ZeroGPU sells to AI agents - no signup

The agent storefront at https://agents.zerogpu.ai/llms.txt publishes prices, payment instructions, and setup. Agents can purchase autonomously or with their human's approval, with a range of payment methods; an agent without a payment method can set one up there instantly, no signup. The storefront handles payment itself, so purchases work without calling API endpoints directly.

    Benchmark · 5,000 production prompts · 31 languages · GPT-5.5 gold labels

    zlm-v2-iab-classify-edge-enriched: multilingual IAB classification at 4–9× lower latency

    ZeroGPU's multilingual, enrichment-aware classification model was evaluated against gpt-5.4-nano on 5,000 real production prompts — 3,000 of them non-English across 31 languages — using GPT-5.5-generated gold labels over the IAB Content 2.2 and Audience 1.1 taxonomies. Zero request errors.

    Median latency
    238ms
    515 ms non-English · 4–9× faster than ~2.2 s
    Audience F1@5
    0.292
    vs gpt-5.4-nano 0.164 · stronger audience targeting
    Content top-1
    0.657
    within 11 points of gpt-5.4-nano (0.771)
    Coverage
    31langs
    5,000 prompts · 0 request errors

    01How this benchmark was measured

    Both models saw the same 5,000 prompts under identical conditions, scored against a single frozen gold set and timed from the same co-located worker.

    1
    5,000 real production prompts

    Sampled from live orchestration prompt logs (taskType=iab_classify), deduplicated, median ~62 characters of short ad-copy and headline text. 3,000 of the rows are non-English, spanning 31 languages.

    2
    GPT-5.5 gold labels

    Gold labels were produced by gpt-5.5, constrained to the exact IAB Content 2.2 (704 labels) and Audience 1.1 (1,567 labels) label space and validated locally. Gold reflects an LLM's judgement rather than certified human truth, so figures are best read as relative comparisons on identical data.

    3
    Co-located timing

    All requests were issued from a Fly worker in region iad, co-located with the classification API, so client network latency is negligible (wall − server ≈ 10–70 ms). Accuracy is measured over all 5,000 rows; latency separately at concurrency = 1 after warm-up.

    02IAB classification quality

    Top-1 = the model's first label is among the gold labels. F1@5 / precision@5 / recall@5 = set overlap over the returned lists. Tier-1 top-1 = the first tier-1 category matches gold. v2 lands within 11 points of gpt-5.4-nano on content top-1 and clearly ahead on audience targeting.

    Metriczlm-v2-iab-classify-edge-enriched ZLM v2gpt-5.4-nano
    Content top-1
    		0.6570.771
    Content F1@5
    		0.4380.510
    Content precision@5
    		0.4080.475
    Content recall@5
    		0.4940.567
    Tier-1 top-1
    		0.5740.674
    Audience F1@5
    		0.2920.164

    Audience F1@5 is 78% higher than gpt-5.4-nano (0.292 vs 0.164), the metric that drives audience segmentation and targeting quality.

    03English vs non-English

    The multilingual path holds up on foreign-language traffic: v2 scores 0.647 top-1 on non-English content versus 0.671 on English — a 2.4-point spread.

    MetricSet ZLM v2gpt-5.4-nano
    Content top-1English0.6710.750
    Content top-1Non-English0.6470.785
    Content F1@5English0.4510.475
    Content F1@5Non-English0.4290.533
    Audience F1@5English0.3030.171
    Audience F1@5Non-English0.2850.160

    Content top-1 by language

    LanguageZLM v2 top-1 ZLM v2gpt-5.4-nano
    English
    		0.6710.750
    French
    		0.6020.759
    German
    		0.6190.757
    Italian
    		0.8440.908
    Spanish
    		0.5870.741
    Hebrew
    		0.5490.750
    Dutch
    		0.5780.656
    Portuguese
    		0.5860.741

    04Accuracy by content area

    Each prompt is bucketed by its gold top-1 tier-1 category. Buckets with at least 25 prompts, sorted by volume. n = prompts in the bucket. Green bars mark the areas where v2 leads on top-1.

