DRACO is Perplexity's ten domain deep research benchmark. OpenSquilla 0.5.0 routes four mid tier Chinese models in parallel and averaged about a third of Fable 5's per task cost on DuckDuckGo.
The cheapest way to top Perplexity AI's DRACO deep-research leaderboard today is to skip the smartest model entirely. OpenSquilla 0.5.0 Preview 1, released July 3, 2026, ran four mid-tier Chinese models in parallel and beat GPT-5.5, Opus 4.8, and Anthropic's reported Fable 5 flagship. The routing code, not the weights, is the unit that won.
DRACO is a public, expert-curated deep-research benchmark covering ten domains with roughly forty grading criteria per task. Per the project's Hugging Face dataset card, OpenSquilla's harness averaged 64.09 on the Brave Search track, edging OpenAI's reported GPT-5.5 at 53.28 by more than twenty percent and Anthropic's Opus 4.8 at 59.11 by roughly eight percent. On the DuckDuckGo track, the harness averaged 60.85 against Anthropic's reported Fable 5 at 59.80, nearly tied on quality at roughly a third of the cost.
The cost column is where the story tilts. Qbitai's coverage lists reported average task cost at $0.12 on Brave and $0.39 on DuckDuckGo, against $1.50-ish for an Opus-class call and $1.21 for the reported Fable 5. The harness is not subsidizing a free-tier API. It is buying four mid-tier Chinese model tokens, paying the traffic, and assembling the answer inside the routing code itself.
The mechanism, documented in the project's GitHub repository and previewed in the companion technical report "Agentic Routing: The Harness-Native Data Flywheel," is diversity sampling plus consensus aggregation. OpenSquilla 0.5.0 organizes DeepSeek v4, GLM-5.2, Kimi K2.7, and Qwen3.7 as parallel proposers, each independently running the search and reasoning loop. A designated aggregator model then blends the four outputs into one final answer. Per the project's own framing, no overseas flagship sits inside the ensemble; the four models are interchangeable, named components of a single distributed system, with the harness deciding which one drafts which sub-question and how their disagreements are resolved.
Deep-research tasks are the highest-leverage workloads for AI agents because they require long evidence chains, contradiction handling, and verified citations, exactly the failure modes where single models still bleed points. Perplexity's DRACO research write-up describes the rubric weighting that makes the test hard to game with shallow retrieval. If a four-model Chinese ensemble can post the lowest cost and highest score on that rubric, the gap between "best model" and "best system" is open for the first time at production scale.
Three caveats matter before any buyer wires this up. First, all DRACO scores above are reported by OpenSquilla via qbitai; independent leaderboard pulls, and any rebuild on the public DRACO site, will be the deciding evidence. Second, the named comparisons (Opus 4.8, GPT-5.5, Fable 5) are reported model labels rather than independently verified product names, and Fable 5's Brave-search run was still in progress at publication. Third, the company is reportedly valued around $100 million after a first funding round, a financial signal that tells you more about venture taste than benchmark architecture, and it should be read as "public reports," not balance-sheet truth.
What the release does document, on the vendor side, is a deliberate version cadence. Per qbitai, OpenSquilla's roadmap has emphasized cost discipline and verifiable delivery since v0.1.0: smart routing by task difficulty, one-click migration from other agent frameworks, a self-organizing skill protocol, and red-green regression evidence for code edits. The 0.5.0 release, published on the GitHub releases page, is the first one where the architecture itself is the product story. TokenRhythm (基元律动), the developer behind the harness, is leaning into the claim that harness-native routing is the durable moat, not base-model capability.
The next DRACO leaderboard refresh is the real test: if the $0.12 Chinese ensemble keeps its top score, the edge in deep research moves from weights to wiring, and TokenRhythm's bet that harness-native routing is the durable moat pays off.