Publié le 2026-09-05 · 林启明
Réponse directe
OpenAI's MRCR v2 results put GPT-6 Astra at 96.3% needle retrieval past 512K tokens. What that means for RAG-vs-long-context decisions — and what it does not cover. Ce guide s’adresse aux équipes produit et plateforme qui comparent la qualité des modèles, le coût, la politique de routage et le risque de déploiement.
The published numbers
On OpenAI's MRCR v2 8-needle evaluation, Astra scores 100% in the 256K-512K range and 96.3% at 512K-1M, where the predecessor Sol drops to double-digit losses. Multi-needle retrieval — finding several distinct facts buried in a huge context — is the benchmark closest to real document-analysis work.
What retrieval scores do not measure
Needle benchmarks test finding, not reasoning across findings. Synthesizing a contract position across eight documents is harder than locating eight clauses, and no published MRCR variant scores that. Long-context recall also does not fix stale or wrong source material — a 1M window of bad context produces confident, well-read nonsense.
A decision rule for RAG vs long context
Most production systems end up hybrid: retrieval to nominate candidates, a long-context model to read them together. The catalog tags Astra for exactly this long-context and agent mix.
- Knowledge changes slowly and fits under ~250K tokens: stuff the context, and revisit the tier boundary before you ship.
- Knowledge changes fast or exceeds the window: retrieval still wins, because re-indexing beats re-billing a 1M-token prompt.
- Needle-style extraction over huge stable corpora: this is Astra's published sweet spot — 96.3% past 512K is the headline capability.
- Tight unit economics per request: retrieval. A 5K-token retrieved prompt lists at pennies; a 1M-token prompt approaches the $20 high-tier input rate.
What is unverified
MRCR and OSWorld figures are OpenAI's published evaluations. First-party long-context quality measurement for this model does not exist yet in the NextModel bench table.
FAQ
Should I delete my RAG pipeline because of 1M context?
Rarely. Use long context where sources are stable and co-reading matters; keep retrieval where knowledge churns or per-request cost dominates. Hybrid is the common answer.
How reliable is retrieval at the full 1M tokens?
OpenAI publishes 96.3% on the 8-needle MRCR v2 benchmark at 512K-1M. That is their evaluation, not a guarantee for your document shape — sample it with your own needles.