v0.3 — 162 specs across 5 formats — 10.3x overall compression (4.37M → 423K tokens)
| API | Raw Tokens | Lean Tokens | Compression |
|---|---|---|---|
| notion | 68,587 | 1,733 | 39.6x |
| snyk | 201,205 | 5,193 | 38.7x |
| zoom | 848,983 | 22,474 | 37.8x |
| digitalocean | 345,401 | 12,896 | 26.8x |
| plaid | 304,530 | 18,930 | 16.1x |
| box | 232,848 | 17,559 | 13.3x |
| asana | 97,427 | 8,257 | 11.8x |
| hetzner | 167,308 | 14,242 | 11.7x |
| vercel | 136,159 | 13,472 | 10.1x |
| slack2 | 126,517 | 13,316 | 9.5x |
| linode | 203,653 | 21,795 | 9.3x |
| stripe-full | 1,034,829 | 118,758 | 8.7x |
| launchdarkly | 31,522 | 4,731 | 6.7x |
| 61,043 | 9,474 | 6.4x | |
| netlify | 20,142 | 3,127 | 6.4x |
| gitlab | 88,242 | 15,092 | 5.8x |
| resend | 21,890 | 3,776 | 5.8x |
| vonage | 1,889 | 412 | 4.6x |
| github-core | 2,190 | 531 | 4.1x |
| circleci | 5,725 | 1,464 | 3.9x |
| stripe-charges | 1,892 | 490 | 3.9x |
| petstore | 4,656 | 1,217 | 3.8x |
| twilio-core | 2,465 | 688 | 3.6x |
| openai-core | 1,730 | 524 | 3.3x |
| google-maps | 941 | 316 | 3.0x |
| discord | 909 | 308 | 3.0x |
| cloudflare | 763 | 265 | 2.9x |
| spotify | 826 | 331 | 2.5x |
| sendgrid | 518 | 209 | 2.5x |
| slack | 762 | 316 | 2.4x |
Median: 5.2x — Large, verbose specs see the biggest gains (up to 39.6x). Small specs with minimal redundancy still achieve 2–3x.
| Format | Specs | Median Compression |
|---|---|---|
| OpenAPI | 30 | 5.2x |
| Postman | 36 | 4.1x |
| Protobuf | 35 | 1.5x |
| AsyncAPI | 31 | 1.4x |
| GraphQL | 30 | 1.3x |
LAP delivers the highest compression on verbose formats (OpenAPI, Postman) where JSON/YAML schemas carry significant structural redundancy. Compact formats like Protobuf and GraphQL are already information-dense, so compression is modest — but still meaningful at scale.
Real-world validation: five scenarios where an LLM agent generates working API integration code. The agent produces identical output regardless of whether it receives raw OpenAPI or LAP — it just costs fewer tokens.
| Scenario | OpenAPI Tokens | LAP Tokens | Savings |
|---|---|---|---|
| Stripe | 3,166 | 1,736 | 45% |
| GitHub | 3,384 | 1,743 | 49% |
| Twilio | 3,623 | 1,857 | 49% |
| Slack | 2,163 | 1,639 | 24% |
| Hetzner | 168,388 | 15,045 | 91% |
Hetzner's 91% reduction is notable — the raw spec barely fits in most context windows, while the LAP version fits comfortably.
With 10.3x average compression, a 100K-token API spec drops to ~10K tokens. At current model pricing, that translates to ~90% cost reduction per API call that includes spec context.
For teams making hundreds of agent calls per day against large API specs, this compounds to significant savings — potentially thousands of dollars per month.
-
Verbose specs benefit most. Notion (39.6x), Snyk (38.7x), and Zoom (37.8x) have massive schemas with repetitive patterns that LAP compresses aggressively.
-
Small specs still compress. Even minimal specs like Slack (762 tokens) achieve 2.4x — the floor is meaningful, not zero.
-
Format matters. OpenAPI and Postman carry structural overhead that compresses well. GraphQL and Protobuf are already concise — LAP still helps, but expect 1.3–1.5x rather than 5–10x.
-
Agent output is preserved. Implementation tests confirm that agents produce identical working code with LAP input — the compression is lossless for practical purposes.
-
v0.3 trades ~4% size for reliability. Inline schemas and completeness signals add a small number of tokens but eliminate common agent failure modes (broken
$refresolution, missing required fields).
- Token counting: cl100k_base tokenizer (GPT-4/Claude compatible)
- Specs: 162 real-world API specifications sourced from public repositories and official documentation
- Formats: OpenAPI 3.x, Postman Collections, Protocol Buffers, AsyncAPI, GraphQL SDL
- Implementation test: Claude Sonnet generates integration code from both raw and LAP specs; output correctness verified manually
- Compression ratio: Raw tokens ÷ Lean tokens
- All results reproducible from the benchmark suite in this repository