{ "window": "30d", "license": { "id": "CC-BY-4.0", "url": "https://creativecommons.org/licenses/by/4.0/", "attribution": "Measured by batchwatch - https://batchwatch.dev", "attribution_html": "Measured by batchwatch", "source_url": "https://batchwatch.dev/license" }, "models": [ { "provider": "openai", "model": "gpt-5-nano", "mode": "batch", "requested_service_tier": "standard", "n": 3863, "completed_calls": 7695, "contributors": 0, "probes": 7695, "crowdsourced": false, "probe_only": true, "p50_s": 73, "p90_s": 1776, "p95_s": 6655, "latest_at": 1792214141, "last_measured_age_s": 728, "answerable": true, "confidence": "very_low" }, { "provider": "openai", "model": "gpt-5-nano", "mode": "sync", "requested_service_tier": "flex", "n": 3938, "completed_calls": 4194, "contributors": 0, "probes": 4194, "crowdsourced": false, "probe_only": true, "p50_s": 2, "p90_s": 3, "p95_s": 3, "latest_at": 1792214574, "last_measured_age_s": 295, "answerable": true, "confidence": "very_low", "measurement_basis": { "measured_at_input_tokens": 8, "floor_clause": "Flex is measured at 8 input tokens and is synchronous, so that figure is the model's own work and a longer prompt makes it larger: read it as a floor. A batch figure is dominated by the wait to be scheduled, so the same extra work is a much smaller part of it.", "note": "Flex latency is measured on a fixed 8-input-token probe, the same fixed prompt size our batch probes use. A flex call is synchronous, so the figure is the model's own work and it grows with the prompt: read it as the floor a larger prompt starts from. A batch figure is dominated by queue time -- the wait to be scheduled -- so the same extra work is a much smaller part of it. The refusal rate is unaffected by prompt size: admission control happens before inference, so this probe measures it exactly as well as a large one would." } }, { "provider": "google", "model": "gemini-3.6-flash", "mode": "sync", "requested_service_tier": "flex", "n": 3312, "completed_calls": 3913, "contributors": 0, "probes": 3913, "crowdsourced": false, "probe_only": true, "p50_s": 1, "p90_s": 2, "p95_s": 2, "latest_at": 1792214576, "last_measured_age_s": 293, "answerable": true, "confidence": "very_low", "measurement_basis": { "measured_at_input_tokens": 8, "floor_clause": "Flex is measured at 8 input tokens and is synchronous, so that figure is the model's own work and a longer prompt makes it larger: read it as a floor. A batch figure is dominated by the wait to be scheduled, so the same extra work is a much smaller part of it.", "note": "Flex latency is measured on a fixed 8-input-token probe, the same fixed prompt size our batch probes use. A flex call is synchronous, so the figure is the model's own work and it grows with the prompt: read it as the floor a larger prompt starts from. A batch figure is dominated by queue time -- the wait to be scheduled -- so the same extra work is a much smaller part of it. The refusal rate is unaffected by prompt size: admission control happens before inference, so this probe measures it exactly as well as a large one would." } }, { "provider": "openai", "model": "gpt-5.6-sol", "mode": "sync", "requested_service_tier": "flex", "n": 3764, "completed_calls": 3768, "contributors": 0, "probes": 3768, "crowdsourced": false, "probe_only": true, "p50_s": 1, "p90_s": 6, "p95_s": 8, "latest_at": 1792214573, "last_measured_age_s": 296, "answerable": true, "confidence": "very_low", "measurement_basis": { "measured_at_input_tokens": 8, "floor_clause": "Flex is measured at 8 input tokens and is synchronous, so that figure is the model's own work and a longer prompt makes it larger: read it as a floor. A batch figure is dominated by the wait to be scheduled, so the same extra work is a much smaller part of it.", "note": "Flex latency is measured on a fixed 8-input-token probe, the same fixed prompt size our batch probes use. A flex call is synchronous, so the figure is the model's own work and it grows with the prompt: read it as the floor a larger prompt starts from. A batch figure is dominated by queue time -- the wait to be scheduled -- so the same extra work is a much smaller part of it. The refusal rate is unaffected by prompt size: admission control happens before inference, so this probe measures it exactly as well as a large one would." } }, { "provider": "openai", "model": "gpt-5.6-luna", "mode": "batch", "requested_service_tier": "standard", "n": 3230, "completed_calls": 3231, "contributors": 0, "probes": 3231, "crowdsourced": false, "probe_only": true, "p50_s": 79, "p90_s": 525, "p95_s": 1559, "latest_at": 1792214042, "last_measured_age_s": 827, "answerable": true, "confidence": "very_low" }, { "provider": "anthropic", "model": "claude-haiku-4-5", "mode": "batch", "requested_service_tier": "standard", "n": 2583, "completed_calls": 2807, "contributors": 0, "probes": 2807, "crowdsourced": false, "probe_only": true, "p50_s": 117, "p90_s": 252, "p95_s": 364, "latest_at": 1792213508, "last_measured_age_s": 1361, "answerable": true, "confidence": "very_low" }, { "provider": "google", "model": "gemini-3.7-flash", "mode": "batch", "requested_service_tier": "standard", "n": 2298, "completed_calls": 2314, "contributors": 0, "probes": 2314, "crowdsourced": false, "probe_only": true, "p50_s": 136, "p90_s": 182, "p95_s": 184, "latest_at": 1792213558, "last_measured_age_s": 1311, "answerable": true, "confidence": "very_low" }, { "provider": "openai", "model": "gpt-5.6-sol", "mode": "batch", "requested_service_tier": "standard", "n": 1648, "completed_calls": 1726, "contributors": 0, "probes": 1726, "crowdsourced": false, "probe_only": true, "p50_s": 51, "p90_s": 295, "p95_s": 802, "latest_at": 1792214107, "last_measured_age_s": 762, "answerable": true, "confidence": "very_low" } ] }