← All changes | includes/ai/class-claude-client.php +104 -6 2.2.0 → 2.7.0 View file →

		@@ -243,19 +243,113 @@
243	243	return $this->model;
244	244	}
245	245
246	246	/**
	247	+ * Output-token ceiling per model family, longest prefix wins.
	248	+ *
	249	+ * Matched by prefix so a dated snapshot (`claude-haiku-4-5-20251001`) and a
	250	+ * point release (`claude-fable-5-1`) resolve to their family. Order matters
	251	+ * only in that lookup walks longest-first, which is what keeps
	252	+ * `claude-fable-5-1` from matching `claude-fable-5`.
	253	+ *
	254	+ * @since 2.7.0
	255	+ * @var array<string, int>
	256	+ */
	257	+ private const MODEL_OUTPUT_LIMITS = [
	258	+ // 128K output.
	259	+ 'claude-fable-5-1' => 128000,
	260	+ 'claude-fable-5' => 128000,
	261	+ 'claude-mythos-5-1' => 128000,
	262	+ 'claude-mythos-5' => 128000,
	263	+ 'claude-opus-5' => 128000,
	264	+ 'claude-opus-4-8' => 128000,
	265	+ 'claude-opus-4-7' => 128000,
	266	+ 'claude-opus-4-6' => 128000,
	267	+ 'claude-sonnet-5' => 128000,
	268	+ 'claude-sonnet-4-6' => 128000,
	269	+ // 64K output.
	270	+ 'claude-haiku-4-5' => 64000,
	271	+ ];
	272	+
	273	+ /**
	274	+ * Upper bound per use case, applied after the percentage.
	275	+ *
	276	+ * Two reasons these exist rather than letting the percentage run against a
	277	+ * 128K ceiling.
	278	+ *
	279	+ * Requests here are a single blocking HTTP call with a 120s timeout and no
	280	+ * streaming, so 0.9 x 128000 would risk running past the timeout instead of
	281	+ * returning — trading a truncation failure for a timeout failure. 16000
	282	+ * leaves room for the brief's JSON plus reasoning tokens while staying
	283	+ * answerable; raise it only alongside streaming.
	284	+ *
	285	+ * And correcting the ceiling would otherwise inflate every other use case
	286	+ * as a side effect — seo_metadata would jump from ~1,229 tokens to ~19,200
	287	+ * purely because this bug was fixed. Metadata generation already works, so
	288	+ * it keeps its cost profile (#665).
	289	+ *
	290	+ * @since 2.7.0
	291	+ * @var array<string, int>
	292	+ */
	293	+ private const USE_CASE_TOKEN_CAPS = [
	294	+ 'content_brief' => 16000,
	295	+ 'llms_txt' => 16000,
	296	+ 'analysis' => 8000,
	297	+ 'seo_metadata' => 4000,
	298	+ 'optimization' => 4000,
	299	+ 'default' => 4000,
	300	+ ];
	301	+
	302	+ /**
	303	+ * Ceiling for a model this table does not know.
	304	+ *
	305	+ * The previous behaviour for every model, kept for older and unrecognised
	306	+ * ones: 8192 is accepted without an extended-output beta header, so it is
	307	+ * the safe answer when we cannot identify the family.
	308	+ *
	309	+ * @since 2.7.0
	310	+ * @var int
	311	+ */
	312	+ private const FALLBACK_OUTPUT_LIMIT = 8192;
	313	+
	314	+ /**
247	315	* Maximum completion (output) tokens accepted for a single Claude request.
248	316	*
249		- * 8192 is accepted by every current Claude model without the extended-output
250		- * beta header, so it is a safe per-request ceiling. Kept as a method (rather
251		- * than a constant) to mirror the other clients and allow per-model tuning.
	317	+ * This returned a flat 8192 for every model and ignored $model entirely, so
	318	+ * Content Brief was capped at a fraction of the available budget and
	319	+ * truncated before its structured JSON completed — on every Claude model,
	320	+ * every time. Current models also emit reasoning tokens from the same
	321	+ * output budget, which is why it failed so reliably rather than
	322	+ * intermittently (#665).
252	323	*
253		- * @param string $model Model ID (reserved for future per-model limits).
	324	+ * @param string $model Model ID.
254	325	* @return int Maximum output tokens.
255	326	*/
256	327	private function get_max_completion_tokens(string $model): int {
257		- return 8192;
	328	+ $model = strtolower(trim($model));
	329	+
	330	+ if ('' === $model) {
	331	+ return self::FALLBACK_OUTPUT_LIMIT;
	332	+ }
	333	+
	334	+ $limits = self::MODEL_OUTPUT_LIMITS;
	335	+
	336	+ // Longest prefix first, so a point release never matches the shorter
	337	+ // family id that is a prefix of it.
	338	+ uksort(
	339	+ $limits,
	340	+ static function (string $a, string $b): int {
	341	+ return strlen($b) <=> strlen($a);
	342	+ }
	343	+ );
	344	+
	345	+ foreach ($limits as $prefix => $limit) {
	346	+ if (0 === strpos($model, $prefix)) {
	347	+ return $limit;
	348	+ }
	349	+ }
	350	+
	351	+ return self::FALLBACK_OUTPUT_LIMIT;
258	352	}
259	353
260	354	/**
261	355	* Recommended output-token budget for a given use case.
		@@ -278,9 +372,13 @@
278	372	'optimization' => 0.15,
279	373	];
280	374	$percentage = $recommendations[$use_case] ?? 0.15;
281	375
282		- return (int) ($max_tokens * $percentage);
	376	+ $budget = (int) ($max_tokens * $percentage);
	377	+
	378	+ $cap = self::USE_CASE_TOKEN_CAPS[$use_case] ?? self::USE_CASE_TOKEN_CAPS['default'];
	379	+
	380	+ return max(1, min($budget, $cap));
283	381	}
284	382
285	383	/**
286	384	* Test API connection