Production tasks
7272production tasks
11task suites
15market categories
4data environments
GTM-Bench is an open benchmark that measures whether AI agents can do real go-to-market work: understanding what a seller offers, inferring who should buy it, and retrieving the right accounts with evidence to back every recommendation. As AI makes it trivially cheap to generate more outbound, we built GTM-Bench to measure whether agents reduce noise or add to it: rewarding evidence-backed matches instead of raw volume.
Leaderboard
GTM Bench Leaderboard
Results are derived from the GTM-Bench paper. The table reports volume-weighted net score, active A-grade rate, efficiency metrics, and trace context for the evaluated purpose-built and generalist systems.
Best net score
26,615.6Best generalist
37.7%Filter by system type
Showing 7 entries
Data source: GTM-Bench paper Tables 6, 8, 9, and 12.
| Rank | Runner | Harness | Net Score | A-grade % | Net / $ | Cost | Composite Index | Returned Records | Trace Tendency |
|---|---|---|---|---|---|---|---|---|---|
| 1 | Blackpearl RTSAblackpearl-rtsa | Blackpearl RTSA | 26,615.6 | 40.9% | 56.63 | $470.03 | +14.18 | 38,823 | Purpose-built offer, ICP, retrieval, and activation system |
| 2 | GPT-5.5 (Codex)openai_gpt-5.5 | codex-cli / openai:gpt-5.5 | 4,040.9 | 37.7% | 27.44 | $147.27 | +16.63 | 1,753 | Compact, script-driven, selective |
| 3 | Claude Sonnet 4.6anthropic_claude-sonnet-4.6 | claude-code-cli / sonnet-4.6 | 400.1 | 27.3% | 2.51 | $159.61 | +0.53 | 5,678 | High-recall SQL/tgrep explorer |
| 4 | Claude Opus 4.7anthropic_claude-opus-4.7 | claude-code-cli / opus-4.7 | -2,476.6 | 31.7% | -9.17 | $270.20 | -3.64 | 4,904 | Deliberative, evidence-heavy, costly |
| 5 | DeepSeek V4 Proopenrouter_deepseek-v4-pro | hermes-agent-cli / deepseek-v4-pro | -3,398.0 | 21.8% | -99.26 | $34.23 | -10.77 | 2,531 | Low-cost, web-assisted, moderate recall |
| 6 | Gemini 3.5 Flashgoogle_gemini-3.5-flash | hermes-agent-cli / gemini-3.5-flash | -10,671.9 | 13.6% | -148.35 | $71.94 | -42.36 | 2,015 | Fast but schema/identity-inconsistent |
| 7 | Kimi K2.6openrouter_kimi-k2.6 | hermes-agent-cli / kimi-k2.6 | -15,402.3 | 22.8% | -171.56 | $89.78 | -22.74 | 5,496 | Slow, expansive, overproduces |
Blackpearl RTSA
26,615.6
GPT-5.5 (Codex)
4,040.9
Claude Sonnet 4.6
400.1
Claude Opus 4.7
-2,476.6
DeepSeek V4 Pro
-3,398
Gemini 3.5 Flash
-10,671.9
Kimi K2.6
-15,402.3
Benchmark analysis
Production catalog composition
The paper reports a 72-task production catalog designed from a source query taxonomy, then stratified across task suites, market categories, and pressure tags.
Task suites
Offer-grounded lead lists8.3%
Named-domain offer extraction11.1%
Offer-to-lead list12.5%
Vertical / geo / firmographic search11.1%
Persona / contact activation12.5%
Intent / trigger evidence9.7%
Technographic search6.9%
Offer-grounded lookalikes4.2%
Market-to-lead list4.2%
Buyer search with contact details13.9%
Limited-context prompts5.6%
Pressure tags
Geo/firmographic constraints44.4%
Compliance or policy sensitivity38.9%
Persona/contact fit37.5%
Intent or recency evidence26.4%
End-to-end activation18.1%
Technographic evidence12.5%
ICP inference11.1%
Comparator/lookalike reasoning5.6%
Source query taxonomy
Domain based lead search43.8%
Generic lead search17.6%
Persona search6.8%
Offer to ICP discovery6.2%
Geographic/local search5.1%
Firmographic filtering4.5%
Industry category search4.4%
Technographic search4.1%
Intent/trigger search3.6%
Lookalike search3.5%
Market research1.7%
Non-GTM / Ambiguous1.4%
GTM-Bench is the first step in a longer effort to define how AI should be evaluated for go-to-market work. To collaborate, open an issue on GitHub or email us at research@blackpearl.com — we'd love to build the next version with you.
Paper and release artifacts
GTM Bench: Evaluating AI Systems for Go-to-Market Workflows
The site includes the current paper PDF, the leaderboard, and the open-source release for the benchmark harness.