'gpt-5.6-sol', 'label' => 'GPT-5.6 Sol', 'reason' => 'because OpenAI retires the GPT-5.x Chat Latest models on August 10, 2026'), | |
| 264 | + 'gpt-5.3-chat-latest' => array('to' => 'gpt-5.6-sol', 'label' => 'GPT-5.6 Sol', 'reason' => 'because OpenAI retires the GPT-5.x Chat Latest models on August 10, 2026'), | |
| 265 | + // OpenAI — mini tier. | |
| 266 | + 'gpt-4o-mini' => array('to' => 'gpt-5-mini', 'label' => 'GPT-5 Mini', 'reason' => 'due to OpenAI deprecating GPT-4 series models'), | |
| 267 | + 'gpt-4.1-mini' => array('to' => 'gpt-5-mini', 'label' => 'GPT-5 Mini', 'reason' => 'due to OpenAI deprecating GPT-4 series models'), | |
| 268 | + // DeepSeek — removed at the vendor 2026-07-24 (hard cutoff, every request 400s). | |
| 269 | + 'deepseek-chat' => array('to' => 'deepseek-v4-flash', 'label' => 'DeepSeek V4 Flash', 'reason' => 'because DeepSeek retired its older API models on July 24, 2026'), | |
| 270 | + 'deepseek-reasoner' => array('to' => 'deepseek-v4-flash', 'label' => 'DeepSeek V4 Flash', 'reason' => 'because DeepSeek retired its older API models on July 24, 2026'), | |
| 271 | + ); | |
| 272 | + } | |
| 273 | + | |
| 274 | + /** | |
| 188 | 275 | * Returns the display label for the given chat-or-embedding provider slug. |
| 189 | 276 | * Chat providers checked first; embedding-only providers fall through. |
| 190 | 277 | */ |
| 191 | 278 | public static function provider_label($slug) { |
| @@ -209,8 +296,142 @@ | ||
| 209 | 296 | return ''; |
| 210 | 297 | } |
| 211 | 298 | |
| 212 | 299 | /** |
| 300 | + * The provider slug that owns a chat model id, or '' when the catalog | |
| 301 | + * doesn't list it (plan b65e8d). | |
| 302 | + * | |
| 303 | + * NOTE for callers doing retirement work: a model the provider has already | |
| 304 | + * retired is frequently ABSENT from this catalog while still saved on an | |
| 305 | + * install — that is precisely the state the liveness check exists to catch. | |
| 306 | + * Treat '' as "catalog doesn't know", not "not a real model", and fall back | |
| 307 | + * to a family-prefix guess if you need one. | |
| 308 | + */ | |
| 309 | + public static function provider_for_chat_model($model_id) { | |
| 310 | + $model_id = (string) $model_id; | |
| 311 | + foreach (self::chat_models() as $provider_slug => $provider) { | |
| 312 | + if (isset($provider['models'][$model_id])) { | |
| 313 | + return $provider_slug; | |
| 314 | + } | |
| 315 | + } | |
| 316 | + return ''; | |
| 317 | + } | |
| 318 | + | |
| 319 | + /** | |
| 320 | + * URL + auth headers for a provider's public model-listing endpoint | |
| 321 | + * (plan b65e8d — the daily liveness check). | |
| 322 | + * | |
| 323 | + * Only providers that actually publish a listing are mapped. OpenRouter is | |
| 324 | + * deliberately excluded: its 'openrouter' catalog entry is a meta-selector | |
| 325 | + * that resolves to a per-account sub-model at request time, so "is the | |
| 326 | + * configured id listed" is not a meaningful question. DeepSeek and the | |
| 327 | + * custom OpenAI-compatible provider are excluded too — a self-hosted or | |
| 328 | + * proxied endpoint's /models is not authoritative about retirement. | |
| 329 | + * | |
| 330 | + * @param string $provider Provider slug. | |
| 331 | + * @param string $api_key The install's key for that provider. | |
| 332 | + * @return array|null array('url' => string, 'headers' => array) or null. | |
| 333 | + */ | |
| 334 | + public static function models_listing_request($provider, $api_key) { | |
| 335 | + $api_key = (string) $api_key; | |
| 336 | + if ($api_key === '') { | |
| 337 | + return null; | |
| 338 | + } | |
| 339 | + | |
| 340 | + switch ($provider) { | |
| 341 | + case 'openai': | |
| 342 | + return array( | |
