diff --git a/src/tools/web.ts b/src/tools/web.ts
index 17b4301..6a059fd 100644
--- a/src/tools/web.ts
+++ b/src/tools/web.ts
@@ -843,8 +843,70 @@ function isLowValueNewsResult(url: string): boolean {
   }
 }
 
+// ── Comparison-query splitting ────────────────────────────────────────────────
+//
+// "A vs B <spec>" queries reliably return review/overview pages that name both products and
+// give the figures for neither, while searching each product separately returns the numbers
+// immediately. Verified 2026-08-10: "RTX 3090 vs RTX 4080 Super memory bandwidth" produced a
+// generic overview with no figure; the two separate searches produced 936 GB/s and 736 GB/s.
+//
+// That finding went into the web_search tool description as an instruction the same day, and the
+// model ignored it — production log 2026-08-11 shows three consecutive turns issuing the exact
+// combined form it was told not to use ("RTX 4080 vs RTX 5060 performance comparison specs"),
+// each returning dead technical.city links, each producing an answer of pure generalities with no
+// number in it. Same lesson as news_search's country/category params and weather's multi-city
+// batching earlier that day: a tool description cannot enforce query construction on this model,
+// so the split happens here in code instead ([[feedback_local_model_needs_code_backstop]]).
+//
+// The combined query still runs — when a comparison page IS alive it is genuinely the best source
+// for this question, and dropping it would trade one failure mode for another. The split searches
+// are additive and capped tighter so the extra grounding does not blow up the context.
+const COMPARISON_SEPARATOR = /\s+(?:vs\.?|versus)\s+|\s*와\s+|\s*과\s+/i;
+/** Words describing WHAT is being compared — kept, since they narrow each single-entity search. */
+const COMPARISON_ATTRIBUTE = /\b(performance|specs?|specifications?|benchmarks?|review|memory\s*bandwidth|bandwidth|tflops|vram|price|speed)\b|성능|스펙|사양|벤치마크|대역폭|가격|속도/gi;
+/** Words meaning "compare these" — dropped, since they are meaningless in a single-entity search. */
+const COMPARISON_VERB = /\b(comparison|compare[ds]?|versus|vs\.?|difference|diff)\b|비교|차이/gi;
+
+export interface ComparisonSplit { a: string; b: string; }
+
+/**
+ * Split "RTX 4080 vs RTX 5060 performance comparison specs" into
+ * "RTX 4080 performance specs" + "RTX 5060 performance specs", or null when the query is not a
+ * spec comparison. Deliberately conservative: "Lakers vs Celtics" has no attribute word and no
+ * model numbers, so it is left alone rather than turned into two unrelated searches.
+ */
+export function splitComparisonQuery(query: string): ComparisonSplit | null {
+  const q = String(query || '').trim();
+  if (!q) return null;
+  const parts = q.split(COMPARISON_SEPARATOR);
+  if (parts.length !== 2) return null;
+
+  const attributes = Array.from(new Set(
+    (q.match(COMPARISON_ATTRIBUTE) || []).map(s => s.trim().toLowerCase()),
+  ));
+  const strip = (s: string) => s
+    .replace(COMPARISON_ATTRIBUTE, ' ')
+    .replace(COMPARISON_VERB, ' ')
+    .replace(/\s+/g, ' ')
+    .trim();
+  const [entityA, entityB] = parts.map(strip);
+  if (entityA.length < 2 || entityB.length < 2) return null;
+
+  // Either an explicit attribute ("performance", "스펙") or two model-number-shaped entities.
+  // Without one of those this is not a spec comparison and splitting would just lose meaning.
+  const bothLookLikeModels = /\d/.test(entityA) && /\d/.test(entityB);
+  if (!attributes.length && !bothLookLikeModels) return null;
+
+  const tail = attributes.length ? attributes.join(' ') : 'specs';
+  return { a: `${entityA} ${tail}`.trim(), b: `${entityB} ${tail}`.trim() };
+}
+
 // ── Main web_search tool ──────────────────────────────────────────────────────
-export async function executeWebSearch(args: { query: string; max_results?: number }): Promise<ToolResult> {
+export async function executeWebSearch(args: { query: string; max_results?: number; _noSplit?: boolean }): Promise<ToolResult> {
+  if (!args._noSplit) {
+    const split = splitComparisonQuery(args.query || '');
+    if (split) return runComparisonSearch(args, split);
+  }
   if (!args.query?.trim()) return { success: false, error: 'query is required' };
   let limit = Math.min(args.max_results ?? 5, 10);
   if (isPriceQuery(args.query)) limit = Math.max(limit, 5);
@@ -1031,6 +1093,39 @@ async function extractAnswerFromResults(query: string, rawResults: string): Prom
   }
 }
 
+/**
+ * Runs the combined query plus one search per entity, and hands the model all three labelled.
+ * Partial failure is fine — any section that came back with results is still grounding the model
+ * did not have before, so this only ever falls back to whatever the combined query alone returned.
