DeepDog 2: Maliyetin çok altında Gemini ve OpenAI derin araştırmadan daha iyi performans gösterir

Turn a question into a cited Markdown report. A supervisor plans the research, delegates work to platform agents, reviews their findings, and writes the final report.

Deep Dog 2 installs directly from GitHub as a Python package in your existing project. Call run_research() with just a question to use the defaults, or pass a per-run configuration to choose models, search, agent types and research budgets. See the quickstart for pip and uv installation, or clone the source to edit the engine and run the repository examples.

At the time of publication, Deep Dog 2 ranked 5th overall and 1st among open-source research agents on the DeepResearch Bench . The published run used a relatively economical profile: a 15-minute research window, a 20-iteration supervisor cap, at most 3 Exa searches per sub-agent, DeepSeek V4 Pro as supervisor, and DeepSeek V4 Flash for sub-agents.

Deep Dog 2 retains the draft-first, iterative refinement idea from Deep Dog 1, while making reflection, delegation, and source handling explicit:

Supervisor reflection is used for planning and control; it is not copied into the final research report. Optional subtopic evaluation and parallel subtopic reports can run after the main report when enabled in deep_research/config.py .

The supported prompt family is OPEN . Older configurations using LEGACY must switch to OPEN .

At the time of publication, Deep Dog 2 ranked 5th overall and 1st among open-source research agents on the DeepResearch Bench benchmark.

These results were achieved with a relatively economical configuration: a maximum of 3 Exa searches per sub-agent, 15 minutes of research time, 20 total iterations, and DeepSeek V4 Pro and DeepSeek V4 Flash as the supervisor and sub-agent base models.

These results are a historical reproducibility profile, not a promise about current defaults. The current package defaults to DeepSeek V4 Flash for all model roles and uses a different supervisor iteration default. Benchmark rankings and scores may change as the leaderboard changes.

The following is a pricing-based estimate for one research task. It is not a controlled cost benchmark: the systems use different architectures, search providers and token budgets.

Deep Dog 2’s range is an estimate based on the DeepSeek V4 pricing (off-peak) and Exa pricing , using off-peak DeepSeek rates.