| 343 | + 'url' => 'https://api.openai.com/v1/models', | |
| 344 | + 'headers' => array('Authorization' => 'Bearer ' . $api_key), | |
| 345 | + ); | |
| 346 | + case 'claude': | |
| 347 | + return array( | |
| 348 | + 'url' => 'https://api.anthropic.com/v1/models?limit=1000', | |
| 349 | + 'headers' => array( | |
| 350 | + 'x-api-key' => $api_key, | |
| 351 | + 'anthropic-version' => '2023-06-01', | |
| 352 | + ), | |
| 353 | + ); | |
| 354 | + case 'gemini': | |
| 355 | + return array( | |
| 356 | + 'url' => 'https://generativelanguage.googleapis.com/v1beta/models?pageSize=1000', | |
| 357 | + 'headers' => array('x-goog-api-key' => $api_key), | |
| 358 | + ); | |
| 359 | + case 'xai': | |
| 360 | + return array( | |
| 361 | + 'url' => 'https://api.x.ai/v1/models', | |
| 362 | + 'headers' => array('Authorization' => 'Bearer ' . $api_key), | |
| 363 | + ); | |
| 364 | + } | |
| 365 | + | |
| 366 | + return null; | |
| 367 | + } | |
| 368 | + | |
| 369 | + /** | |
| 370 | + * A minimal 1-token chat request used ONLY to confirm whether a model id | |
| 371 | + * still exists (plan b65e8d). Never used for content — the response is | |
| 372 | + * thrown away; only the HTTP status and error code are read. | |
| 373 | + * | |
| 374 | + * WHY THIS EXISTS: a provider's model LISTING is not authoritative about | |
| 375 | + * retirement, in either direction (all probed live 2026-08-13): | |
| 376 | + * - OpenAI KEEPS retired ids listed — gpt-5.3-chat-latest is in | |
| 377 | + * /v1/models and 404s "model_deprecated" on use. | |
| 378 | + * - Anthropic DROPS working aliases — claude-opus-4-5 is absent from | |
| 379 | + * /v1/models and answers HTTP 200 perfectly. | |
| 380 | + * - xAI's listing is partial — grok-4-0709 is absent and answers 200. | |
| 381 | + * So "absent from the listing" is a suspicion, not a verdict. This probe is | |
| 382 | + * the verdict, and it is the only thing allowed to raise a warning. | |
| 383 | + * | |
| 384 | + * @return array|null array('url','headers','body') or null when unsupported. | |
| 385 | + */ | |
| 386 | + public static function inference_probe_request($provider, $api_key, $model) { | |
| 387 | + $api_key = (string) $api_key; | |
| 388 | + $model = (string) $model; | |
| 389 | + if ($api_key === '' || $model === '') { | |
| 390 | + return null; | |
| 391 | + } | |
| 392 | + | |
| 393 | + $msg = array(array('role' => 'user', 'content' => 'hi')); | |
| 394 | + | |
| 395 | + switch ($provider) { | |
| 396 | + case 'openai': | |
| 397 | + return array( | |
| 398 | + 'url' => 'https://api.openai.com/v1/chat/completions', | |
| 399 | + 'headers' => array('Authorization' => 'Bearer ' . $api_key, 'Content-Type' => 'application/json'), | |
| 400 | + 'body' => array('model' => $model, 'messages' => $msg, 'max_completion_tokens' => 1), | |
| 401 | + ); | |
| 402 | + case 'xai': | |
| 403 | + return array( | |
| 404 | + 'url' => 'https://api.x.ai/v1/chat/completions', | |
| 405 | + 'headers' => array('Authorization' => 'Bearer ' . $api_key, 'Content-Type' => 'application/json'), | |
| 406 | + 'body' => array('model' => $model, 'messages' => $msg, 'max_tokens' => 1), | |
| 407 | + ); | |
| 408 | + case 'claude': | |
| 409 | + return array( | |
| 410 | + 'url' => 'https://api.anthropic.com/v1/messages', | |
| 411 | + 'headers' => array( | |
| 412 | + 'x-api-key' => $api_key, | |
| 413 | + 'anthropic-version' => '2023-06-01', | |
| 414 | + 'Content-Type' => 'application/json', | |
| 415 | + ), | |
| 416 | + 'body' => array('model' => $model, 'max_tokens' => 1, 'messages' => $msg), | |
| 417 | + ); | |
| 418 | + case 'gemini': | |
| 419 | + return array( | |
| 420 | + 'url' => 'https://generativelanguage.googleapis.com/v1beta/models/' | |
| 421 | + . rawurlencode($model) . ':generateContent', | |
