Stop burning cash on bloated prompts
Paste a prompt. Token-Trim audits the waste, finds your prompt-caching savings, and rewrites it — while scoring output-quality risk so you never silently degrade results.
No signup for your first 3 audits.
Find the waste
Deterministic token counting plus rules that flag filler, duplicated instructions, and bloated personas — with exact tokens saved per fix.
Cache the static parts
We detect the stable prefix of your prompt and tell you where to set a cache breakpoint. Cached reads cost ~10% on Anthropic — usually the biggest win.
Don't degrade output
Every cut gets a quality-risk score. Mechanical cuts auto-apply; risky ones are flagged for review. Pro runs a real original-vs-compressed output delta.
Audit a prompt now
Free, in-browser, no account. Results in a second.
Paste your prompt or system message
Your audit appears here: token count, monthly cost per model, bloat findings, a caching recommendation, and a safe rewrite — with a quality-risk score.
Simple, revenue-honest pricing
Free to try. Pay only when token savings pay for it.
Free
Kick the tires
Free forever
- 3 audits / month
- Token + cost breakdown
- Bloat findings & safe rewrite
- Caching recommendation
- Quality-risk score
Pro
For shipping devs
≈ $12/mo
- Unlimited audits
- Savings dashboard
- Real output quality-delta
- Per-model cost projections
- Priority rules updates
Team
For SaaS teams
≈ $30/mo
- Everything in Pro
- Up to 5 team seats
- Shared prompt library
- Per-model cost projections
- GST invoice
Prices in INR (GST-inclusive invoicing for Team). USD shown for reference. Cancel anytime — see our refund policy.
Get a weekly prompt-cost tip
One concrete way to cut your Claude/GPT bill, every week. No spam.
FAQ
How is this different from just counting tokens?
Counters tell you the bill. Token-Trim tells you which specific lines are wasted, where to add prompt caching, and — critically — whether cutting them changes your output. It measures quality risk before recommending a cut.
What's the 'compression illusion'?
Aggressive prompt compression often looks fine on a glance but silently degrades results on edge cases. We separate mechanical, lossless cuts (filler, duplicates, whitespace) from content cuts that need review, and Pro runs a real output comparison.
Do you store my prompts?
No. Prompts are analyzed in-memory to produce the report and are not persisted. We store only anonymous usage counts and, if you opt in, your email.
Which models are supported?
Token counts are exact for OpenAI o200k models (GPT-4o/4.1/o-series) and approximate (≈) for Claude, since Anthropic ships no public tokenizer. Cost projections cover Claude Opus/Sonnet/Haiku and GPT-4o/4o-mini/4.1.