+ */
+async function runComparisonSearch(
+  args: { query: string; max_results?: number },
+  split: ComparisonSplit,
+): Promise<ToolResult> {
+  const splitLimit = Math.min(args.max_results ?? 5, 3);
+  const [combined, a, b] = await Promise.all([
+    executeWebSearch({ ...args, _noSplit: true }),
+    executeWebSearch({ query: split.a, max_results: splitLimit, _noSplit: true }),
+    executeWebSearch({ query: split.b, max_results: splitLimit, _noSplit: true }),
+  ]);
+
+  const sections: string[] = [];
+  const push = (label: string, r: ToolResult) => {
+    if (r.success && String(r.stdout || '').trim()) sections.push(`[${label}]\n${String(r.stdout).trim()}`);
+  };
+  push(`검색: ${args.query}`, combined);
+  push(`검색: ${split.a}`, a);
+  push(`검색: ${split.b}`, b);
+
+  if (!sections.length) return combined;
+  console.log(`[v2] web_search comparison split: "${args.query}" → "${split.a}" + "${split.b}"`);
+  return {
+    success: true,
+    data: { ...(combined.data as any || {}), comparison_split: [split.a, split.b] },
+    stdout: `비교 질문이라 각 대상을 따로 검색했습니다. 아래 세 검색 결과를 모두 근거로 쓰세요.\n\n${sections.join('\n\n')}`,
+  };
+}
+
 export async function executeWebSearchWithExtraction(args: { query: string; max_results?: number }): Promise<ToolResult> {
   const res = await executeWebSearch(args);
   if (!res.success || !res.stdout) return res;
diff --git a/tests/web-comparison-split.test.ts b/tests/web-comparison-split.test.ts
new file mode 100644
index 0000000..0e15999
--- /dev/null
+++ b/tests/web-comparison-split.test.ts
@@ -0,0 +1,73 @@
+/**
+ * splitComparisonQuery — 비교 질문을 대상별 검색으로 쪼개기
+ *
+ * 2026-08-10에 "A vs B로 합치지 말고 따로 검색하라"를 web_search 설명에 넣었는데, 08-11 프로덕션
+ * 로그에서 모델이 세 턴 연속으로 정확히 금지된 형태를 그대로 날렸다:
+ *   web_search({"query":"RTX 4080 vs RTX 5060 performance comparison specs"})
+ * 매번 technical.city 링크가 죽어 있었고(2/5 dead) 답변은 수치 0건의 일반론이었다. 설명으로
+ * 강제가 안 되니 코드에서 쪼갠다.
+ *
+ * 이 테스트가 지키는 선: 쪼갤 만한 것만 쪼갠다. "Lakers vs Celtics"를 "Lakers"/"Celtics" 두
+ * 검색으로 만들면 원래 질문의 의미가 사라진다 — 오탐 쪽이 더 비싸다.
+ */
+
+import { test, describe } from 'node:test';
+import assert from 'node:assert/strict';
+import { splitComparisonQuery } from '../src/tools/web';
+
+describe('쪼개야 하는 쿼리', () => {
+  test('실제 사고 쿼리 — 비교 동사는 버리고 속성어는 남긴다', () => {
+    const r = splitComparisonQuery('RTX 4080 vs RTX 5060 performance comparison specs');
+    assert.deepEqual(r, { a: 'RTX 4080 performance specs', b: 'RTX 5060 performance specs' });
+  });
+
+  test('속성어가 없어도 양쪽이 모델명 꼴이면 쪼갠다', () => {
+    const r = splitComparisonQuery('RTX 4080 vs RTX 5060');
+    assert.deepEqual(r, { a: 'RTX 4080 specs', b: 'RTX 5060 specs' });
+  });
+
+  test('어제 검증한 대역폭 케이스', () => {
+    const r = splitComparisonQuery('RTX 3090 vs RTX 4080 Super memory bandwidth');
+    assert.ok(r);
+    assert.ok(r!.a.startsWith('RTX 3090'));
+    assert.ok(r!.b.startsWith('RTX 4080 Super'));
+    assert.ok(r!.a.includes('bandwidth') && r!.b.includes('bandwidth'));
+  });
+
+  test('한국어 "와/과 ... 비교"도 쪼갠다', () => {
+    assert.deepEqual(
+      splitComparisonQuery('RTX 4080과 5060 성능 비교'),
+      { a: 'RTX 4080 성능', b: '5060 성능' },
+    );
+  });
+
+  test('vs. / versus 표기도 인식한다', () => {
+    assert.ok(splitComparisonQuery('M4 Max vs. M3 Ultra benchmark'));
+    assert.ok(splitComparisonQuery('A100 versus H100 tflops'));
+  });
+});
+
+describe('쪼개면 안 되는 쿼리', () => {
+  test('스포츠 대진 — 속성어도 모델번호도 없다', () => {
+    assert.equal(splitComparisonQuery('Lakers vs Celtics'), null);
+  });
+
+  test('일반 검색은 건드리지 않는다', () => {
+    assert.equal(splitComparisonQuery('오늘 서울 날씨'), null);
+    assert.equal(splitComparisonQuery('RTX 4080 memory bandwidth'), null);
+  });
+
+  test('비교 대상이 셋이면 판단하지 않는다', () => {
+    assert.equal(splitComparisonQuery('A vs B vs C specs'), null);
+  });
+
+  test('"결과"처럼 과/와로 끝나는 단어에 잘못 걸리지 않는다', () => {
+    assert.equal(splitComparisonQuery('검색결과 스펙'), null);
+    assert.equal(splitComparisonQuery('효과 성능'), null);
+  });
+
+  test('빈 입력', () => {
+    assert.equal(splitComparisonQuery(''), null);
+    assert.equal(splitComparisonQuery(null as any), null);
+  });
+});