| 422 | + 'headers' => array('x-goog-api-key' => $api_key, 'Content-Type' => 'application/json'), | |
| 423 | + 'body' => array( | |
| 424 | + 'contents' => array(array('parts' => array(array('text' => 'hi')))), | |
| 425 | + 'generationConfig' => array('maxOutputTokens' => 1), | |
| 426 | + ), | |
| 427 | + ); | |
| 428 | + } | |
| 429 | + | |
| 430 | + return null; | |
| 431 | + } | |
| 432 | + | |
| 433 | + /** | |
| 213 | 434 | * Flat allowlist of every chat model id across all providers. Used by the |
| 214 | 435 | * sanitize() flow in class-mxchat-admin.php and the AJAX autosave validator |
| 215 | 436 | * in admin/class-ajax-handler.php so a new model id only needs editing in |
| 216 | 437 | * the catalog above. |
| @@ -271,13 +492,22 @@ | ||
| 271 | 492 | * Retired models (past 'deprecated_on') are hidden — EXCEPT the install's |
| 272 | 493 | * currently saved model, which is re-injected with a "retired" marker so |
| 273 | 494 | * the settings screen never renders a blank select (plan e46b8f). |
| 274 | 495 | * |
| 275 | - * @param string $current_model Optionally, the install's saved model id. | |
| 496 | + * @param string $current_model Optionally, the install's saved model id. | |
| 497 | + * @param array|null $provider_slugs Optional allowlist of provider slugs | |
| 498 | + * (catalog keys, e.g. 'openai', 'claude'). | |
| 499 | + * null = every provider. Lets a consumer | |
| 500 | + * that can only dispatch to some providers | |
| 501 | + * (the content generator — plan ccb751) | |
| 502 | + * derive its picker from the same catalog. | |
| 276 | 503 | */ |
| 277 | - public static function settings_dropdown_groups($current_model = '') { | |
| 504 | + public static function settings_dropdown_groups($current_model = '', $provider_slugs = null) { | |
| 278 | 505 | $out = array(); |
| 279 | 506 | foreach (self::chat_models() as $provider_slug => $provider) { |
| 507 | + if (is_array($provider_slugs) && !in_array($provider_slug, $provider_slugs, true)) { | |
| 508 | + continue; | |
| 509 | + } | |
| 280 | 510 | $out[$provider['label']] = array(); |
| 281 | 511 | foreach ($provider['models'] as $id => $entry) { |
| 282 | 512 | $label = self::picker_label($entry); |
| 283 | 513 | if ($label === null && $id !== $current_model) { |
| @@ -432,12 +662,15 @@ | ||
| 432 | 662 | * stated replacement for the gpt-5.x-chat-latest aliases retired |
| 433 | 663 | * 2026-08-10 (which this map previously pinned; plan e46b8f). |
| 434 | 664 | * - claude: the current Haiku — image analysis wants fast + cheap; Haiku |
| 435 | 665 | * is vision-capable and the catalog's newest Haiku generation. |
| 436 | - * - xai: the catalog's dated id for what the bare 'grok-4' alias resolves | |
| 437 | - * to today — documented vision-capable. The grok-4-1-fast pair is newer | |
| 438 | - * but its image-input support is unconfirmed; promote it HERE once | |
| 439 | - * confirmed and every asking add-on follows. | |
| 666 | + * - xai: grok-4.6, the current flagship. Promoted 2026-08-14 (plan | |
| 667 | + * e1aa4a) from grok-4-0709, which was two generations behind and no | |
| 668 | + * longer appears in xAI's /v1/models listing. Image input is CONFIRMED, | |
| 669 | + * not assumed: a 32x32 fixture probe on 2026-08-14 returned the correct | |
| 670 | + * colour on grok-4.6 / 4.5 / 4.3 and both 4.20 variants (all HTTP 200). | |
| 671 | + * grok-4.3 is the cheaper vision option (roughly half the image-token | |
| 672 | + * price, 1M context) if cost ever outweighs capability here. | |
| 440 | 673 | * - gemini: current stable Flash — vision-capable, no per-image surcharge. |
| 441 | 674 | * |
| 442 | 675 | * NOT covered: image-GENERATION models (Gemini Imagen / Flash Image etc.). |
