AI & Agent Dev Bug Sandbox Back to Radar

Count_tokens_approximately Inflates Estimates 5-10X For Reasoning/Thinking Blocks

count_tokens_approximately has no dedicated case for reasoning or thinking content blocks, causing them to fall through to the catch-all len(repr(block)) branch. This stringifies the entire block dict, including provider metadata such as extras.encrypted_content, inflating token estimates by 5-10x and breaking context-window management.

highConfidence 93%Langchain-Core Affected V1.5.5

Origin Analysis

Missing explicit handling for content blocks with type 'reasoning' or 'thinking' in count_tokens_approximately. The fallback len(repr(block)) serializes the entire block dict, including long encrypted_content fields inside extras, rather than extracting only the readable reasoning/thinking text.

from langchain_core.messages import AIMessage from langchain_core.messages.utils import count_tokens_approximately reasoning_text = "I should inspect the staging model." ciphertext = "g" * 4000 block_with_extras = { "type": "reasoning", "id": "rs_abc", "reasoning": reasoning_text, "extras": {"content": [], "encrypted_content": ciphertext}, } block_without_extras = { "type": "reasoning", "id": "rs_abc", "reasoning": reasoning_text, } print(count_tokens_approximately([AIMessage(content=[block_with_extras], id="ai-1")])) # ~1070 print(count_tokens_approximately([AIMessage(content=[block_without_extras], id="ai-1")])) # ~15

Fixing Code Block

0x7F 45 4C 46 02 01 01 000x90 12 AB 04 C1 7E 33 08
0xA4 1C 90 3E 77 0B D2 580x11 F6 6A 2C 80 19 DE 44
0x00 00 00 00 03 00 3E 000x01 00 B7 4D 55 90 0C 61
0x6C 69 62 63 2E 73 6F 2E0x36 00 8E 17 42 AA 05 F3
0xDE C0 DE 01 8B 45 FC 890xC7 5D 20 6F 91 03 E4 78
0x13 37 00 4F 9A 2D E1 700x08 B4 5C 22 71 D0 4A 16
0x48 8D 3D 22 01 00 00 E80x91 FF 0A 63 2E 57 B8 04
0x5A 91 C3 0E 44 70 1B 880x02 6F 39 C4 15 80 ED 27
0xFF 15 40 20 60 00 48 890xC3 90 7A 01 4C 28 B1 56
0x0F 1F 44 00 00 55 48 890xE5 C3 18 6D 92 0B 33 A7

Edge Case Audit

This hotfix only covers top-level 'reasoning' and 'thinking' fields. If providers embed encrypted or long payloads inside other block types or under different keys (e.g., nested inside extras for 'text' blocks), the fallback len(repr(block)) will still inflate counts. Future-proof by routing content extraction through get_buffer_string or maintaining a central block-type registry. Rollback is straightforward: restore the original function body; no data migration is required. Test with both OpenAI reasoning blocks and Anthropic thinking blocks to ensure no regression.