Updated Nov 10, 2026 · every number cited
The GTM model leaderboard.Which AI is best at getting customers?
Small models trained for marketing are starting to beat the giants at narrow jobs. We collect the public benchmarks here, with sources. Soon: a real-world arena ranked on actual launch results from GrowthRouter users.
[ 01 / 03 ] adval: running X ad campaigns
AdVal by loglog.ai. A $0.40 open model beat every frontier model.
Agents get a client brief and budget ($100–2,700) and run a campaign on a simulated X Ads market through an emulated X Ads API. 45 held-out briefs, one run each. Spend is simulated, not real.
| # | Model | Maker | Weights | Share of best result | Score vs default campaign | Cost / run |
|---|---|---|---|---|---|---|
| 1 | yvligo | loglog.ai | open (MIT LoRA on DeepSeek-V4.1-Flash) | 77% | 1.31 | $0.40 |
| 2 | Gemini 4 Argon | closed | 70% | 1.20 | $2.14 | |
| 3 | GPT-6.1 Sol | OpenAI | closed | 67% | 1.15 | $0.39 |
| 4 | DeepSeek V4.1 Flash (base) | DeepSeek | open | 64% | 1.10 | $0.02 |
| 5 | Grok 4.7 | xAI | closed | 64% | 1.09 | $2.05 |
| 6 | Claude Opus 5.5 | Anthropic | closed | 63% | 1.07 | $2.50 |
| 7 | GPT-6 Astra | OpenAI | closed | 62% | 1.06 | $2.11 |
| 8 | Muse Spark 1.3 | Meta | closed | 61% | 1.04 | $0.30 |
| 9 | Claude Haiku 5.5 | Anthropic | closed | 59% | 1.01 | $0.08 |
| 10 | Claude Sonnet 5.5 | Anthropic | closed | 55% | 0.93 | $0.75 |
| 11 | GPT-6 Luna | OpenAI | closed | 43% | 0.74 | $0.02 |
| 12 | Nemotron 3.5 Lightning | NVIDIA | open | 13% | 0.21 | $0.09 |
Sources: loglog.ai (share of best known result, across their runs), github.com/loglogai/AdVal (geometric mean of conversions ÷ par, where par = one default ad group at X's AUTO bid; cost per run), huggingface.co/loglogai/yvligo. Model names as the authors label them. Caveats from the authors: one run per brief, about 10% run-to-run variation, and the simulator may reward some targeting more than real X does.
[ 02 / 03 ] open simulators for X and Reddit
Simulate a crowd yourself. Open-source engines.
| Engine | Repo | License | What it simulates | How heavy |
|---|---|---|---|---|
| OASIS | camel-ai/oasis | Apache-2.0 | X and Reddit, up to 1M agents, 23 actions | Python + LLM calls; ~50–170k tokens for a small run |
| MiroFish | 666ghj/MiroFish | AGPL-3.0 | Builds a world from documents, runs X/Reddit sims on OASIS, writes a report | Heavy; Docker; ~85K stars |
| AgentSociety | tsinghua-fib-lab/AgentSociety | Apache-2.0 | City- and society-scale social simulation | Research platform, many LLM calls |
| Y Social | YSocialTwin/YSocial | GPL-3.0 | Microblog digital twin with a REST server | Python, local or API LLM |
| Generative Agents | joonspk-research/generative_agents | Apache-2.0 | Believable agents with memory and reflection (the original) | Background reference |
What the research says: AI alone predicts the winning post only a little better than chance; grounded in real community data and checked against small real tests, it works. That's how GrowthRouter's pre-flight check is built.
[ 03 / 03 ] list your model
Built a model or benchmark for marketing, sales or GTM? Get it listed.
← Back to GrowthRouter: get your first users from Reddit and X