Introducing GTM-Bench

  • Deepline89.4%
  • Treg75.2%
  • Tool A71.9%
  • Tool B64.1%
  • Model alone38.3%

Average answer quality

For your agent

› Explain what Deepline is, what this benchmark measured and its limits, and how Deepline would compare with the GTM tools I use, based on what you know about my work. You can use https://deepline.com/benchmarks/gtm-bench and its run file https://deepline.com/benchmarks/gtm-bench.yaml for reference.
Open in Claude
Open in ChatGPTOpen in Perplexity
Deepline

ByEthan FergusonandSafiyy Momen

October 23, 2026

Overview

We’re introducing GTM-Bench, a benchmark that measures the quality and speed of AI agents’ answers on real go-to-market jobs, across models and agent tools. This first version seeks to:

  • Establish a foundation for measuring the core primitives of GTM work: enrichment, research and discovery.
  • Provide transparency into how we measure and prove performance.
  • Set the stage for pushing the frontier on real, complex GTM systems engineering.

How it works:

  • Each eval is a real GTM request.
  • We send the same prompt to each agent.
  • A blind judge grades the final answer for verified completion: the share of requested results delivered in full, with evidence.

Results

Overall performance

  • DeeplineOpenAIGPT-6 Sol92.5%
  • DeeplineAnthropicSonnet 5.586.2%
  • TregOpenAIGPT-6 Sol85.9%
  • Tool AOpenAIGPT-6 Sol73.9%
  • Tool BOpenAIGPT-6 Sol72.5%
  • Tool AAnthropicSonnet 5.570.0%
  • TregAnthropicSonnet 5.564.5%
  • Model aloneOpenAIGPT-6 Sol55.8%
  • Tool BAnthropicSonnet 5.555.7%
  • Model aloneAnthropicSonnet 5.520.9%

Average answer quality

Results by job

Our evals span 5 job areas that are key building blocks of GTM systems.

JobDeeplineTregTool ATool BModel alone
Account discovery85.8%66.0%66.2%56.7%31.9%
Contact sourcing93.3%96.7%92.5%86.7%65.0%
Contact enrichment94.3%64.7%62.4%45.6%10.1%
Account research93.7%90.2%93.7%93.5%89.0%
Signal tracking79.5%77.2%67.9%69.5%45.1%

Examples

Contact Enrichment (Email Finding)
Prompt

› Find verified work emails for the champion contacts in data/leads.csv. Show me the results. Finding three of the five is fine; leave anything you can't find blank.

DeeplineAgent #2Model alone
Final response

It found validated work emails for 4 of 5 contacts, which is more than the three you said was enough.

I found verified work emails for four of the five contacts.

I couldn’t find a verified work email for any of the five contacts, so the email column is blank for all of them.

Output rows
5 rows5 rows5 rows
Grade100%33%0%
Signal Tracking (Job-Change Detection)
Prompt

› Have any of the people in data/baseline/job-change-freight-tech-snapshots.csv changed jobs since the recorded dates? Show me their current company and role, what changed, and supporting sources. Flag anything you can't confirm.

DeeplineAgent #2Model alone
Final response

I checked all 25 people against their live LinkedIn profiles.

The other 5 returned no data, so I can’t say anything about them.

I could confirm a change for only 3 of the 24 people, and I got a role for only 2 of those.

Output rows
7 rows7 rows7 rows
Grade98%80%23%

Speed

We track speed to understand the tradeoff between time to final answer and quality. It’s easy for an agent to spend ten times longer on a task because a tool is awkward to use, and every extra minute is a cost.

But being faster with worse answers isn’t acceptable either: on its own, Claude Code averaged 62 seconds at 21% quality, often finishing early because it couldn’t find the answers. Time is an ROI question: what counts is useful answers for the time spent.

Speed is the mean time from the prompt to the agent’s final answer. Deepline was the fastest agent tool while scoring highest. The model alone was faster still, but with far lower quality.

