A 5% adjustment accounted for improved code survival rates.
See my blog post with @mmooritz for additional details on these analyses: https://bigspin.ai/resources/the-decline-of-token-level-purchasing-power
Does a token buy you more or less now than it did a few months ago? We built a consumer price index (CPI) for AI coding output from Anthropic's Opus 4.6 model in SWE-chat, Feb 5–Apr 15, 2026. What we find looks like tokenflation:
Does a token buy you more or less now than it did a few months ago? We built a consumer price index (CPI) for AI coding output from Anthropic's Opus 4.6 model in SWE-chat, Feb 5–Apr 15, 2026. What we find looks like tokenflation:
The goods in our CPI basket are standard engineering outcomes like PRs, but we also included knowledge capture (any time work is stored somewhere durable), which shows the most positive trend and which resonates with us as heavy users of these tools.
See my blog post with @mmooritz for additional details on these analyses: https://bigspin.ai/resources/the-decline-of-token-level-purchasing-power
Nominating “Tokenflation” for early-decision word-of-the-year for 2026.
Does a token buy you more or less now than it did a few months ago? We built a consumer price index (CPI) for AI coding output from Anthropic's Opus 4.6 model in SWE-chat, Feb 5–Apr 15, 2026. What we find looks like tokenflation:
We ourselves derive significant value from these tools and wonder whether measuring other relevant outcomes would change the CPI. That said, it seems clear that AI is becoming more expensive in ways that might be outpacing its current value.
What is causing the decline in token-level purchasing power? One relevant thing we know is that the job of an Opus 4.6 token changed dramatically in this period, from purely writing code in February to coding, thinking, and explaining in mid-April:
What is causing the decline in token-level purchasing power? One relevant thing we know is that the job of an Opus 4.6 token changed dramatically in this period, from purely writing code in February to coding, thinking, and explaining in mid-April:
To give you a concrete sense for what the above declines mean, here are the estimated outcomes for a 10K-token session:
Feb 2026: a third of a PR, 630 lines of code drafted, 8 files touched, 1 knowledge capture.
Apr 2026: a sixth of a PR, 91 lines of code drafted, 4 files touched, less than half a knowledge capture.
Our CPI includes a hedonic adjustment based on the code survival rate. This goes from about 90% to 95% in our time period, so we incorporate a uniform 5% adjustment. This softens the tokenflation; code output seems to be improving!
The goods in our CPI basket are standard engineering outcomes like PRs, but we also included knowledge capture (any time work is stored somewhere durable), which shows the most positive trend and which resonates with us as heavy users of these tools.
To give you a concrete sense for what the above declines mean, here are the estimated outcomes for a 10K-token session:
Feb 2026: a third of a PR, 630 lines of code drafted, 8 files touched, 1 knowledge capture.
Apr 2026: a sixth of a PR, 91 lines of code drafted, 4 files touched, less than half a knowledge capture.
Another link to the associated blog post: https://bigspin.ai/resources/the-decline-of-token-level-purchasing-power
We would love to expand our analysis to other models, providers, and time periods. We've created a Claude Code tool/plugin that you can use to analyze your own usage. The tool shares no data with us, but perhaps we could all pool our anonymized data to learn more about the true value of our tokens: https://github.com/bigspinai/plugins
We would love to expand our analysis to other models, providers, and time periods. We've created a Claude Code tool/plugin that you can use to analyze your own usage. The tool shares no data with us, but perhaps we could all pool our anonymized data to learn more about the true value of our tokens: https://github.com/bigspinai/plugins
@ChrisGPotts the framing that might matter more: what's the total cost to reach a fixed quality bar, not cost per line of output. if the first draft gets worse but the fix loop is faster, the CPI looks bad while the real outcome improves. time-to-acceptable-PR feels like the right unit
A 5% adjustment accounted for improved code survival rates.
See my blog post with @mmooritz for additional details on these analyses: https://bigspin.ai/resources/the-decline-of-token-level-purchasing-power
Does a token buy you more or less now than it did a few months ago? We built a consumer price index (CPI) for AI coding output from Anthropic's Opus 4.6 model in SWE-chat, Feb 5–Apr 15, 2026. What we find looks like tokenflation:
