CoChat Deep Research Mode Beats Every Frontier Model on DRACO

TL;DR. Today we’re launching Deep Research Mode, a way for scientists, researchers, and anyone who lives in sources to run rigorous, fully-cited investigations on any topic. Under the hood it pairs CoChat’s citation-and-evidence harness with multi-model synthesis, and that combination lifts quality across every axis of Perplexity’s DRACO benchmark. The result: CoChat’s Deep Research Mode posts a normalized DRACO composite of 71.1, ahead of every frontier model and every fusion configuration published to date.

What the DRACO benchmark tests

DRACO (Perplexity AI, February 2026) is the most serious public test of deep-research systems we’ve found (paper on arXiv). It runs 100 real-world research tasks across 10 domains, each graded against an expert-written rubric with 3,934 weighted criteria in total, spanning factual accuracy (52% of total weight), breadth and depth, presentation, and citation quality.

The part that makes it hard to game: 415 criteria carry negative weights. Confidently stating something wrong costs you points, with penalties reaching −500 for dangerous medical advice. You cannot verbosity your way to a good DRACO score, padding a report with unsupported claims actively lowers it. This is where CoChat’s harness earns its keep: it caught 82.6% of those negative-weight traps overall, and 95.9% on factual-accuracy criteria and 98.1% on breadth-and-depth which is the axes DRACO penalizes hardest.

Our DRACO benchmark results

CoChat’s Deep Research Mode leads the field on the normalized DRACO benchmark composite (0–100 scale):

Scores for rows 2 and 4–11 are drawn from OpenRouter’s Fusion Beats Frontier writeup.

† Perplexity results: https://arxiv.org/html/2602.11685. Perplexity used its production Deep Research endpoint with Opus 4.6 or Opus 4.5 as the base model. Gemini-3-Pro judged all 100 tasks over 5 independent grading runs. The reported normalized scores were 70.5% (SD 0.3) and 67.2% (SD 0.3), respectively. These results used a different judge and evaluation harness, so they are not directly comparable to the CoChat or OpenRouter scores.
‡ Fable 5 was graded on only 93 of 100 tasks, the remaining 7 were refused by its content filters (per OpenRouter).

DRACO scores by domain

CoChat’s strongest showings are in the domains where source discipline matters most, Law and Academic research, with Medicine close behind despite DRACO’s brutal penalties for unsafe claims.

By axis: Where the score comes from

The DRACO benchmark composite rolls up four scoring axes. CoChat’s strongest axis is citation quality (78.5), fitting, since verified, source-backed citations are the product’s whole reason for existing, followed by presentation quality (73.5), factual accuracy (69.9), and breadth and depth (69). Every axis clears the field’s midline, which is what the TL;DR means by lifting quality across the board rather than on one headline metric.

Trap-catch rate = share of the axis’s negative-weight (“trap”) criteria CoChat correctly avoided [2].

How CoChat’s Deep Research Mode works

One request fans out into three layers of independent investigation and comes back as a single cited report. Nothing here is benchmark-specific, this is the production pipeline every CoChat user gets.

What the Deep Research flag actually toggles. The citation-and-evidence engine is always on, for every user, on every research request, flag or no flag. Decomposition, the parallel slice-researchers, the evidence pool, the CrossRef-verified citation registry, and the skeptical verification pass all run in standard mode too. The Deep Research flag adds exactly two things on top: the blind independent pass (step 4) and the reconciliation round (step 5). With the flag off, the pipeline is byte-identical to standard mode — you still get verified citations and evidence-backed claims; you just don’t get the second, independent investigation merged in.

flowchart TD
    Q([Research question])
    Q --> P["1 - Plan & decompose<br/><i>Opus 4.8 lead</i>"]
    P --> R1["Researcher<br/><i>Gemini 3.5 Flash</i>"]
    P --> R2["Researcher<br/><i>Gemini 3.5 Flash</i>"]
    P --> R3["Researcher<br/><i>Gemini 3.5 Flash</i>"]
    R1 --> POOL[("2 - Evidence pool +<br/>Citation registry<br/><i>CrossRef-verified</i>")]
    R2 --> POOL
    R3 --> POOL
    POOL --> D["3 - Draft report<br/><i>Opus 4.8</i>"]
    Q -. "blind, parallel" .-> IP["4 - Independent pass<br/><i>GPT-5.5 full investigation</i>"]
    D --> REC{"5 - Reconcile<br/>consensus - conflicts<br/>unique - blind spots"}
    IP --> REC
    REC --> V["6 - Skeptical verification<br/><i>claim-by-claim vs sources</i>"]
    V --> OUT([Cited report])

    style Q fill:#0e7490,color:#fff
    style OUT fill:#0e7490,color:#fff
    style POOL fill:#f1f5f9
    style REC fill:#fef3c7
    style IP fill:#fef3c7

Amber steps (4–5) are what the Deep Research flag adds. Everything else like plan, research, evidence pool, citation registry, verification runs for every CoChat user in standard mode too.

Step 1: Plan and decompose

The lead agent (Claude Opus 4.8) breaks the question into sub-domains and lays out a task plan. A question about, say, staggered-adoption econometrics becomes parallel slices: the foundational sources, the estimator variants, the adoption data, the critiques.

Step 2: Parallel research sub-agents

Each slice goes to a researcher sub-agent (Gemini 3.5 Flash) equipped with web search, page fetching, and academic search across CrossRef, Semantic Scholar, and arXiv. Sub-agents don’t just summarize. Instead, they write into two shared structures on the parent conversation: an evidence pool (each load-bearing fact paired with the exact source passage that supports it) and a citation registry (every source verified against CrossRef before it can be cited, so invented DOIs get rejected at write time).

