Parallel’s launch covers 24 popular open- and closed-source models. It defines Search Intelligence Score as accuracy across three existing public benchmark suites, while its Search Efficiency ranking orders models by cost among those at or above the median accuracy score. The company says GPT-6 (Astra) posts the highest Search Intelligence Score, 70.3, at $301 per 1,000 tasks; it lists GPT-5.6 Luna as the efficiency leader at 55.0 and $23 per 1,000 tasks. It also reports that DeepSeek V4.1 Flash gains 45 points with search.
The new leaderboard makes web-search competence a separately visible model-selection variable rather than a hidden assumption inside general capability rankings. Search performance is composite: a model must decide when more evidence is required, formulate useful queries, select and inspect relevant URLs, and synthesize supported answers. Holding the search layer constant is meant to isolate the model’s contribution to those choices; separating accuracy from cost then turns deployment into a threshold-and-budget decision.
This matters now because agents increasingly use web research as an operational dependency, while model reports and search-provider comparisons do not consistently expose this interaction. Teams can use the leaderboard to form hypotheses about a candidate stack, then validate them against their own task mix, latency limits, sources, and error tolerance. The reported cost figures are especially useful as directional deployment inputs, not as a universal bill of materials.
The evidence is limited to Parallel’s published methodology, selected public benchmark suites, and its September 15, 2026 snapshot; it does not independently establish performance on every production workflow or prove that the ranking will persist as models, prices, prompts, and search behavior change.
A first implication is that teams building search-heavy agents should benchmark the model-and-search combination, not select a model solely from general-purpose leaderboards. A second implication is that an accuracy target can be a more useful procurement constraint than an abstract preference for the top score.
Watch whether Parallel publishes task-level methodology, benchmark composition, reproducible runs, and updates that show how rankings move as model versions and pricing change. Also watch whether other evaluators adopt comparable model-plus-search measurements, which would test whether this framing generalizes beyond one provider’s harness.
This begins a distinct analytical record because no prior New Runtime record was supplied for this specific public leaderboard launch and its model-plus-search evaluation framing.