Claude Opus 4.6-Level Performance Will Cost as Much as Claude 3.5 Haiku in 12 Months
Feb 8, 2026
TL;DR: Claude Opus 4.6 costs $10/M blended* today. Using HumanEval benchmark data and historical pricing trends, equivalent coding performance will cost $1.50-$2.00/M in 12 months—same as Claude 3.5 Haiku ($1.60/M) today.
*Blended pricing assumes 3:1 input/output token ratio. Claude Opus 4.6 is $5 input / $25 output per 1M tokens.
HumanEval: Price History
Writing Python functions from docstrings.
(Obviously the price won’t go to zero)
In March 2023, GPT-4 achieved 67% HumanEval for $37.50/M. By September 2024, Qwen2.5-Coder achieved 93% for $0.09/M. That’s a 417x price decline in 18 months for far superior capability.
You can repeat the experiment with different benchmarks (GPQA Diamond, MMLU, etc), and it tells the same story.
Data sources (HumanEval score, pricing at 3:1 input/output ratio)
- GPT-4 (Mar 2023): 67.0% , $30/$60 = $37.50/M
- Claude 2 (Jul 2023): 71.2% , $8/$24 = $12.00/M
- DeepSeek-Coder-33B (Nov 2023): 74.4% , ~$1.00/$2.00 = $1.25/M
- Claude 3 Opus (Mar 2024): 84.9% , $15/$75 = $30.00/M
- GPT-4o (May 2024): 90.2% , $2.50/$10 = $4.38/M
- Claude 3.5 Sonnet (Jun 2024): 92.0% , $3/$15 = $6.00/M
- GPT-4o mini (Jul 2024): 87.2% , $0.15/$0.60 = $0.26/M
- o1-mini (Sep 2024): 92.4% , $3/$12 = $6.00/M
- Qwen2.5-Coder-32B (Sep 2024): 92.7% , $0.09/$0.09 = $0.09/M
- Claude 3.5 Haiku (Nov 2024): 88.1% , $0.80/$4.00 = $1.60/M
- DeepSeek V3 (Dec 2024): 82.6% , $0.27/$1.10 = $0.48/M
Our users keep asking for more usage. This is why I tell them to expect 10x cheaper usage in a year.
See also
- Epoch AI: Estimated 9-900x annual price decline (source )
- NeurIPS 2025: 5-10x annual decline estimate (source )
- Introl: Estmated GPT-4 50x cheaper in 3 years (source )