CLAUDE OPUS 5 — ARENA.AI PRELIMINARY SCORES (JULY 28, 2026)
● Frontend Code Arena: #1 at 1,725 Elo (Opus 5 Max) — Kimi K3 Max at 1,682
● Text Arena (factuality enabled): #1 at 1,512 Elo
● Design Arena: #2 at 1,358 Elo — matching GPT-5.6 Sol
● Status: Preliminary — scores are still accumulating votes
● Caveat: Arena.ai scores require significant vote volume to stabilise — early rankings can shift
● Price: $5/$25/M — same as Opus 4.8, half of Fable 5 at $10/$50/M
● What Arena measures: Crowd-sourced blind A/B votes on real tasks — not vendor-run benchmarks
What These Scores Mean and What to Watch For
Arena.ai (formerly LMSys Chatbot Arena) uses blind pairwise comparison — users see two model responses side by side without knowing which model generated each, then vote for the better one. The Elo score is derived from millions of these votes across real tasks. According to Arena.ai's own X post, Opus 5 was added to all Arena boards on July 25 with scores still accumulating. The preliminary 1,725 Frontend Code Elo is a strong early signal — but early Arena scores for newly added models can shift by 30-60 Elo points in either direction as vote volume grows. The scores are worth tracking, not treating as settled.
The Frontend Code Arena matters specifically because it captures frontend UI coding quality via blind human votes — not synthetic benchmarks. As Propel Code's analysis explains, the board measures front-end web development and agentic build tasks with real prompts, making it a useful complement to automated benchmarks like SWE-bench that target backend code changes. A model ranking #1 in Frontend Code Arena has demonstrated to actual developers — via their votes — that it produces better frontend code than its competitors. That is a different kind of evidence than a lab-run benchmark.
How Opus 5 Compares to What Came Before It
According to BenchLM's leaderboard history, the top Arena Elo score rose from 1,094 (vicuna-13b) in May 2023 to 1,501 (claude-opus-4-6-thinking) by mid-July 2026 — a gain of 407 points over 39 months. A preliminary 1,725 Frontend Code Elo for Opus 5 Max would be a significant jump from that baseline. The "Max" qualifier in the score matters: Arena often tracks different inference configurations separately, and Opus 5 Max likely represents the highest-reasoning configuration. As LocalAI Master notes, the July 2026 wave of frontier releases — Fable 5, Grok 4.5, GPT-5.6 family, Kimi K3, and now Opus 5 — means the board is in flux with all new entrants still accumulating votes.
What the Leaderboard Position Does Not Tell You
Arena scores measure what general users prefer in blind votes — they do not directly measure production coding accuracy on real GitHub issues (that is SWE-bench), factual accuracy under pressure (where Anthropic's own system card notes Opus 5 has a higher hallucination rate than Opus 4.8 on one factual benchmark), or cost-efficiency under agentic load. According to FelloAI's July model rankings, Fable 5 still leads on the boards that publish automated results — Terminal-Bench 2.1 at 83.8% in Claude Code, LiveBench Coding at 86.0%, and the Remote Labor Index. The Arena #1 for Opus 5 is a meaningful user preference signal. It does not displace Fable 5's automated coding benchmark lead.
The Pricing Signal — Why This Matters for Teams Still on Fable 5
The combination of preliminary Arena #1 and $5/M pricing creates a specific question for teams paying $10/M for Fable 5: if Opus 5 is preferred by users in blind votes on frontend coding tasks, is the Fable 5 premium still justified? The honest answer is it depends on your workload. For automated production code changes where SWE-bench Pro results matter, Fable 5's published 80.4% score remains the documented benchmark leader. For teams doing mixed frontend and UI work where human preference for code quality is the relevant measure, Opus 5's Arena #1 at half the price is a serious reason to run your own evaluation now.
Sources: Arena.ai X post (July 25) · FelloAI July model rankings · BenchLM leaderboard history · LocalAI Master Arena guide · Related: Claude Opus 5 full launch review · Opus 5 vs Fable 5 — when to use each