GTMBENCHv2026.07 · 30 models · refreshed monthly

Results · July 2026 run

Nico 2.5 is the first model to clear 88 on GTM Workbench.

GTMBench scores language models on what revenue teams actually do: research the account, build the dossier on the company and the buying committee, write the outbound, run the call, handle the objection, keep the CRM clean, and plan a named account end to end. Eight task families, 2,946 graded tasks, rebuilt every month from accounts and transcripts that did not exist when the models were trained.

01Nico 2.5GTM foundation model

MacroDeep’s GTM foundation model leads three of the eight families at roughly a third the blended cost of the closed frontier. It cedes account research, company & persona research, call analysis, crm extraction and sequence execution, the families where raw frontier scale still beats specialisation.

88.4
Top GTM score
2,946
Graded tasks
8
Task families
10.5
Closed → open gap, best of each
12/30
Open-weight entries

The leaderboard

Sort any column. Select a row for the full profile. Scores are 0–100; the GTM score is the unweighted mean of the eight families.

License
#License
Showing 30 of 30 modelsBlended cost is USD per million tokens, assuming a 3:1 input-to-output ratio at July 2026 list prices.

Open vs. closed

The gap is not uniform, and most of it belongs to one specialist.

On account research the best open-weight model is 7.1 points off the closed frontier. On outbound copy the same comparison costs 21.1. Strip the purpose-built GTM model out of the closed column and the widest gap roughly halves: what is left is a pricing question, not a capability one.

FamilyClosedOpenGap
Account research87.980.8+7.1
Company & persona research90.680.2+10.4
Outbound copy92.971.8+21.1
Call analysis88.880.7+8.1
Objection handling91.275.2+16.0
CRM extraction86.578.5+8.0
Sequence execution90.179.9+10.2
End-to-end account plan91.580.2+11.3

Best-in-class per license, July 2026 run.

Score against price

Blended cost per million tokens, log scale. Filled marks are closed models; outlined marks are open weights.

Scatter plot of GTM score against blended cost per million tokens on a log scale. Score rises with price, but the leading GTM specialist sits well above the trend line at mid-range cost, and the strongest open-weight models cluster near 78 at under a dollar.90807060$0.10$1$10$100Nico 2.5Fable 5DeepSeek-V4-ProDeepSeek-V4-Flash

The eight families

Every task is scored 0–100 against a rubric written by working GTM operators, then blind-graded three times.

01 · Res · 412 tasks

Account research

Given a company URL, a 10-K excerpt and four news items, produce the account brief a rep would open before a first call: what changed, who owns the problem, what the trigger is. Graded on cited specifics and on penalising invented facts.

Rubric graders: 3 · Hallucination penalty: −12

02 · Deep · 288 dossiers

Company & persona research

Twenty tool-using minutes on one company and one named person: what the company just changed, who sits on the buying committee, what that persona is measured on, which of their own public words to quote back. Graded on cited primary sources, on committee coverage, and on a hard penalty for inferred job scope.

Sources per dossier: 30+ · Uncited claim: −15

03 · Out · 468 tasks

Outbound copy

Write the first-touch email and the follow-up under real constraints: a named persona, a 90-word ceiling, one ask, no unearned familiarity. Graded on specificity to the account, on the strength of the ask, and against a spam-signal classifier.

Human preference pairs: 1,404

04 · Call · 386 tasks

Call analysis

Turn a 40-minute discovery transcript into qualification fields, risks and next steps. Graded field-by-field against an operator's own notes, with credit for flagging what the rep failed to ask.

Exact-match fields: 14 per call

05 · Obj · 340 turns

Objection handling

Six turns against a scripted buyer who is polite, budget-constrained and already talking to a competitor. Graded on whether the model holds price, surfaces the real blocker, and stops selling when the buyer says yes.

Buyer personas: 22 · Turns per run: 6

06 · CRM · 460 records

CRM extraction

Read a messy thread and emit a schema-strict object: accounts, contacts, amounts, stage, close date. Graded on schema validity, on deduplication against existing records, and on abstaining rather than guessing.

Schema violations: hard zero

07 · Seq · 346 runs

Sequence execution

Plan and execute fourteen days of multi-threaded follow-up with eight tools available: enrichment, calendar, email, CRM write, pricing lookup. Graded on plan coherence, tool-call validity and recovery after a failed call.

Tools: 8 · Max steps: 30

08 · Plan · 246 plans

End-to-end account plan

One named account, three personas, four channels, thirty days: where to enter, how to sequence the committee, what each persona hears, and what to do when the champion goes quiet. Graded head-to-head against the plan an operator wrote for the same account, and on whether the plan survives its own first failure.

Operator head-to-head · Personas per plan: 3

Contamination control

Every account, transcript and CRM snapshot in a run is drawn from material first published in the 30 days before that run, or generated with operators under NDA. Nothing in the July 2026 set was on the public web when any evaluated model finished training. Prior sets are released after the following run.

Grading

Rubric items are graded three times: twice by a model panel that excludes the model under test, once by a human operator on a 20% sample. Panel and human disagree on 4.1% of items; those items are re-graded by a second operator and the human label wins.