| 443 | 676 | * Those are a different capability the catalog does not model; add-ons that |
| @@ -450,10 +683,10 @@ | ||
| 450 | 683 | public static function default_vision_model($provider) { |
| 451 | 684 | $map = array( |
| 452 | 685 | 'openai' => 'gpt-5.6-sol', |
| 453 | 686 | 'claude' => 'claude-haiku-4-5-20251001', |
| 454 | - 'xai' => 'grok-4-0709', | |
| 455 | - 'gemini' => 'gemini-3.5-flash', | |
| 687 | + 'xai' => 'grok-4.6', | |
| 688 | + 'gemini' => 'gemini-3.8-flash', | |
| 456 | 689 | ); |
| 457 | 690 | $model = isset($map[$provider]) ? $map[$provider] : ''; |
| 458 | 691 | return (string) apply_filters('mxchat_default_vision_model', $model, $provider); |
| 459 | 692 | } |
| @@ -471,17 +704,22 @@ | ||
| 471 | 704 | * |
| 472 | 705 | * The mapping is DELIBERATELY per-context because the surfaces genuinely |
| 473 | 706 | * differ today (they are NOT a single clamped ladder): |
| 474 | 707 | * - 'chat' — /v1/chat/completions main chat. gpt-5.5 and gpt-5.4 use |
| 475 | - * 'none'; the *-chat-latest / mini / nano / gpt-5.2 ids | |
| 476 | - * send NO reasoning_effort. | |
| 708 | + * 'none'; the *-chat-latest / 5.4-mini / 5.4-nano / | |
| 709 | + * gpt-5.2 ids send NO reasoning_effort; the original | |
| 710 | + * gpt-5 / gpt-5-mini / gpt-5-nano send 'minimal' (their | |
| 711 | + * generation requires it — 25b972 probe). Unknown future | |
| 712 | + * gpt-5* ids send NOTHING (allowlist fall-through). | |
| 477 | 713 | * - 'content' — content generation (content-generator, AC ai-client, |
| 478 | 714 | * AC content-calendar). gpt-5.5 / gpt-5.4 use 'low'; |
| 479 | - * gpt-5.3-chat-latest / gpt-5.4-mini / gpt-5.4-nano use | |
| 480 | - * 'minimal'; only gpt-5.2 and gpt-5.1-chat-latest send | |
| 481 | - * nothing. This is an ALLOWLIST (5621dd safe-default): | |
| 482 | - * a future gpt-5* id not listed gets NO param rather than | |
| 483 | - * a guessed value that might 400. | |
| 715 | + * gpt-5.4-mini / gpt-5.4-nano use 'none' (they never | |
| 716 | + * supported 'minimal' — 25b972); gpt-5.3-chat-latest and | |
| 717 | + * the original gpt-5 / -mini / -nano use 'minimal'; only | |
| 718 | + * gpt-5.2 and gpt-5.1-chat-latest send nothing. This is | |
| 719 | + * an ALLOWLIST (5621dd safe-default): a future gpt-5* id | |
| 720 | + * not listed gets NO param rather than a guessed value | |
| 721 | + * that might 400. | |
| 484 | 722 | * - 'websearch' — /v1/responses web-search. Only the gpt-5.6 trio, 5.5, |
| 485 | 723 | * 5.4, and 5.1-2025-11-13 send 'low'; every other gpt-5 id |
| 486 | 724 | * (including gpt-5 / mini / nano / chat-latest) sends |
| 487 | 725 | * nothing. |
| @@ -493,10 +731,29 @@ | ||
| 493 | 731 | */ |
| 494 | 732 | public static function reasoning_effort_for($model, $context = 'chat') { |
| 495 | 733 | $model = (string) $model; |
| 496 | 734 | |
| 497 | - // Only the gpt-5 family carries reasoning_effort. Everything else omits. | |
| 498 | - if (strpos($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)) { | |
| 499 | 756 | return null; |
| 500 | 757 | } |
| 501 | 758 | |
| 502 | 759 | switch ($context) { |
| @@ -503,8 +760,11 @@ | ||
| 503 | 760 | case 'websearch': |
| 504 | 761 | // integrator web-search (Responses API). Explicit allowlist, no |
| 505 | 762 | // "else" — anything not listed omits reasoning entirely. |
| 506 | 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', | |
| 507 | 767 | 'gpt-5.6-sol' => 'low', |
| 508 | 768 | 'gpt-5.6-terra' => 'low', |
| 509 | 769 | 'gpt-5.6-luna' => 'low', |
| 510 | 770 | 'gpt-5.5' => 'low', |
| @@ -515,9 +775,22 @@ | ||
| 515 | 775 | |
| 516 | 776 | case 'content': |
| 517 | 777 | // content-generator + AC class-ai-client + AC content-calendar. |
| 518 | 778 | // ALLOWLIST (5621dd): a gpt-5* id absent here sends nothing. |
