Claude Opus 5.5's Best Score Is a Tie. Its Real Win Is Elsewhere
Claude Opus 5.5’s Best Score Is a Tie. Its Real Win Is Elsewhere
Claude Opus 5.5 scores 9.93 on structured output extraction [^1], its highest quality figure in this material, yet on that task the score marks a tie at the ceiling rather than a lead. Its clearest win is elsewhere: matching profiles to a pool, where it was judged best on 18 of 26 occasions [^2]. Telling those two kinds of result apart shows where Opus 5.5 earns a route on quality and where cost and latency should decide instead.
Two kinds of result
The benchmark reports Opus 5.5’s work in two ways. A quality score rates the model’s answer out of 10. A best-answer share is the fraction of judged occasions on which the model’s answer was picked as best among the models competing on that occasion.
The two can tell different stories. A high quality score shows the model does the work well, but it may not show that the model does it better than anyone else. The first two examples below are exactly that case.
The extraction score sits at the ceiling
On structured output extraction, Opus 5.5’s 9.93 is fractionally above Claude Sonnet 5.5 at 9.88 ±0.16 [^1]. That margin should not be read as an advantage. Across 119 judged occasions on this task, the top two answers scored exactly the same 54.6% of the time, with a median gap of 0.0 [^1].
The task separates models poorly. The correct reading is that Opus 5.5 is at the ceiling on pulling structured fields from text, not ahead of the field. This is also the only direct comparison between Opus 5.5 and Sonnet 5.5 in the material.
The social-post score is a four-way tie
On selecting a portfolio of social posts, Opus 5.5 scores 8.6 ±0.21 [^3]. Three other models land in the same band: Tencent Hy4 Preview at 8.65 ±0.16, GLM-5.3 at 8.6 ±0.17 and Grok 4.7 at 8.49 ±0.22 [^3]. Their intervals overlap, so the quality evidence cannot separate these four models.
This is the only Opus 5.5 quality figure in the material that carries a stated interval, and that interval is what makes the tie readable at all. Its other figures, including the 9.93 on extraction, are reported without intervals, so their sensitivity to measurement noise cannot be assessed.
The real lead: matching profiles to a pool
Matching profiles to a pool is different. Opus 5.5 was judged best on 18 of 26 occasions, or 69.2% [^2]. That is a genuine head-to-head win on this work, not a score pressed against the top of a scale.
Promotional campaign writing shows what a weaker preference looks like. There Opus 5.5 was judged best on 8 of 25 occasions, or 32.0% [^4], a real share but less than a third of judged occasions. Both shares describe preference among the models competing on those occasions and carry over to nothing else [^2][^4].
What to do with it
Send profile matching to Opus 5.5: on this evidence, its 69.2% best-answer share [^2] is the clearest quality win in its profile. Treat promotional campaign writing as a competitive but not leading route, since 32.0% [^4] is a share of preference rather than dominance.
Structured output extraction can go to Opus 5.5 too, but price the decision accordingly. With the top two answers tied on 54.6% of occasions [^1], a cheaper model may deliver the same result. On the social-post portfolio, the four-way tie [^3] makes the choice a matter of cost and latency rather than quality.
The broader habit is simple. Before treating a top score as a reason to route, check whether the task separates models at all, through a stated interval or a tie rate, and give more weight to head-to-head preference where it exists.
The live benchmark at https://fronset.ai/benchmark/ lets you set the quality bar your workload actually needs. Check the numbers there before committing a route.