Settings

Temperature 0 where supported, no system-prompt tuning per model, no retries. Reasoning models run at their default effort. Every model sees the same tool schemas. Runs are single-shot: the score you see is the first answer, not the best of five.

The arithmetic

Every number on this page is reproducible from the published task set and grader prompts. This is the whole chain, from one graded rubric item up to the composite.

01 · From rubric item to score

A task is not scored as a whole. It is decomposed into weighted rubric items, each graded on a three-point scale (met, partially met, missed), and the task score is the weighted proportion met, minus whatever penalties the family charges. Weights are fixed when the rubric is written and do not vary by model.

GTM  =  (1/8) · Σ F_k                 the eight family means

F_k  =  clip( (1/n_k) · Σ s_i, 0, 100 )   over n_k tasks

s_i  =  100 · ( Σ w_j·r_ij / Σ w_j ) − Σ p_ij

        r_ij ∈ {0, 0.5, 1}   item j: met / partial / missed
        w_j  ∈ [1, 3]        item weight, fixed with the rubric
        p_ij                 penalty charged against task i

Two consequences worth naming. Penalties are subtracted after the proportion is scaled, so a single invented fact costs the same 12 points whether the rest of the brief was excellent or threadbare. The benchmark treats fabrication as a floor violation, not a quality gradient. And because s_i is clipped at 0 rather than allowed to go negative, a catastrophic answer and a merely bad one are indistinguishable at the bottom of the scale.

02 · Who does the grading

Each rubric item is graded three times: twice by a model panel and once, on a 20% sample, by a working GTM operator. The model under test is excluded from the panel that grades it, so no model ever contributes to its own score. Where the panel and the human disagree, the item is re-graded by a second operator and the human label wins outright: the panel never overrules a person.

r_ij  =  human_ij                       if task i is in the 20% human sample
         mode(panel_ij)                 otherwise, if the two panel graders agree
         second_operator_ij             otherwise, the tie goes to a human

panel  =  2 models drawn from the grader pool, model under test excluded
agreement  =  95.9% panel-vs-human on the sampled items

The 4.1% disagreement rate is the number to be suspicious of first. It is low because most rubric items are near-mechanical (did the model cite a source for this claim, yes or no), and it concentrates in the two families where items are judgement calls: outbound copy and the end-to-end account plan. Read those two columns with wider error bars than the table draws.

03 · How precise the numbers actually are

A family score is a sample mean, so it carries a standard error of σ∕√n. Per-task rubric scores are dispersed with σ ≈ 18 points across the set; every figure in this table follows from that and the task counts, and nothing else.

FamilyTasks (n)Items/taskGraded judgementsStd. error95% CIHeadline penalty
Account research412145,768±0.89±1.74−12 per invented fact
Company and persona deep research288226,336±1.06±2.08−15 per uncited claim
Outbound copy468115,148±0.83±1.63spam-classifier gate
Call analysis386145,404±0.92±1.80exact-match, no partial credit
Objection handling34093,060±0.98±1.91−8 per concession past the floor
CRM extraction460125,520±0.84±1.64schema violation → 0
Sequence execution346165,536±0.97±1.90−10 per invalid tool call
Account plan246194,674±1.15±2.25operator head-to-head
GTM score (composite)2,946n/a41,446±0.34±0.66unweighted mean of the eight

The composite is sharper than any family that feeds it: averaging eight independent means divides the variance by eight, which is the entire reason the headline number is a mean and not a single flagship task family. The practical upshot is the last number in that row. Two models must differ by at least 0.9 points on the GTM score before the gap clears a 95% two-sample test. Anything tighter than that is a tie, however the rows happen to be ordered.

Why the mean is unweighted

A weighted composite would encode a claim about which GTM work matters most, and that claim is different for every revenue team. An unweighted mean makes the weighting the reader’s job: the eight family columns are published precisely so a team that lives in outbound can rank on that column and ignore ours. It also removes the most obvious lever for gaming the headline number.

Blended cost, and why 3:1

The cost column is (3·input + 1·output) ÷ 4 at list price. GTM work is read-heavy: a long account brief, a 40-minute transcript, a CRM thread in, a short email out, so a token-symmetric average would flatter models that charge little for input and a great deal for output. The ratio is a stated assumption, not a measurement; recompute it against your own traffic before trusting the ordering.

What the score does not capture

Single-shot, temperature 0, no per-model prompt tuning: this measures the first answer from an untuned model, which is the floor of what a team would deploy, not the ceiling. Latency, rate limits, refusal behaviour and fine-tuning headroom are not scored at all. A model two points down the table with half the p95 latency may be the better production choice, and this benchmark will not tell you that.

Submit a model

Labs and vendors can request an eval for the next monthly run (2026-08-14). Open-weight models are run at our cost; hosted models need an API key with a 40M-token allowance. Results publish whether they flatter you or not.

Cite this run

@misc{gtmbench2026,
  title  = {GTMBench: benchmarking language models on
            go-to-market work},
  author = {GTMBench Collective},
  year   = {2026},
  note   = {Run 2026.07, 2946 tasks, 30 models},
  url    = {https://gtmbench.org}
}