Production tasks
72Results updated · September 2026
GTM Bench
A benchmark for measuring whether AI systems can turn an offer into useful go-to-market work: clear ICPs, grounded account matches, reachable contact paths, and activation-ready outputs.
72 production tasks
11 task suites
15 market categories
4 data 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
The June leaderboard now includes RTSA 2.0 and Fable 5.1, with and without Pearl Engine data, evaluated on the same 72 published tasks. Updated September 13, 2026.
Best net score
205,731.5Blackpearl RTSA 2.0
Best generalist · A-grade
56.1%Fable 5.1 + Pearl Engine data
September results · 72 / 72 tasks complete
RTSA 2.0 leads with 205,731.5 net points
4.36× Fable 5.1 with Pearl Engine data and 7.73× RTSA 1.0. Fable scores 47,136.6 with data and −3,463.7 web-only. RTSA 2.0 records 85.7 net points per dollar, compared with Fable’s 56.1.
Filter by system type
Showing 11 entries
June paper + September 2026 benchmark results.
| Rank | Runner | Harness | Net Score | A-grade % | Net / $ | Cost | Composite Index | Returned Records | Trace Tendency |
|---|---|---|---|---|---|---|---|---|---|
| 1 | Blackpearl RTSA 2.0 12 Sep 2026 | Blackpearl RTSA 2.0 | 205,731.5 | 59.3% | 85.69 | $2401.00 | - | 24,416 | Purpose-built offer, ICP, retrieval, and activation system |
| 2 | Fable 5.1 + Pearl Engine data 11 Sep 2026 | claude-code-cli / fable-5.1 / Pearl Engine | 47,136.6 95% CI 45,601 to 48,717 | 56.1% | 56.09 | $840.37 | +46.06 | 7,001 | 72 tasks; four output-fault tasks scored as delivered |
| 3 | Blackpearl RTSA 1.0 June 2026 | Blackpearl RTSA | 26,615.6 95% CI 24,281 to 28,950 | 40.9% | 56.63 | $470.03 | +14.18 | 38,823 | Purpose-built offer, ICP, retrieval, and activation system |
| 4 | GPT-5.5 (Codex) June 2026 | codex-cli / openai:gpt-5.5 | 4,040.9 95% CI 3,364 to 4,713 | 37.7% | 27.44 | $147.27 | +16.63 | 1,753 | Compact, script-driven, selective |
| 5 | Codex GPT-5.5 (web-only) June 2026 | codex-cli / openai:gpt-5.5 / web-only | 1,015.4 95% CI 495 to 1,549 | 31.6% | - | - | - | - | Web-only baseline; no Pearl Engine data |
| 6 | Claude Sonnet 4.6 June 2026 | claude-code-cli / sonnet-4.6 | 400.1 95% CI -1,138 to 1,942 | 27.3% | 2.51 | $159.61 | +0.53 | 5,678 | High-recall SQL/tgrep explorer |
| 7 | Claude Opus 4.7 June 2026 | claude-code-cli / opus-4.7 | -2,476.6 95% CI -3,788 to -1,182 | 31.7% | -9.17 | $270.20 | -3.64 | 4,904 | Deliberative, evidence-heavy, costly |
| 8 | DeepSeek V4 Pro June 2026 | hermes-agent-cli / deepseek-v4-pro | -3,398.0 95% CI -4,252 to -2,544 | 21.8% | -99.26 | $34.23 | -10.77 | 2,531 | Low-cost, web-assisted, moderate recall |
| 9 | Fable 5.1 web-only 11 Sep 2026 | claude-code-cli / fable-5.1 / web-only | -3,463.7 95% CI -4,353 to -2,593 | 22.8% | -6.79 | $510.20 | -13.38 | 1,711 | Web-only baseline; no Pearl Engine data |
| 10 | Gemini 3.5 Flash June 2026 | hermes-agent-cli / gemini-3.5-flash | -10,671.9 95% CI -11,280 to -10,051 | 13.6% | -148.35 | $71.94 | -42.36 | 2,015 | Fast but schema/identity-inconsistent |
| 11 | Kimi K2.6 June 2026 | hermes-agent-cli / kimi-k2.6 | -15,402.3 95% CI -16,698 to -14,165 | 22.8% | -171.56 | $89.78 | -22.74 | 5,496 | Slow, expansive, overproduces |
Blackpearl RTSA 2.0 205,731.5
Fable 5.1 + Pearl Engine data 47,136.6
Blackpearl RTSA 1.0 26,615.6
GPT-5.5 (Codex) 4,040.9
Codex GPT-5.5 (web-only) 1,015.4
Claude Sonnet 4.6 400.1
Claude Opus 4.7 -2,476.6
DeepSeek V4 Pro -3,398
Fable 5.1 web-only -3,463.7
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 lists 8.3%
Named-domain offer extraction 11.1%
Offer-to-lead list 12.5%
Vertical / geo / firmographic search 11.1%
Persona / contact activation 12.5%
Intent / trigger evidence 9.7%
Technographic search 6.9%
Offer-grounded lookalikes 4.2%
Market-to-lead list 4.2%
Buyer search with contact details 13.9%
Limited-context prompts 5.6%
Pressure tags
Geo/firmographic constraints 44.4%
Compliance or policy sensitivity 38.9%
Persona/contact fit 37.5%
Intent or recency evidence 26.4%
End-to-end activation 18.1%
Technographic evidence 12.5%
ICP inference 11.1%
Comparator/lookalike reasoning 5.6%
Source query taxonomy
Domain based lead search 43.8%
Generic lead search 17.6%
Persona search 6.8%
Offer to ICP discovery 6.2%
Geographic/local search 5.1%
Firmographic filtering 4.5%
Industry category search 4.4%
Technographic search 4.1%
Intent/trigger search 3.6%
Lookalike search 3.5%
Market research 1.7%
Non-GTM / Ambiguous 1.4%
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 June paper PDF documents the original benchmark. The leaderboard also includes the September results.