Install once.
nerfd connects to your coding tools.
You work as usual.
Independent benchmarks, built by the community
INSTALL NERFDFREE & OPEN SOURCE
Start contributing.
Connect your coding tools with one command.
nerfd records the results as you work.
curl -fsSL https://nerfd.org/install.sh | shShares session measurements. Your prompts and code stay local.
No account · Node.js 22.13+ · No lab funding
How it works
Install nerfd. Keep working.
Measure the session.
Results, corrections and cost,
measured on your machine.
Compare with others.
Share session measurements
to help compare models.
Your prompts and code stay on your machine. Sharing is optional.
Explore the measurements
How much help did it need?
Corrections, repeated requests and pushback reveal how much direction a model needed. nerfd reduces these to counts on your machine and compares rates per user turn, so a long session is not automatically worse.
Also records frustration markers, clarifying questions and abandonment signals. These are heuristic indicators; phrase detection is English-only.
Available signals vary by tool. Shared records use weekly rotating pseudonymous IDs. Privacy · Open method ↗
Gemini CLI
Qwen Code01 / What you can learn
See how models perform
on everyday tasks.
Lab tests are useful, but they cannot tell you everything about using a model. nerfd adds results from real coding work, including successful sessions, errors and repeated corrections.
Compare models by task.
Compare ten kinds of work, from debugging to design. Each result shows how many sessions it is based on.
Compare tasks ↗EXAMPLEMODEL AMODEL BMODEL CDebugging908568Review838980Refactoring877572
Track performance over time.
See whether a model’s results improve or decline. Compare its weekly scores with its earlier results and other models.
View model histories ↗See what worked.
See how people rated the work, how much of the code they kept, and how often they had to correct the model.
View session results ↗63 / 84SUCCESSFUL SESSIONS
ILLUSTRATIVE DATA
ILLUSTRATIVE DATA
02 / The community benchmark
The community’s results.
Weekly resultsIllustrative data
Model scores by week
Model AModel BModel C
Example scores over eight weeks. Visit a model page for its recorded results.
The community scorecard
ExampleCONTRIBUTED SESSIONS · LAST 4 WEEKS
01
Claude Opus 4.628 sessions88
Claude Opus 4.628 sessions8802
GPT-5.424 sessions81
GPT-5.424 sessions8103
Kimi K2.512 sessions74
Kimi K2.512 sessions74Results vary by task.↗
Loading rankings…
A public rank requires at least 10 sessions. Scores combine human ratings, retained code and sessions without errors or interruptions. Tasks and users differ, so compare similar work.
03 / Your personal report
Your own results,
in one report.
You also get a report of your own sessions: which models worked best, what your subscriptions provided, and where you spent time correcting mistakes.
Import your existing session history to get started. Your report works whether or not you share results with the community.
nerfd reportnerfd / PERSONAL REPORTEXAMPLE · 4 WEEKS
Your monthly summary
84sessions
62.4hours of work
63successful
Claude Max 5× · API-equivalent value3.4×
GPT-5.4 · code survival88%
Claude Opus 4.6 · your score88 / 100
Saved as an HTML file on your machine.
API-equivalent value is an estimate, not cash saved.
04 / How it works
From your coding tools
to the public rankings.
nerfd measures sessions on your machine. If sharing is on, it sends those measurements to the public dataset. The same published scoring method applies to every model.
Install once.
The installer connects to supported tools it finds. Start with existing history using nerfd backfill --since 90d.
Work as usual.
Session hooks collect tokens, timing and errors. Conversation signals become counts locally. nerfd check measures code survival later.
Share session results.
Sharing adds your session measurements to the rankings. Prompts, code, file paths and project names stay on your machine.
Compare the results.
Browse community rankings and model histories. Run nerfd report to see your own results and costs.
05 / Help build the benchmark
Add your results
to the benchmark.
Help others compare models using results from your work. Successful sessions and failed attempts are both useful.
Free and open source. No account. Node.js 22.13 or newer.
Sharing is your choice; your personal report works either way.
TERMINAL~/ your machine
curl -fsSL https://nerfd.org/install.sh | sh$ nerfd backfill --since 90d
$ nerfd report
Sharing is on. Session measurements contribute to the rankings. Turn it off with nerfd share off.
Copy the command, then run it in your terminal.
Common questions
Before you install.
How is my work kept private?
Sharing sends session measurements, including the model, tool, plan, task category, timing and results. It does not send prompts, code, file paths or project names. Shared records use an ID that changes weekly. They are pseudonymous, not guaranteed anonymous; combinations of metadata may still identify someone. Inspect the full privacy statement.
How is this different from a lab benchmark?
Lab benchmarks test models under controlled conditions. nerfd compares results from everyday coding sessions. This shows how models perform in practice, but tasks, tools and users vary. Check the task filters and session counts when comparing models.
Does it work with local models?
Yes, through supported tools such as OpenCode and Goose. Different hosts and model variants are tracked separately, including quantised versions run locally.
How do I remove it?
Run nerfd share off to stop sharing and nerfd init --remove to remove the hooks. Your local report remains a file you can keep.
Does it include good and bad results?
Yes. With sharing on, supported completed sessions are included whether they went well or badly. Errors, corrections and interruptions are recorded too. Automated sessions are excluded from quality rankings by default.
How can I help?
Contribute session metrics, report an issue, improve an adapter or review the scoring method. Start with the contribution guide.