OpenAI Codex Lead Reveals Tokenizer Inefficiency Can Spike API Bills by 34%
新智元 · wechat · 2026-08-19
OpenAI Codex lead Tibo revealed that tokenizer efficiency varies wildly between models. The same text costs 766 tokens on GPT-5.6Sol but 1,170 on Claude Opus 5, meaning the latter could cost 34.5% more despite identical unit prices.
Real costs depend on four factors beyond tokenization:
- Caching: GPT-5.6Sol offers cached input at 10% of the standard price.
- Output Costs: GPT outputs cost $30/1M tokens vs. Claude's $25+, heavily impacting agent workflows.
- Long Context Multipliers: GPT-5.6Sol doubles input prices and increases output prices by 1.5x for requests over 272K tokens.
- Effective Window: Users may need to manually configure Codex to unlock the full 1M token window.
Tibo advises focusing on the "price per successful outcome" rather than token price, urging users to benchmark their specific workloads considering tokenization, caching, and context multipliers.
More from Infra
- 16-bit Model Requires 60GB VRAM; 4-bit Quantized Fits on 24GB — LeviTurk · 2026-08-19
- Why Removing the Vision Encoder Can Be Better — From an Infra Perspective — liuziwei7 · 2026-08-19
- Matmul Optimization Bottleneck: Data Movement, Not Multiplications — yaroslavvb · 2026-08-19
- Alibaba Cloud Opens 3rd Data Center in Korea; Doubao Adds PC Control — 创业邦 · 2026-08-19
- China opens world's largest AI data center targeting 1 million GPUs — teortaxesTex · 2026-08-19
- Morgan Stanley: US data centers need 68GW power by 2028 — tctjr · 2026-08-19