Arena Launches Coding Index: GPT-6.2 Astra Ranks #3 with 62% Lower Cost
arena · x · 2026-11-22
Arena introduced its Coding Index, measuring which agents perform best at real coding work, alongside the new Agent Arena Index for overall agent performance built from millions of real-world task traces.- GPT-6.2 Astra (Max) scored 1338 on the Coding Index (#3) and 1291 overall (#6), OpenAI's best model; it gains +50 points over GPT-6 Astra while cutting median per-task cost from $12.55 to $4.75 (62% lower), landing on the cost Pareto frontier- GPT-6.2 Sol (Max) ranked #11 at 1241 points, also on the cost frontier- Overall cost dropped 40% vs the previous variant ($2.94 vs $4.91 per task); user satisfaction (Praise vs Complaint) is its strongest signal at #5- Results are based on 7.1K+ real agentic sessionsThe Coding Index leaderboard image shows Claude Fable 5.2 (1489) and Claude Opus 5.5 (1373) in the top two spots.
More from Models
- Julia 1.5 and Julia 2 launch: two models built for decision-making in software — Scobleizer · 2026-11-22
- LLMs increasingly write non-human objects as agents — an SVO-compression tell — Rioting-Flamingo · 2026-11-22
- Intel's 350W D270 claims 5.9x lead over RTX PRO 6000 — but the fine print reveals benchmark games — StefanoGogioso · 2026-11-22
- rakyll: Neon Is My Favorite Model Yet — Fast, Capable, Sounds Human — rakyll · 2026-11-22
- Qwen4 Max 30% off on Together AI: fixed Tornado bug for $0.69 in one request — togethercompute · 2026-11-22
- LlamaIndex's ParseBench-2 ranks 59 document parsers; KDL tops open-weight at 78.89 — victormustar · 2026-11-22