Cursor removed cost information from the usage page and CSV export

5 hours ago

Cursor replaced dollar amounts with token amounts on the usage page for non-enterprise plans, frustrating users who tracked spending.
Enterprise plans still show dollar amounts; individual and self-serve Teams plans show tokens, with on-demand costs only in the Cost column.
Users can find dollar totals on Dashboard > Spending and per-user on-demand on Dashboard > Members, but per-model breakdown is unavailable.
The change is intentional: dollar values were removed from usage endpoints and CSV exports to avoid confusion about included vs. billed usage.
Users criticize the lack of transparency, asking for a toggle to switch between tokens and dollars for better budget tracking.
The API no longer returns per-request cost fields for self-serve plans, even historically, breaking independent cost tracking tools.
Users request restoration of dollar-based usage analysis or improved visualization to distinguish included usage from actual charges.

FCC voted to eliminate a Biden-era rule requiring ISPs to list all passthrough fees on broadband price labels.
ISPs can now display an aggregate 'up to' amount instead of individually listing fees, making price comparison harder for consumers.
Chairman Carr proposed allowing ISPs to provide links to price labels instead of displaying them prominently on ordering pages.

CSV Unwrap was not focused on parsing CSV, but on helping users build a mental model of unfamiliar datasets during the first few minutes of exploration.
The project assumed clean datasets, but discovered dirty, inconsistent data reflecting human decisions and edge cases.
The table viewer became a major challenge, requiring on-demand loading and efficient rendering without noticeable complexity.

Claude Code uses significantly more tokens than OpenCode for the same tasks, with baseline overhead around 33,000 tokens vs 7,000 for a simple reply.
Cache inefficiency is a major factor; Claude Code rewrites tens of thousands of prompt-cache tokens mid-session, leading to up to 54x more cache writes than OpenCode.
Configuration elements like instruction files and MCP servers add substantial token overhead, with a real setup reaching 75,000–85,000 tokens before user input.

AI agents often overthink simple greetings like 'Hi', leading to high time costs despite low token costs.
Benchmarking 14 models shows that waiting time, not token pricing, is the real expense; waiting can cost over 20 times more than tokens.
Some models (e.g., Sonnet) respond to 'Hi' with excessive tool calls, auditing repositories and making unsolicited commits.

Claude subscribers sued over limits in 2026, but limits actually increased that year.
Three metering systems exist: a rolling five-hour window, a weekly cap, and a premium-model meter.
Anthropic publishes usage estimates as ranges (bands), not exact numbers, causing opacity.

FCC voted to eliminate a Biden-era rule requiring ISPs to list all passthrough fees on broadband price labels.
ISPs can now display an aggregate 'up to' amount instead of individually listing fees, making price comparison harder for consumers.
Chairman Carr proposed allowing ISPs to provide links to price labels instead of displaying them prominently on ordering pages.

CSV Unwrap was not focused on parsing CSV, but on helping users build a mental model of unfamiliar datasets during the first few minutes of exploration.
The project assumed clean datasets, but discovered dirty, inconsistent data reflecting human decisions and edge cases.
The table viewer became a major challenge, requiring on-demand loading and efficient rendering without noticeable complexity.

Claude Code uses significantly more tokens than OpenCode for the same tasks, with baseline overhead around 33,000 tokens vs 7,000 for a simple reply.
Cache inefficiency is a major factor; Claude Code rewrites tens of thousands of prompt-cache tokens mid-session, leading to up to 54x more cache writes than OpenCode.
Configuration elements like instruction files and MCP servers add substantial token overhead, with a real setup reaching 75,000–85,000 tokens before user input.

AI agents often overthink simple greetings like 'Hi', leading to high time costs despite low token costs.
Benchmarking 14 models shows that waiting time, not token pricing, is the real expense; waiting can cost over 20 times more than tokens.
Some models (e.g., Sonnet) respond to 'Hi' with excessive tool calls, auditing repositories and making unsolicited commits.

Claude subscribers sued over limits in 2026, but limits actually increased that year.
Three metering systems exist: a rolling five-hour window, a weekly cap, and a premium-model meter.
Anthropic publishes usage estimates as ranges (bands), not exact numbers, causing opacity.

Hasty Briefsbeta