The Decline of Token-Level Purchasing Power – Bigspin AI
We analyzed thousands of conversations in a first-of-its-kind study. The biggest failures were invisible.
See what they are
Invisible Failures in AI
See report
About us
Resources
Research
Contact us
About us
Resources
Research
Contact us
Models
The Decline of Token-Level Purchasing Power
The Decline of Token-Level Purchasing Power
Coding assistants like Claude and Codex generate more tokens than ever before, and the cost of these tokens is generally rising.
Coding assistants like Claude and Codex generate more tokens than ever before, and the cost of these tokens is generally rising.
June 8, 2026
Reading time:
min
Coding assistants like Claude and Codex generate more tokens than ever before, and the cost of these tokens is generally rising. We all sense these trends as we watch the token counts climb in the apps, and we feel them viscerally when we review our monthly bills.<br>Are all these tokens actually leading to better results for us? This question is increasingly on the minds of individual users and large organizations, because the frontier LLM providers have started to confront us with the true cost of all our usage. The sticker shock can be significant; the days of token-maxxing may be numbered.<br>What is the purchasing power of a token? Does a token buy you more or less now than it did a few months ago? To address these questions, we build a consumer price index (CPI) for AI coding output from Anthropic’s Opus 4.6 model. Unlike the usual sort of CPI, ours contains engineering events rather than milk and rent, but the logic is the same. Our sample comes from SWE-chat, specifically, 5,851 real Opus 4.6 sessions from February 5 to April 15, 2026.<br>The short answer to our question is sobering:<br>The purchasing power of an Opus 4.6 token fell dramatically from February 5 to April 15, 2026. The downward trend holds for all the goods in our engineering basket.<br>This looks like tokenflation : your Opus 4.6 token budget is not going to take you as far as it did even just a few months ago. We do not know the full causal story behind this trend, though we strongly suspect that changes in the job mix of tokens (away from pure code generation and toward thinking and explanation) are a significant factor (see below for details).<br>Important note: our analysis covers only Anthropic’s Opus 4.6 model for the period February 5 to April 15, 2026. We do not have sufficient data from other models, providers, or time periods to broaden the empirical picture right now. We hope this post serves as a call to action for people to begin tracking these trends industry-wide, since they are critical to the economics of AI.
Breakdown by goods<br>Our basket of deliverables includes some standard engineering outcomes: lines of code and documentation, both drafted and committed; files touched; and PRs shipped. However, we feel strongly that these outcomes alone fail to capture the true value of our coding agent sessions, which range from quick check-ins to extended collaborations that are more about exploration and discovery than shipping code. (If you are interested in better understanding your own Claude Code sessions, check out our free, anonymous plugin.)<br>To help acknowledge these intangibles, we defined a novel knowledge capture good: an instance where work is stored somewhere durable (a CLAUDE.md, an AGENTS.md, a plan doc, a skill definition, or a project rule). This is still a purely mechanical signal, but it helps us track how knowledge accumulates in coding agent sessions.<br>The following figure provides a detailed look at our CPI. The x-axis tracks time, and the y-axis tracks output units per token, normalized to Phase A. All of the deliverables trend downward. The black line is the average of all goods, which indicates that a token in mid-April bought about $0.23 on the February $1.00.
We have followed standard CPI practices for this analysis: our basket of engineering goods has five items in it (PRs, files touched, etc.). In each phase P, the price of one of these goods G is the total output tokens in P divided by the total units of G produced in P. To convert these prices to dollars, we used Opus 4.6’s current list price ($75 per million output tokens), applied to all periods to isolate token-quantity inflation from any pricing changes Anthropic may have made. We also implemented a hedonic adjustment by tracking code survival, i.e., the fraction of lines committed during each phase that still exist in the repository 4 days later. We find that the code survival rate is about 90% in Phase A and 95% in Phase E, so we incorporate a uniform 5% hedonic adjustment. This has the effect of softening the tokenflation trends.<br>The time periods for our analysis were chosen deliberately (see the timeline annotations at the bottom of the figure). The choices mainly reflect our attempt to navigate the exogenous events that Anthropic described in their April postmortem....