Pareta routes cheap LLM tasks to small models, 620x cheaper than GPT-5.5
D33B · reddit · 2026-09-15
- The team launched Pareta: an OpenAI-compatible endpoint where model="auto" routes each request to a small specialist model, escalating to a frontier model only when a check fails.
- On their own benchmark of 60 gold-labelled classification/extraction items (ticket intent, triage, PII, sentiment, moderation, language routing) vs GPT-5.5 with same prompts and schemas: 55/60 vs 60/60 accuracy, 620x cheaper, 1.6s vs 2.1s latency.
- Cost ratio across the six job types ranged from 308x to 1601x; the authors admit results are unknown for summarization, RAG, and agent loops.
- $30 free credit for new users; team is soliciting feedback.
More from coding & agent
- GPT-6 Astra + Hyper3D Rodin MCP Turns One Image Into a Full 3D Scene — heyshrutimishra · 2026-09-15
- RAG chunking can strip the context your answer needs: attach doc titles and section metadata — gethackteam · 2026-09-15
- Shop owner vibe-codes an accounting app: one HTML file, 100 test suites, and the rules that keep agents honest — anime_king_1 · 2026-09-15
- Pydantic AI claims 20% speed win over LangChain with similar accuracy in third-party agent benchmark — solyarisoftware · 2026-09-15
- Man runs agent to auto-post on his FB page every 30 min overnight, using just 1% of quota — TheMoonMidas · 2026-09-15
- Qwen Code ships CUA Driver v0.20.9 with signed macOS binaries and MCP payload filtering — github-actions[bot] · 2026-09-15