{ "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": 3882, "completed_calls": 7718, "contributors": 0, "probes": 7718, "crowdsourced": false, "probe_only": true, "p50_s": 74, "p90_s": 4709, "p95_s": 24325, "latest_at": 1791590478, "last_measured_age_s": 628, "answerable": true, "confidence": "very_low" }, { "provider": "openai", "model": "gpt-5-nano", "mode": "sync", "requested_service_tier": "flex", "n": 3979, "completed_calls": 4194, "contributors": 0, "probes": 4194, "crowdsourced": false, "probe_only": true, "p50_s": 2, "p90_s": 3, "p95_s": 3, "latest_at": 1791590921, "last_measured_age_s": 185, "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": 3403, "completed_calls": 4067, "contributors": 0, "probes": 4067, "crowdsourced": false, "probe_only": true, "p50_s": 1, "p90_s": 2, "p95_s": 2, "latest_at": 1791590923, "last_measured_age_s": 183, "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": 3019, "completed_calls": 3508, "contributors": 0, "probes": 3508, "crowdsourced": false, "probe_only": true, "p50_s": 1, "p90_s": 3, "p95_s": 3, "latest_at": 1791590919, "last_measured_age_s": 187, "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": 3246, "completed_calls": 3248, "contributors": 0, "probes": 3248, "crowdsourced": false, "probe_only": true, "p50_s": 86, "p90_s": 956, "p95_s": 2199, "latest_at": 1791590140, "last_measured_age_s": 966, "answerable": true, "confidence": "very_low" }, { "provider": "anthropic", "model": "claude-haiku-4-5", "mode": "batch", "requested_service_tier": "standard", "n": 2589, "completed_calls": 2839, "contributors": 0, "probes": 2839, "crowdsourced": false, "probe_only": true, "p50_s": 150, "p90_s": 409, "p95_s": 624, "latest_at": 1791590157, "last_measured_age_s": 949, "answerable": true, "confidence": "very_low" }, { "provider": "google", "model": "gemini-3.7-flash", "mode": "batch", "requested_service_tier": "standard", "n": 2494, "completed_calls": 2584, "contributors": 0, "probes": 2584, "crowdsourced": false, "probe_only": true, "p50_s": 124, "p90_s": 170, "p95_s": 190, "latest_at": 1791589016, "last_measured_age_s": 2090, "answerable": true, "confidence": "very_low" }, { "provider": "openai", "model": "gpt-5.6-sol", "mode": "batch", "requested_service_tier": "standard", "n": 1885, "completed_calls": 1913, "contributors": 0, "probes": 1913, "crowdsourced": false, "probe_only": true, "p50_s": 30, "p90_s": 1000, "p95_s": 4736, "latest_at": 1791587669, "last_measured_age_s": 3437, "answerable": true, "confidence": "very_low" } ] }