HN Debrief

Cursor removed cost information from the usage page and CSV export

  • AI
  • Developer Tools
  • Startups
  • Economics

The post pointed out that Cursor had stopped showing cost information in places users relied on to understand usage, including the usage page and CSV export. A Cursor employee said the missing dollar values in CSV were an accidental bug introduced while removing an old feature flag and that export had already been fixed. The bigger change was intentional. Cursor removed a dollar-denominated usage graph because it showed included plan usage as if it were spend at API rates, and people misread that as what they were actually billed. The practical effect, according to users, is that the product now tells you less about the cost of a given model or session right when you need to make choices.

If your team buys AI tooling through a bundled IDE or subscription, insist on per-user cost visibility and exportable usage data before renewal. Opaque metering turns model selection and workflow tuning into guesswork, which quickly becomes a finance and vendor-trust issue.

Discussion mood

Strongly negative. People saw the change as deliberately reducing cost transparency, then took the employee explanation as confirmation that the most useful cost view was removed on purpose. Underneath the anger was a broader loss of trust in Cursor’s pricing, product direction, and fit as a long-term team vendor.

Key insights

  1. 01

    Admin-only billing breaks local decision making

    Admin-only visibility turns cost from a developer feedback loop into a management report. If the person choosing between models or workflows cannot see their own usage, they cannot optimize a session, compare runs, or learn which tasks are worth handing to the agent in the first place.

    Treat per-user usage visibility as a product requirement, not a nice-to-have. If only admins can see spend, expect waste to rise because the people creating it cannot see it in time to change behavior.

      Attribution:
    • EugeneOZ #1
    • jonjohnsen #1
  2. 02

    Harness design drives token cost as much as the model

    The model is only part of the bill. Prompt scaffolding, tool catalogs, memory systems, and compaction behavior can multiply token use by an order of magnitude on the same task. The most useful framing here is not that one harness is universally better. It is that teams need to inspect traces, prompts, and cache behavior because a lot of spend comes from the wrapper around the model, not the model itself.

    Audit your agent harness before blaming model pricing. Instrument prompts, tool injections, and cache hits so you can cut waste at the orchestration layer instead of only shopping for cheaper models.

  3. 03

    Cursor still wins on IDE polish

    Even many frustrated users still gave Cursor real credit for being a better editor experience than most extensions and terminal-agent setups. Cursor Tab, agent UI, and remote SSH workflows were cited as genuinely better integrated. That matters because the replacement stack often splits coding, reviewing, and agent control across several tools.

    Do not assume a raw switch to terminal agents preserves developer throughput. If you are evaluating alternatives, measure edit-review flow and remote workflows, not just model quality or subscription price.

      Attribution:
    • redox99 #1
    • RussianCow #1
    • tzone #1
    • PufPufPuf #1
    • teaearlgraycold #1
  4. 04

    Opaque usage turns into a vendor risk issue

    Once a vendor sits between your team and multiple model providers, usage transparency is no longer just about avoiding surprise bills. It is how you verify that the broker is adding value instead of markup, and whether your engineers are using the tool efficiently. Removing that visibility makes it harder to justify renewals and easier for finance and security concerns to compound.

    If you buy through an AI platform rather than direct APIs, require exportable data that lets you reconcile usage, pricing, and outcomes. Without that, you cannot tell whether the vendor is saving time or just hiding costs.

      Attribution:
    • reilly3000 #1
    • bogometer #1
    • xp84 #1

Against the grain

  1. 01

    Dollar graphs were genuinely misleading on plans

    Showing included plan usage as dollars blurred the line between reference pricing and actual billing. If users interpreted API-rate equivalents as real spend, removing that graph can be defended as cleaning up a bad metric rather than hiding charges.

    Separate three numbers in your own tooling and procurement reviews: actual billed spend, included-plan consumption, and API-rate equivalent. Mixing them in one graph creates avoidable confusion.

      Attribution:
    • jonjohnsen #1
  2. 02

    Cheap short tasks can hide long-session tradeoffs

    The dramatic token savings from minimalist harnesses were measured on simple tasks with known outputs. Several people cautioned that this can overstate the benefit if longer sessions need compaction, richer memory, or broader tool support to stay effective over hours or days. Efficiency claims are real, but they need task shapes that match production use.

    Benchmark agent setups on the kinds of sessions your team actually runs, especially long-lived ones. A harness that dominates on short chores may lose its edge when context management and recovery matter.

      Attribution:
    • applfanboysbgon #1 #2
    • zuzululu #1
    • tosh #1 #2
    • xienze #1
  3. 03

    You may not need a special AI IDE

    A meaningful slice of people have already rebuilt the workflow with Neovim, Zed, VS Code, terminal agents, and standard diff tools. Their point is not that Cursor is bad. It is that once code generation shifts into external agents, the editor itself becomes less differentiated than it looked a year ago.

    Test a lightweight stack before committing to an AI IDE renewal. If your team mainly prompts, reviews diffs, and applies fixes, the premium IDE layer may no longer be where the value sits.

      Attribution:
    • leonvoss #1
    • urbsgpw #1
    • PufPufPuf #1
    • throw1234567891 #1

In plain english

API ↩
Application Programming Interface, a way for software to call another service or model programmatically.
Claude Code ↩
Anthropic's command-line coding agent product that can read, edit, and run code-related tasks.
Codex ↩
OpenAI's coding agent and related tooling for writing and modifying code.
compaction ↩
A process where an agent summarizes or compresses prior context so it can keep working within the model's context limit.
CSV ↩
Comma-separated values, a plain text file format commonly used for exporting tabular data into spreadsheets or other tools.
harness ↩
The software layer around a model that adds prompts, tools, memory, routing, and other behavior for a specific workflow.
IDE ↩
Integrated Development Environment, a code editor bundled with features like debugging, file navigation, and source control tools.
Neovim ↩
A highly configurable terminal-based text editor popular with programmers.
SSH ↩
Secure Shell, a protocol used to connect to and work on remote machines from a local computer.
token ↩
A chunk of text a language model processes for input and output, which is commonly used as the unit for pricing and context limits.
VS Code ↩
Visual Studio Code, Microsoft's widely used code editor.
Zed ↩
A modern code editor focused on speed and collaboration.

Reference links

Primary source and vendor responses

Token usage and harness efficiency