{ "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": 3888, "completed_calls": 7741, "contributors": 0, "probes": 7741, "crowdsourced": false, "probe_only": true, "p50_s": 50, "p90_s": 1718, "p95_s": 6612, "latest_at": 1792763036, "last_measured_age_s": 489, "answerable": true, "confidence": "very_low" }, { "provider": "openai", "model": "gpt-5-nano", "mode": "sync", "requested_service_tier": "flex", "n": 3945, "completed_calls": 4196, "contributors": 0, "probes": 4196, "crowdsourced": false, "probe_only": true, "p50_s": 2, "p90_s": 3, "p95_s": 3, "latest_at": 1792763162, "last_measured_age_s": 363, "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": 3668, "completed_calls": 4059, "contributors": 0, "probes": 4059, "crowdsourced": false, "probe_only": true, "p50_s": 1, "p90_s": 2, "p95_s": 3, "latest_at": 1792763160, "last_measured_age_s": 365, "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": 3465, "completed_calls": 3952, "contributors": 0, "probes": 3952, "crowdsourced": false, "probe_only": true, "p50_s": 2, "p90_s": 3, "p95_s": 3, "latest_at": 1792763164, "last_measured_age_s": 361, "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": 3333, "completed_calls": 3337, "contributors": 0, "probes": 3337, "crowdsourced": false, "probe_only": true, "p50_s": 61, "p90_s": 233, "p95_s": 1154, "latest_at": 1792762494, "last_measured_age_s": 1031, "answerable": true, "confidence": "very_low" }, { "provider": "anthropic", "model": "claude-haiku-4-5", "mode": "batch", "requested_service_tier": "standard", "n": 2582, "completed_calls": 2797, "contributors": 0, "probes": 2797, "crowdsourced": false, "probe_only": true, "p50_s": 114, "p90_s": 204, "p95_s": 204, "latest_at": 1792762562, "last_measured_age_s": 963, "answerable": true, "confidence": "very_low" }, { "provider": "google", "model": "gemini-3.7-flash", "mode": "batch", "requested_service_tier": "standard", "n": 2158, "completed_calls": 2170, "contributors": 0, "probes": 2170, "crowdsourced": false, "probe_only": true, "p50_s": 127, "p90_s": 166, "p95_s": 175, "latest_at": 1792762607, "last_measured_age_s": 918, "answerable": true, "confidence": "very_low" }, { "provider": "openai", "model": "gpt-5.6-sol", "mode": "batch", "requested_service_tier": "standard", "n": 1730, "completed_calls": 1794, "contributors": 0, "probes": 1794, "crowdsourced": false, "probe_only": true, "p50_s": 50, "p90_s": 55, "p95_s": 123, "latest_at": 1792758336, "last_measured_age_s": 5189, "answerable": true, "confidence": "very_low" } ] }