# 斩杀线 · Kill Line
Base URL: https://www.heigeai.com/llmkill
Read-only AI model comparison, with 10 task scenarios.
GET https://www.heigeai.com/llmkill/api/health for scheduler state and latest price snapshot.
GET https://www.heigeai.com/llmkill/api/models?scenario=overall for model IDs, prices, scores, source timestamps and mapping evidence.
GET https://www.heigeai.com/llmkill/api/scenarios for scenarios and score coverage.
GET https://www.heigeai.com/llmkill/api/analyze?model_id=openai%2Fgpt-6.1-sol&scenario=coding&input_tokens=4000&output_tokens=1500&cache_ratio=0.3&min_win_rate=0.45&min_saving=0.3&open_only=0
GET https://www.heigeai.com/llmkill/api/events for observed changes.
GET https://www.heigeai.com/llmkill/api/track with the same task parameters for its changes.
GET https://www.heigeai.com/llmkill/feed.xml with the same task parameters for a personal RSS feed; omit parameters for the global feed.
Resolve the feed_url returned by api/track against https://www.heigeai.com.
Prices: models.dev (MIT) listing of OpenRouter routes, USD per million tokens. OpenRouter confirms availability only.
Scores: Arena leaderboard dataset (lmarena-ai/leaderboard-dataset), CC BY 4.0. Mapping to catalogue IDs by 斩杀线.
Supplementary evaluation_references include a benchmark version, rank scope, date and source URL. Vals Index and Epoch Capabilities Index refresh automatically. Epoch AI data is CC BY 4.0, with ranks calculated across the full source snapshot including historical/configuration rows. AA lives only in evaluation_history as a dated historical snapshot and requires no periodic manual review. Never treat these as Arena Elo or use them in the replacement formula.
official_release.official_scores contains vendor-reported results for that exact revision, with official announcement and model-card links.
GET https://www.heigeai.com/llmkill/monitor/status.json for the independent server monitor heartbeat, and https://www.heigeai.com/llmkill/api/monitor-health for the second server's observer heartbeat.
Compare only the same scenario and benchmark version. min_win_rate is estimated blind-test preference, not task accuracy.
Use mapping_type and effort to disclose alias, reasoning-effort and date differences. Missing scores are scenario-specific. Check scored_scenarios; never reuse an older model revision score.
GET https://www.heigeai.com/llmkill/api/versions?model_id=PROVIDER/ID for separate canonical versions and their recorded evidence.
Costs are estimates for input/output/cache tokens, displayed per 1000 tasks. Reasoning tokens can raise actual bills.
Check meta.price_updated_at, benchmark_publish_date, sources and refresh_status before describing results as current.
Validate candidates on the user's real tasks before replacing a production model.