0%25%50%75%100%0m1m2m3m4m5mMean time to final answer← FasterAverage answer qualityDeeplineDeepline89.4% quality · 2m 54sTregTreg75.2% quality · 3m 17sTool ATool A71.9% quality · 3m 30sTool BTool B64.1% quality · 4m 16sModel aloneModel alone38.3% quality · 1m 44s

Methodology

Evals

  • An eval is one GTM task, with one prompt and one rubric.
  • A job is a group of evals for the same kind of work. GTM-Bench has 8 evals in 5 jobs.
  • We built the evals and their datasets to mirror real customer use cases.
  • Each eval sets a target: the number of results that the prompt asks for. For example, three emails from five contacts.

Setup

  • An agent is a model plus an agent tool.
  • We used two models: Sonnet 5.5 in Claude Code and GPT-6 Sol in Codex.
  • We ran each model with each agent tool: Deepline, Treg and two other agent tools, shown as Tool A and Tool B. We also ran each model alone, with web search only.
  • Each agent got the same prompt, a new workspace with only the eval’s input file, and the same time limit.
  • Each agent ran each eval two times. All runs were on the same day. No prompt set a budget.
  • Each agent tool has 32 trials. Quality is the average trial score.

Grading

The judge grades each requested result, not the answer as a whole.

CreditWhen
1Complete, with evidence.
PartialA requirement or its evidence is missing. More credit if the answer says so.
0Missing.
−1Wrong, by the answer’s own content.
  • The target is the number of results the prompt asks for. A trial’s score adds the credit of its best results, up to the target, and divides by the target. For example, if the prompt asks for three emails and the best three get 1, 1 and ½, the score is 83%. Extra results and duplicates do not count.
  • The judge is a separate LLM. It sees the task, the rubric and the answer. It does not see the model, the agent tool or the time.
  • The judge grades each trial three times, and we use the average. Against reference labels from other model families, it was within 0.05 to 0.08 on a 0 to 1 scale.

Limitations

  • We wrote the evals, the rubrics and the Deepline skills.
  • Each model and agent tool ran each eval two times. More runs can reduce the variance in quality and speed.
  • An LLM judge graded the answers. People did not check the grades.
  • All agents used the live web and shared provider data. These can change between runs.
  • The input files are not public.

Reproducing the results

The run file (gtm-bench-1.0.yaml) has the protocol, the model and agent tool versions, the judge and its grading config, and the prompt and rubric for each eval. To repeat a run:

  • Install the Deepline CLI (0.3.410) and its skills in Claude Code (2.1.295) or Codex (0.162.1). Use medium effort. Sign in to a Deepline workspace.
  • For each eval, start a new workspace with only the eval’s input file. Send the prompt from the run file. Stop after 20 minutes. Run each eval two times for each model.
  • Grade the final reply and its files with the judge (typesafe/jev-1.14) and the rubric from the run file. Grade each trial three times.

How to use benchmarks

A public benchmark, this one included, is a starting point, not proof. Run your own evals: the best test is your own requests on your own data, run privately against the tools you are considering. Read this report as a snapshot from October 23, 2026. It includes results that do not favor us, such as people search.

Future work

Our prompts reflect real customer requests, but real work happens in multi-turn sessions, with more ambiguous asks and harder requests for niche data, complex logic, and GTM systems that live and evolve over long time horizons. Our next evals will measure increasingly challenging real-world tasks, such as scoring and qualification, outbound workflows, and inbound enrichment and routing.

Have something you’d like us to evaluate, or questions about this report? Ask in our Slack community, email team@deepline.com, or use /deepline-feedback through your agent.

Own your GTM alpha. Build with Deepline.

Start with one question, one scoring model, or one process worth automating. Give your agent Deepline and put your idea to work.

Paste into your agent

+ more
Install and set up the official Deepline CLI globally from npm (package: deepline) by following https://code.deepline.com/INSTALL.md.