inference/research
verifiedhigh · 0.85updated 2026-10-17

OpenRouter hosted API router

openrouter

Fields

id
services/svc-openrouter
Provider
openrouter
Endpoint
https://openrouter.ai/api/v1
Input $/Mtok
0.59
Output $/Mtok
0.79
Price basis
USD per 1M TOKENS, representative model meta-llama/llama-3.3-70b-instruct ROUTED TO GROQ, re-verified 2026-10-17. OpenRouter is a router, not a hoster: the same model id fans out to DeepInfra, Novita, AkashML, Parasail, Cloudflare, SambaNova, Groq, CoreWeave, Google Vertex and Together, each with its own upstream price that OpenRouter passes through. The 2026-10-17 provider table still spans $0.10/$0.32 (DeepInfra Turbo) to $1.04/$1.04 (Together). price_input/price_output here stay the GROQ row ($0.59/$0.79) — the highest-throughput backend — not the floor. The page headline still prints the floor route ($0.10/$0.32) as 'the' price; that is a different claim and is named here rather than written into the scalar fields. Cache-read on Groq is still $0.295. SambaNova's list $0.60/$1.20 now renders struck through next to a 25% off badge at $0.45/$0.90; that discount is recorded on svc-sambanova-cloud, not folded into this Groq-row price. No single 'OpenRouter price' exists for a model; quote provider+model or nothing.
Context window
131,072
TTFT (ms)
520
Output tok/s
191
Measured by
self
Regions
global
notes
Fifth re-read later on 2026-10-17, Groq row on llama-3.3-70b-instruct: 0.52s P50 latency, 191 tok/s, 100.00% uptime. That replaces the fourth 2026-10-17 Groq row (0.51s, 198 tok/s), the third (0.53s, 201 tok/s), the second (0.49s, 209 tok/s), the earlier 2026-10-17 row (0.50s, 198 tok/s), the 2026-10-16 row (0.52s, 193 tok/s) and the 2026-10-14 row (0.40s, 233 tok/s). Same source, same model, same day: slightly slower and lower throughput than the previous capture. List prices did not move, including the Groq row this record's price fields are anchored to ($0.59/$0.79). Other rows that moved, not promoted into the scalar fields: DeepInfra Turbo 0.65s / 16 tps / 99.85% (was 1.71s / 6 tps / 99.92% on the fourth 2026-10-17 read); Novita 1.30s / 35 tps / 99.35% (was 1.05s / 41 tps / 99.29%); SambaNova Turbo 0.89s / 86 tps / 99.87% with the list price still struck through at 25% off $0.45/$0.90 (was 2.58s / 162 tps); CoreWeave 0.13s / 85 tps / $0.71/$0.71 (was 0.11s / 88 tps); Vertex US row 2.26s / 103 tps at $0.72/$0.72, and the second Vertex row has no latency cell again (was 1.22s / 96 tps, second row 1.40s / 96 tps); Together 0.94s / 28 tps / 96.72% at $1.04/$1.04 (was 0.55s / 41 tps / 96.68%). Best-across-providers widget: 191 tok/s (Groq), 0.13s (CoreWeave). OpenRouter's own routing telemetry is the measurement: its model pages publish per-provider P50 latency, throughput and uptime measured over real routed traffic. Fourth re-read later on 2026-10-17, Groq row on llama-3.3-70b-instruct: 0.51s P50 latency, 198 tok/s, 100.00% uptime. That replaces the third 2026-10-17 Groq row (0.53s, 201 tok/s), the second (0.49s, 209 tok/s), the earlier 2026-10-17 row (0.50s, 198 tok/s), the 2026-10-16 row (0.52s, 193 tok/s) and the 2026-10-14 row (0.40s, 233 tok/s). Same source, same model, same day: slightly faster and slightly lower throughput than the previous capture, back at the morning's 198 tok/s. List prices did not move, including the Groq row this record's price fields are anchored to ($0.59/$0.79). Other rows that moved, not promoted into the scalar fields: DeepInfra Turbo 1.71s / 6 tps / 99.92% (was 0.55s / 16 tps / 99.94% on the third 2026-10-17 read); Novita 1.05s / 41 tps / 99.29% (was 1.08s / 41 tps / 99.25%); SambaNova Turbo 2.58s / 162 tps / 99.86% with the list price still struck through at 25% off $0.45/$0.90 (was 0.78s / 110 tps); CoreWeave 0.11s / 88 tps / $0.71/$0.71 (was 0.16s / 86 tps); Vertex US row 1.22s / 96 tps at $0.72/$0.72, and the second Vertex row now has a cell at 1.40s / 96 tps (was 2.32s / 92 tps, second row empty); Together 0.55s / 41 tps / 96.68% at $1.04/$1.04 (was 0.71s / 39 tps / 96.30%). Best-across-providers widget: 198 tok/s (Groq), 0.11s (CoreWeave). OpenRouter's own routing telemetry is