#3 of 15
55.29 on the official leaderboard1
A benchmark for deep research agents. It has 100 PhD-level research tasks, written by experts in 22 fields, in Chinese and English (ICLR 2026).4
45 organizations evaluate performance based on it, including:5
Agentic Research
85%
of input tokens are rereading session history
Frontier research agents work in a loop: search, open a page, take notes, repeat. Before each new step, they read all their old notes again. So most of their effort goes into rereading, not into new sources.2
GPT-6 Astra Pro · Max Reasoning
Read-Time Compute
2%
of input tokens are rereading session history
Our state-of-the-art orchestrator handles the loop itself, so the model does not reread old notes. We spend the savings on research: 5X more searches,2 and an engine built to read thousands of sources.7
Deep Cove Research · Max Reasoning
On average, frontier research agents cost $15.36 per task. Rereading their old notes is billed at a discount, but still costs $3.76, almost 4X our entire cost. Meanwhile, most of our cost goes to comprehending new sources.2
Agentic Research
GPT-6 Astra Pro Deep Research
$15.36
Read-Time Compute
Deep Cove Research
$0.99
Agentic Research
14
Average Searches per Report
Frontier research agents have limited memory, and the loop uses most of it. So they only have room for about 14 searches per report. For many sources, they read only the short snippet, not the complete content, so details are easy to miss or misread.2
GPT-6 Astra Pro · Max Reasoning
Read-Time Compute
68
Average Searches per Report
Our state-of-the-art context management keeps the model’s memory clear. So it has room for about 68 searches per report, 5X more than frontier agents. We comprehend the whole page and turn it into evidence cards for deep comprehension.2
Deep Cove Research · Max Reasoning
The comprehension engine works from real-time, post-training sources it comprehends at runtime, not from obsolete knowledge. That makes model size a minor factor. We run a model priced about 40X lower than GPT-6 Astra Pro and Claude Fable 5.1, and score higher.8
Cost and Score Sources: 8
Frontier agents decide every step on their own: what to search, what to open and when to stop. These decisions come from the model’s pre-training, including its biases, so the same question can give a different report each time. Our engine runs every task through one research guardrail and builds every claim from post-training sources it has read. This reduces pre-training bias and makes reports more replicable.
Frontier agents keep all their progress in one long session with the model. If the session stops halfway, the progress is lost, and you have to run the task again from the start and pay again. Our engine splits research into stages. If a run stops, you can resume it from the last checkpoint or start a new fork from there, so you never pay for the same work twice.
You can freely scale up your deep research plan without cost pressure. Our state-of-the-art orchestrator saves the model’s context window and dices the work into many small calls, so scaling up your plan makes the cost grow linearly. Meanwhile, frontier agents pay the looping cost on every extra step, so for them the same increase makes the cost grow non-linearly.7
Our engine handles the research process, so a light model is enough, at about 31X lower hardware cost.9
Frontier-Scale Model
2.8T Parameters
16 × NVIDIA B200 / $109/hr
Model Used by Deep Cove Research
117B Parameters
1 × NVIDIA H100 / $3.50/hr
Deep Cove Research is a stealth team based in North Vancouver, British Columbia. We build the comprehension engine and test it in public, on DeepResearch Bench.
Deep Cove Research is opening in stages. Request an invitation for your team. Available as a web app and as an MCP server for your own agents.
Tell us what you research, and we will reply.
Every report Deep Cove Research submitted to DeepResearch Bench, with its overall score, 4 sub-scores and cost.
Per-report scores are our own evaluation with the official RACE method. Costs use public API prices.
The same research task, judged the same way.
Field
Compare With
Scored with the official RACE method and GPT-5.5 judge. Costs are averages per report.