Step 3: Draft

Next, the lead agent writes the report from the accumulated evidence, citing inline from the registry.

Step 4: The independent pass

Here’s the fusion-style step. While the lead drafts, a second model (GPT-5.5) has been running its own complete, blind investigation of the entire question: its own searches, its own source selections, its own reasoning path. Not a slice; the whole thing, unaware of the lead’s work.

Step 5: Reconciliation

Then the lead agent compares the two reports through a structured analysis (consensus points, contradictions, unique findings, blind spots) and applies surgical patches to its draft: pulling in what the independent pass found that it missed, resolving disagreements against sources, keeping the citation trail intact. This mirrors the insight OpenRouter’s Fusion demonstrated: two independent paths through the same question surface different facts, and synthesis captures the union. For example, they measured cross-model fusion (Opus + GPT-5.5) at +2.1 over self-fusion on their harness; our configuration mirrors that pairing.

Step 6: Verification

Finally, before delivery, the draft goes through an independent skeptical review that checks every specific claim against its cited sources and flags what it can’t verify. As a result, flagged claims get fixed, re-sourced, or removed. DRACO’s negative-weight criteria punish confident errors, and so do real users.

Why different models for different jobs? The researchers that gather facts, the model that investigates independently, and the model that reconciles are chosen per role. Flash-class searchers grind out queries cheaply; the frontier models do synthesis and judgment. We A/B tested promoting the researchers to Opus 4.8 — it scored worse (p = 0.005 at the criteria level) while costing more. Fact-gathering rewards search volume, not model size.

Methodology

• System. CoChat’s production Deep Research Mode, exactly as described above. The same features run for every CoChat user; nothing benchmark-specific was added.
• Assistant. We used the Research Assistant available as a template. The prompt was iterated using benchmark feedback (as with every published system); it contains no task- or rubric-specific content.
• Grading. The paper’s protocol — per-criterion binary verdicts by an LLM judge, multiple independent grading runs, mean normalized score. Runs were tight: mean per-task standard deviation of only ~1.0 normalized point across gradings, which is the measured noise floor we compare changes against [2]. Judge: gemini-3.1-pro-preview, temperature 0.2.
• Coverage. All 100 of 100 tasks graded, none blocked or refused [2].
• Memory off, leakage audited. Benchmark runs disable CoChat’s memory system for reproducibility, and every run passes an audit that scans the full agent-side context — including sub-agent transcripts — for rubric text and dataset markers. Search results from domains hosting the benchmark are excluded.
• The finding we’ll shout about since nobody else did. When we gave our agents web search, they occasionally found the DRACO rubric online — OpenRouter reported the same. Anyone benchmarking DRACO with live search needs exclusion lists and a leakage audit, or their factual-accuracy numbers are suspect.ct.

What building this taught us (the reusable part)

Three lessons shaped the architecture — and each one shows up in the numbers.

DRACO lesson 1: Single-agent search is brittle

Point an agent at a research question twice and you get two different sets of sources: which queries it fires, which pages it opens, and which facts it retains all vary run to run. That variance is the ceiling on factual accuracy, and no amount of prompting a single model removes it. Instead, running a second model through the entire question blind, then reconciling, converts that variance into coverage: two independent paths surface different facts, and synthesis captures the union. It’s the same effect OpenRouter measured with cross-model fusion (+2.1 over self-fusion), and it’s why the blind pass exists.

DRACO lesson 2: Verified citations win where it matters

Every fact is paired with its source passage in the evidence pool, and every source is verified against CrossRef before it can be cited. The payoff is measurable: citation quality is CoChat’s single strongest axis at 78.5, ahead of presentation (73.5), factual accuracy (69.9), and breadth (69). On the negative-weight citation traps, the criteria that punish fabricated or mismatched sources, CoChat catches 92.2%. Verify-at-write-time isn’t a compliance checkbox; it’s the difference-maker on the axis researchers care about most. We’ve written before about how researchers can verify sources across different AI tools, and this is the same failure mode those tools are graded on: fabricated citations that survive to the final draft.

DRACO lesson 3: Models forget, so we remind them

A long research run drifts: by the time the model is drafting, the evidence and citations it gathered early are far back in context and effectively fade. So our fix is to re-inject the evidence pool and citation registry into the model’s context throughout the loop, keeping the exact source passage backing each claim in front of the writer at the moment it writes the claim. Verified sources you don’t keep in view get dropped. In other words, keeping them present is what turns a full evidence pool into a fully-cited report.

Bonus: the benchmark became a debugging instrument. Working at rubric-criterion granularity, we also caught and fixed real production bugs — sub-agents silently losing research when a wind-down instruction named a tool they didn’t hold; a stream watchdog reaping healthy runs stuck mid-search at exactly 120 seconds; sub-agents inheriting memory context they should never have seen. Each shipped the same week. The harness — per-criterion verdicts, paired comparisons against a measured rerun-noise floor, and leakage audits — is now how we evaluate every deep-research change we make.

The data

• DRACO benchmark results: CoChatAi/benchmarks
• CoChat Research Assistant
• Try Deep Research Mode

References

• DRACO – https://arxiv.org/abs/2602.11685
• Perplexity Draco Dataset – https://huggingface.co/datasets/perplexity-ai/draco
• OpenRouter Fusion – https://openrouter.ai/blog/announcements/fusion-beats-frontier/