Fable 5 Enters Arena Search at #3 — Claude Opus 4.6 Holds the Summit at 1252 Elo
Arena added Claude Fable 5 and Claude Opus 4.8 to its Search leaderboard on June 15. The initial standings reveal something counterintuitive: Fable 5, the highest-scoring model on the Artificial Analysis Intelligence Index at 64.9 and the first to clear 95% on SWE-bench Verified, is not leading the search rankings. Two older models are.
Where the Models Landed
| Rank | Model | Elo | Battles |
|---|---|---|---|
| 1 | Claude Opus 4.6 Search | 1252 ± 5 | 82,294 |
| 2 | GPT-5.5 Search | 1240 ± 6 | 37,025 |
| 3 | Claude Fable 5 | 1237 ± 11 | 5,007 |
| 4 | Claude Opus 4.7 | 1232 ± 6 | 38,472 |
| 5 | ERNIE 5.1 | 1226 ± 10 | 3,824 |
| 11 | Claude Opus 4.8 | 1203 ± 6 | 18,408 |
Fable 5 enters at 1237 Elo with a ±11 confidence interval — more than twice the spread on Opus 4.6’s settled 1252 ± 5. At 5,007 battles, it has covered roughly 6% of the ground Opus 4.6 has. The score will move.
Claude Opus 4.8 enters at a more stable 1203 Elo with 18,408 battles, which puts it at rank 11 — below the Google Gemini grounding models and the GPT-5.2 Search variant.
What the Gap Means
The Search leaderboard measures user preference on grounded, citation-backed queries — a different evaluation surface than the Intelligence Index or SWE-bench. High reasoning scores do not translate linearly. Arena’s coding, text, and document leaderboards show Fable 5 at or near the top; search evidently rewards a different response style.
Opus 4.6 has held the #1 search position for months across 82,000+ battles. Its lead over the next Anthropic model (Fable 5) is 15 Elo points — outside Fable 5’s current confidence interval. That gap may close as votes accumulate, but it also may not. Opus 4.8, with four times more battles than Fable 5 and a more settled interval, currently sits 34 points behind Opus 4.6.
GPT-5.5 Search at 1240 Elo is the only non-Anthropic model in the top four. Google’s grounding-enabled Gemini models (3.1 Pro at 1214, 3 Pro at 1207) start at rank 7 despite native web access being a structural advantage.
Anthropic Holds the Breadth
Four Anthropic models occupy the top six search slots: Opus 4.6 (#1), Fable 5 (#3), Opus 4.7 (#4), and Sonnet 4.6 (#6 at 1218). OpenAI holds #2 with GPT-5.5. xAI’s Grok 4.20 multi-agent beta enters at #9 (1206 Elo).
ERNIE 5.1 at #5 (1226 Elo) is notable — Baidu’s model outranks both Gemini 3.1 Pro Grounding and GPT-5.4 Search on a leaderboard where native web infrastructure might be expected to matter.
Pricing at the Top
Claude Opus 4.6 Search runs at $5 input / $25 output per million tokens. Claude Fable 5 is $10/$50 — double the cost, three Elo points behind. GPT-5.5 Search sits at $5/$30. The most affordable in the top 20: Grok 4-1 Fast Search at $0.20/$0.50.
Fable 5 will need additional battle volume before its true search ceiling is clear. The fact that it entered below its older siblings — with a wider confidence interval — is expected at this stage, not a definitive verdict.