Publikováno 2026-09-05 · 林启明
Přímá odpověď
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. Tento průvodce je určen pro produktové a platformní týmy, které porovnávají kvalitu modelů, cenu, politiku routování a riziko rolloutů.
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.