    CategorynZLM v2 top-1 ZLM v2GPT top-1 ZLM F1@5GPT F1@5
    Real Estate592
    		0.8360.9210.6550.730
    Medical Health503
    		0.7380.8410.4400.529
    Business and Finance396
    		0.6460.7950.3980.477
    Personal Finance358
    		0.6820.6790.4610.472
    Home & Garden279
    		0.6420.7880.4610.469
    Pop Culture252
    		0.4360.6310.3370.436
    Healthy Living241
    		0.6350.8130.4570.553
    News and Politics225
    		0.7070.8220.3910.474
    Style & Fashion216
    		0.5560.6160.3990.476
    Automotive211
    		0.6210.7390.4070.475
    Technology & Computing210
    		0.7670.8190.5240.493
    Travel198
    		0.7020.8280.4260.491
    Shopping136
    		0.5070.6910.3970.509
    Sports126
    		0.7860.7940.4120.500
    Family and Relationships114
    		0.6140.7720.4580.476
    Pets111
    		0.7570.7390.4820.400
    Food & Drink108
    		0.7500.8240.5060.546
    Education97
    		0.6600.8350.4020.503
    Hobbies & Interests75
    		0.5600.6400.3430.481
    Events and Attractions59
    		0.6100.6780.3640.439
    Careers51
    		0.3920.7650.3020.469
    Science40
    		0.7250.8000.4740.506
    Television39
    		0.5380.6920.2650.549
    Books and Literature37
    		0.4600.7570.3090.482
    Movies36
    		0.8610.8330.3640.566
    Fine Art29
    		0.5170.6550.3010.493

    05Latency

    Measured warm, one request at a time (concurrency = 1) — the latency a single production caller experiences. ZLM figures are the server-reported processing_time_ms; gpt-5.4-nano has no server timing, so its client wall-clock from the same co-located worker is used.

    ModelSetp50 p90p95mean
    ZLM v2 edge-enrichedEnglish238 ms348 ms370 ms243 ms
    ZLM v2 edge-enrichedNon-English515 ms856 ms1061 ms548 ms
    gpt-5.4-nanoEnglish2232 ms3169 ms3505 ms3381 ms
    gpt-5.4-nanoNon-English2173 ms3107 ms3656 ms2510 ms

    Median speed-up

    English
    9×
    238 ms vs 2,232 ms
    Non-English
    4×
    515 ms vs 2,173 ms
    Request errors
    0
    across all 5,000 prompts
    ZLM v2 responds in 238 ms (English) / 515 ms (non-English) at the median, versus ~2.2 s for gpt-5.4-nano — a 4–9× latency advantage while trailing GPT on top-1 accuracy by a small margin.

    The bottom line

    • 4–9× faster at the median than gpt-5.4-nano, on both English and non-English traffic.
    • Audience F1@5 0.292 vs 0.164 — materially stronger audience signal extraction.
    • Content top-1 0.657 vs 0.771 — within 11 points of gpt-5.4-nano on the same gold labels.
    • Consistent across languages: 0.647 top-1 on non-English vs 0.671 on English, across 31 languages.
    • Leads gpt-5.4-nano on top-1 in Personal Finance, Pets and Movies, and on F1@5 in Technology & Computing and Pets.

    Notes & limitations

    Fair-reading notes. gpt-5.4-nano is included as a reference point and remains ahead on content top-1 in most categories. Gold labels are GPT-5.5 judgements rather than certified human truth, so the absolute numbers should be read as relative comparisons on identical data. Latency for gpt-5.4-nano is client wall-clock from a co-located worker, while ZLM figures are server-reported processing time.
    Run details
    Test set5,000 real production prompts · 3,000 non-English across 31 languages · median ~62 chars
    TaxonomiesIAB Content 2.2 (704 labels) · IAB Audience 1.1 (1,567 labels)
    Gold labelsgpt-5.5, constrained to the exact label space, validated locally
    ZeroGPU modelzlm-v2-iab-classify-edge-enriched — multilingual, content + audience
    Reference modelgpt-5.4-nano
    EnvironmentFly worker, region iad, co-located with the classification API
    Latency protocolwarm, concurrency = 1, split by English / non-English
    Date2026-08-03