diff --git a/lib/token-utils.ts b/lib/token-utils.ts
index bba07fa1..7168b7f6 100644
--- a/lib/token-utils.ts
+++ b/lib/token-utils.ts
@@ -2,8 +2,16 @@ import { SessionState, WithParts } from "./state"
 import { AssistantMessage, UserMessage } from "@opencode-ai/sdk/v2"
 import { Logger } from "./logger"
 import * as _anthropicTokenizer from "@anthropic-ai/tokenizer"
-const anthropicCountTokens = (_anthropicTokenizer.countTokens ??
-    (_anthropicTokenizer as any).default?.countTokens) as typeof _anthropicTokenizer.countTokens
+const getAnthropicTokenizer = (_anthropicTokenizer.getTokenizer ??
+    (_anthropicTokenizer as any).default?.getTokenizer) as typeof _anthropicTokenizer.getTokenizer
+// The package's countTokens() builds and frees a new WASM tokenizer on every call (~70 ms each), which
+// blocks the host for seconds when many tool outputs are counted. Build it once and reuse it: the
+// counts are identical (same encoder, same NFKC normalisation and special-token handling).
+let anthropicTokenizer: ReturnType<typeof getAnthropicTokenizer> | undefined
+function anthropicCountTokens(text: string): number {
+    anthropicTokenizer ??= getAnthropicTokenizer()
+    return anthropicTokenizer.encode(text.normalize("NFKC"), "all").length
+}
 import { getLastUserMessage } from "./messages/query"
 
 export function getCurrentTokenUsage(state: SessionState, messages: WithParts[]): number {
diff --git a/tests/token-counting.test.ts b/tests/token-counting.test.ts
index 03e9439c..964c6faa 100644
--- a/tests/token-counting.test.ts
+++ b/tests/token-counting.test.ts
@@ -4,6 +4,7 @@ import type { WithParts } from "../lib/state"
 import {
     COMPACTED_TOOL_OUTPUT_PLACEHOLDER,
     countAllMessageTokens,
+    countTokens,
     countToolTokens,
     estimateTokensBatch,
     extractCompletedToolOutput,
@@ -172,3 +173,25 @@ test("counting uses the compacted tool placeholder for completed outputs", () =>
     assert.equal(extractCompletedToolOutput(part), COMPACTED_TOOL_OUTPUT_PLACEHOLDER)
     assertCounted(part, [JSON.stringify(input), COMPACTED_TOOL_OUTPUT_PLACEHOLDER])
 })
+
+test("countTokens matches @anthropic-ai/tokenizer and reuses one tokenizer", async () => {
+    const tokenizer = await import("@anthropic-ai/tokenizer")
+    const reference = (tokenizer.countTokens ??
+        (tokenizer as any).default?.countTokens) as typeof tokenizer.countTokens
+    const samples = [
+        "hello world",
+        "Crème brûlée — fi ligature ① and emoji 🚀",
+        JSON.stringify({ path: "src/index.ts", edits: [{ line: 1, text: "const x = 1" }] }),
+        "x".repeat(10_000),
+        "<|endoftext|> special tokens are counted, not rejected",
+    ]
+    for (const sample of samples) assert.equal(countTokens(sample), reference(sample))
+
+    // Building a tokenizer per call took ~70 ms; 200 calls with a shared one take a few milliseconds.
+    const start = performance.now()
+    for (let index = 0; index < 200; index++) countTokens(`tool output ${index} `.repeat(50))
+    assert.ok(
+        performance.now() - start < 2_000,
+        "countTokens should not rebuild the tokenizer on every call",
+    )
+})