Deep research as a commodity.
Faster, cheaper, traceable research — produced by a competitive swarm of miners on Bittensor SN67.
Deep costs dollars. Fast loses depth.
Best-in-class engines deliver quality but take minutes and cost dollars per query. Basic search APIs are fast but offer no synthesis, no citations, no coverage guarantees. Web search vs deep research: Even providers distinguish these as fundamentally different tiers — OpenAI classifies web search as “fast and ideal for quick lookups” while describing deep research as a long-running workflow requiring background mode and webhooks. OpenAI Web Search GuideOpenAI Deep Research API docs Agents need both. Right now, neither option delivers.
GPT 5.2 Pro
$1.83
per query · 5–30 min Existing deep research pricing/latency: OpenAI GPT 5.2 PRO ~$1.83/query, 5–30min; Gemini DR ~$3.68; Perplexity DR ~$1.54. Native APIs charge per token; figures represent average cost per completed workflow. Parallel Web Systems Benchmarks (Dec 2025)Parallel DeepSearchQA blog (Dec 2025)
Gemini DR
$3.68
per query · variable Existing deep research pricing/latency: OpenAI GPT 5.2 PRO ~$1.83/query, 5–30min; Gemini DR ~$3.68; Perplexity DR ~$1.54. Native APIs charge per token; figures represent average cost per completed workflow. Parallel Web Systems Benchmarks (Dec 2025)Parallel DeepSearchQA blog (Dec 2025)
Harnyx — Target
10x lower
per query · faster
Not fragments. Completed analytical work.
Builders wire in search APIs because the alternative costs dollars per query. The agent gets fragments and acts anyway — it has no judgment to fill the gap. Agent scaling barriers: 46% of organizations cite integration with existing systems and 42% cite data access/quality as the biggest barriers to scaling AI agents. The 2026 State of AI Agents (Orbislabs.ai, Jan 2026)
Comprehensiveness
Coverage guarantees that surface what a surface-level search would miss.
Synthesis
Not fragments from multiple sources, but a single analytical judgment you can act on.
Traceability
Every claim linked to its origin, not “according to sources.”
Agents don’t pause for a second opinion.
The quality of the research is the quality of the outcome.
One API call. Structured research back.
Stop stitching together search calls, extraction scripts, and verification steps.
curl https://api.harnyx.ai/v1/research \
-d '{
"input": "Compare Salesforce vs HubSpot for mid-market CRM in 2025.",
"format": { "winner": "string", "price_per_seat_usd": "number", "key_reason": "string" }
}' {
"report": {
"winner": "HubSpot",
"price_per_seat_usd": 45,
"key_reason": "HubSpot's bundled AI tier undercuts Salesforce
Einstein by $50/seat [1], with faster onboarding cited by
60% of churned accounts [2]..."
},
"citations": [
{ "index": 1, "title": "HubSpot Q2 Earnings Call", "url": "..." },
{ "index": 2, "title": "Forrester Wave: Sales Force Automation", "url": "..." }
]
} Works with any agent stack
- OpenClaw
- LangChain
- CrewAI
- n8n
The best pipeline is a moving target.
New models, retrieval methods, and cost structures emerge constantly. How fast a provider adapts is the defining competitive axis.
Read our vision statementBig Model Labs
OpenAI DR, Gemini DR
Locked to their own model, even when others excel at specific stages. Harness optimization competes for bandwidth with model training, product features, and safety.
Centralized Harness
Perplexity Sonar
Decouples the harness from any one model — can swap faster. But every optimization still flows through one team’s bandwidth and release cycle.
Open-source Harness
Perplexica
Open development, community contributions. But no economic incentive for continuous optimization — adaptation is voluntary and sporadic.
Bittensor Subnet
Harnyx — SN67
Decoupled harness + open development + persistent economic incentives. Miners adapt in real time because the market rewards whoever finds a better approach first.
The standard is stable.
The execution is adaptive.
Parity first. Then lead.
Harnyx starts by matching the best deep research available — then the competitive pressure of the network takes it further.
Reach Parity
CurrentResearch output at industry benchmark quality — DRACO DRACO (Deep Research Accuracy, Completeness, and Objectivity): Open benchmark for deep research agents — 10 domains, 100 tasks, expert-grounded rubrics. Released by Perplexity. arXiv 2602.11685 (Feb 2026)HuggingFace dataset and DeepSearchQA DeepSearchQA: 900-prompt benchmark for multi-step information-seeking tasks — 17 domains, causal-chain structure. Released by Google DeepMind. DeepSearchQA paper (Dec 2025)Kaggle leaderboard standards — delivered at meaningfully lower cost. External API for agent builders. Drop-in integrations for major agent frameworks.
Lead and Define
Miners compete head-to-head instead of against fixed reference answers. When the goal shifts from matching a known target to outperforming each other, miners surface sources and reasoning paths others missed — novel synthesis emerges from competitive pressure, not from a predefined ceiling.
Enterprise Readiness
Confidential compute via TEE. Zero data retention. SLA guarantees. The infrastructure that lets agent builders serve finance, legal, and consulting — verticals where research is high-value and the cost of a wrong answer is measured in dollars.
Backed by
Run the network
A unified API routes queries across a competitive swarm of miners, ranked by outcome-focused validators.
Miners
Build research pipelines
Compete to produce the highest-quality deep research. Build your own pipeline — choose models, retrieval strategies, synthesis methods. The best results earn the most emissions.
Miner GuideValidators
Secure the network
Grade miner submissions against reference answers and distribute rewards based on quality.
Validator Guide