| 779 | + // | |
| 780 | + // plan-25b972 (probed live 2026-08-13): supported reasoning_effort | |
| 781 | + // VALUES are per-model. gpt-5.4-mini/nano never supported | |
| 782 | + // 'minimal' (their floor is 'none' — the old 'minimal' entries | |
| 783 | + // 400'd on every call); the original gpt-5/-mini/-nano still | |
| 784 | + // REQUIRE 'minimal' and reject 'none'. Do not "harmonize" these | |
| 785 | + // — the two generations genuinely differ. | |
| 519 | 786 | $map = array( |
| 787 | + // plan 45c22d (probed 2026-09-27): gpt-6-astra accepts only | |
| 788 | + // low|medium|high; gpt-6-sol/-luna accept none|low|medium|high. | |
| 789 | + // 'low' for all three, as for their 5.6 siblings. | |
| 790 | + 'gpt-6-astra' => 'low', | |
| 791 | + 'gpt-6-sol' => 'low', | |
| 792 | + 'gpt-6-luna' => 'low', | |
| 520 | 793 | 'gpt-5.6-sol' => 'low', |
| 521 | 794 | 'gpt-5.6-terra' => 'low', |
| 522 | 795 | 'gpt-5.6-luna' => 'low', |
| 523 | 796 | 'gpt-5.5' => 'low', |
| @@ -522,10 +795,10 @@ | ||
| 522 | 795 | 'gpt-5.6-luna' => 'low', |
| 523 | 796 | 'gpt-5.5' => 'low', |
| 524 | 797 | 'gpt-5.4' => 'low', |
| 525 | 798 | 'gpt-5.1-2025-11-13' => 'low', |
| 526 | - 'gpt-5.4-mini' => 'minimal', | |
| 527 | - 'gpt-5.4-nano' => 'minimal', | |
| 799 | + 'gpt-5.4-mini' => 'none', | |
| 800 | + 'gpt-5.4-nano' => 'none', | |
| 528 | 801 | 'gpt-5.3-chat-latest' => 'minimal', |
| 529 | 802 | 'gpt-5' => 'minimal', |
| 530 | 803 | 'gpt-5-mini' => 'minimal', |
| 531 | 804 | 'gpt-5-nano' => 'minimal', |
| @@ -551,12 +824,26 @@ | ||
| 551 | 824 | } |
| 552 | 825 | if ($model === 'gpt-5.4') { |
| 553 | 826 | return 'none'; |
| 554 | 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($model, array('gpt-6-astra', 'gpt-6-sol', 'gpt-6-luna'), true)) { | |
| 832 | + return 'low'; | |
| 833 | + } | |
| 555 | 834 | if (in_array($model, array('gpt-5.6-sol', 'gpt-5.6-terra', 'gpt-5.6-luna'), true)) { |
| 556 | 835 | return 'low'; |
| 557 | 836 | } |
| 558 | - return 'minimal'; | |
| 837 | + // plan-25b972: the original gpt-5 generation REQUIRES 'minimal' | |
| 838 | + // (probed 2026-08-13: 'none' is rejected for these three). | |
| 839 | + if (in_array($model, array('gpt-5', 'gpt-5-mini', 'gpt-5-nano'), true)) { | |
| 840 | + return 'minimal'; | |
| 841 | + } | |
| 842 | + // plan-25b972: allowlist fall-through (5621dd philosophy). An | |
| 843 | + // unknown future gpt-5* id sends NO param — a guessed value is | |
| 844 | + // exactly how the 5.4-mini/nano 400s shipped. | |
| 845 | + return null; | |
| 559 | 846 | } |
| 560 | 847 | } |
| 561 | 848 | |
| 562 | 849 | /** |
| @@ -575,13 +862,35 @@ | ||
| 575 | 862 | * |
| 576 | 863 | * @param string $model Chat model id. |
| 577 | 864 | * @return bool true if a custom temperature may be sent. |
| 578 | 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' => $level); | |
| 886 | + } | |
| 887 | + | |
| 579 | 888 | public static function supports_temperature($model) { |
| 580 | 889 | $model = (string) $model; |
| 581 | 890 | |
| 582 | - // gpt-5 family accepts only the default temperature (1). | |
| 583 | - if (strpos($model, 'gpt-5') === 0) { | |
| 891 | + // gpt-5 / gpt-6 families accept only the default temperature (1). | |
| 892 | + if (self::is_modern_openai($model)) { | |
| 584 | 893 | return false; |
| 585 | 894 | } |
| 586 | 895 | |
| 587 | 896 | // Anthropic flagships that reject the temperature param outright. |
| @@ -610,7 +919,28 @@ | ||
| 610 | 919 | * @param string $model Chat model id. |
| 611 | 920 | * @return string 'max_completion_tokens' | 'max_tokens' |
| 612 | 921 | */ |
| 613 | 922 | public static function openai_token_param($model) { |
| 614 | - 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; | |
| 615 | 945 | } |
| 616 | 946 | } |
