A stable price per token is not the same as a stable bill. That is the arithmetic OpenAI’s GPT-5.6 Sol update obscures: the company kept its listed fresh-input, cached-input, and output prices unchanged from GPT-5.5, but developer Vincent Schmalbach’s own Codex usage logs show the new model consuming roughly 2.25 times as many tokens to do comparable work. Effective cost is price multiplied by consumption, and only the price half of that equation showed up in OpenAI’s announcement.

Schmalbach, writing on his blog on 3 August 2026, compared two matched 14-day windows of deduplicated local Codex logs, one running GPT-5.5 xhigh, one running GPT-5.6 Sol xhigh. Session counts barely moved, 1,667 before versus 1,715 after, a rise of just 2.9 percent. Token consumption did not stay proportional. GPT-5.5 xhigh used 12.17 billion tokens across those sessions, an average of 7.30 million tokens per session. GPT-5.6 Sol xhigh used 28.22 billion tokens across a nearly identical session count, an average of 16.45 million tokens per session. Total usage rose 2.32 times, and usage per session rose 2.25 times, a 125 percent increase.

Since the listed per-token prices for both models match, that consumption jump converts almost directly into cost. A flat price schedule times 2.25 times the tokens produces a bill roughly 2.25 times larger for a similar mix of work, even though nothing on OpenAI’s pricing page announced an increase. The change compounds further. GPT-5.6 Sol also introduces a cache-write fee, a cost line that GPT-5.5 never carried, on token traffic that used to be a savings mechanism.

For quota holders rather than metered API users, Schmalbach reports the effect directly. His three subscriptions used to last roughly a week under heavy use. Since switching to GPT-5.6 Sol, about a day of medium use now exhausts them, and the several banked usage resets he rarely touched under GPT-5.5 have run out on every account.

This is one developer’s measurement of his own Codex workflow, not a controlled study, and token consumption swings widely by task type, prompt structure, and codebase size. The direction still deserves attention. It describes the same mechanism any quota holder would hit, only the size of the multiplier would differ by workload. Today’s issue also carries a cost-per-task ranking that measures exactly this, the total bill to finish a job rather than the sticker price per token; the two findings describe the same arithmetic from opposite directions.

Teams working off a fixed token quota should stop budgeting from list price and start budgeting from logged consumption per completed task, re-measured after every model swap. A published price freeze can still double, or more than double, a monthly bill.

Vincent Schmalbach reported these findings on his personal blog on 3 August 2026, based on his own Codex usage logs.