the measurement: its model pages publish per-provider P50 latency, throughput and uptime measured over real routed traffic. Third re-read later on 2026-10-17, Groq row on llama-3.3-70b-instruct: 0.53s P50 latency, 201 tok/s, 100.00% uptime. That replaces the second 2026-10-17 Groq row (0.49s, 209 tok/s), the earlier 2026-10-17 row (0.50s, 198 tok/s), the 2026-10-16 row (0.52s, 193 tok/s) and the 2026-10-14 row (0.40s, 233 tok/s). Same source, same model, same day: slower and lower throughput than the previous capture the same day, and still below the Oct-14 snapshot. The earlier pairs stay in this note so the drift is visible. measurement_source 'self' means measured-by-the-service-itself over production traffic, a different claim from both a vendor marketing number and the Artificial Analysis medians stored on the individual provider records (Groq's own record carries the AA figure of 305.7 tok/s, not this one). context_window is per-model-per-provider; 131072 is the llama-3.3-70b figure. Balanced/Nitro/Floor/Exacto routing modes pick different providers for the same request, so 'where does it run' is a routing decision, not a property of the endpoint. Other rows that moved on the same page, not promoted into the scalar fields: DeepInfra Turbo 0.55s / 16 tps / 99.94% (was 0.46s / 26 tps / 99.93% on the second 2026-10-17 read); Novita 1.08s / 41 tps / 99.25% (was 1.04s / 48 tps); SambaNova Turbo 0.78s / 110 tps / 99.88% with the list price still struck through at 25% off $0.45/$0.90 (was 0.96s / 117 tps); CoreWeave 0.16s / 86 tps / $0.71/$0.71 (was 0.13s / 85 tps). Vertex's US row: 2.32s / 92 tps at $0.72/$0.72 (was 1.24s / 89 tps). Together: 0.71s / 39 tps at $1.04/$1.04, 96.30% uptime (was 0.91s / 53 tps / 95.86%). List prices did not move, including the Groq row this record's price fields are anchored to ($0.59/$0.79). Billing goes through OpenRouter regardless of which upstream serves the request. OpenRouter's own routing telemetry is the measurement: its model pages publish per-provider P50 latency, throughput and uptime measured over real routed traffic. Second re-read later on 2026-10-17, Groq row on llama-3.3-70b-instruct: 0.49s P50 latency, 209 tok/s, 100.00% uptime. That replaces the earlier 2026-10-17 Groq row (0.50s, 198 tok/s, 100% uptime), the 2026-10-16 row (0.52s, 193 tok/s, 100% uptime) and the 2026-10-14 row (0.40s, 233 tok/s). Same source, same model, same day, faster than the morning capture and still ~10% slower than the Oct-14 snapshot. The earlier pairs stay in this note so the drift is visible. measurement_source 'self' means measured-by-the-service-itself over production traffic, a different claim from both a vendor marketing number and the Artificial Analysis medians stored on the individual provider records (Groq's own record carries the AA figure of 305.7 tok/s, not this one). context_window is per-model-per-provider; 131072 is the llama-3.3-70b figure. Balanced/Nitro/Floor/Exacto routing modes pick different providers for the same request, so 'where does it run' is a routing decision, not a property of the endpoint. Other rows that moved on the same page, not promoted into the scalar fields: DeepInfra Turbo 0.46s / 26 tps / 99.93% (was 0.62s / 19 tps / 99.94% earlier on 2026-10-17, 0.48s / 16 tps / 98.80% on 2026-10-16, and 0.51s / 14 tps on 2026-10-14); Novita 1.04s / 48 tps / 99.16% (was 1.17s / 45 tps); SambaNova Turbo 0.96s / 117 tps / 99.90% with the list price struck through at 25% off $0.45/$0.90 (was 1.01s / 93 tps); CoreWeave 0.13s / 85 tps / $0.71/$0.71 (was 0.13s / 86 tps). Vertex's US row: 1.24s / 89 tps at $0.72/$0.72 (was 1.19s / 92 tps). Together: 0.91s / 53 tps at $1.04/$1.04, 95.86% uptime (was 0.94s / 52 tps / 95.75%). Billing goes through OpenRouter regardless of which upstream serves the request. Earlier 2026-10-17 note, kept: OpenRouter's own routing telemetry is the measurement: its model pages publish per-provider P50 latency, throughput and uptime measured over real routed traffic. Re-read 2026-10-17, Groq row on llama-3.3-70b-instruct: 0.50s P50 latency, 198 tok/s, 100.00% uptime. That replaces the 2026-10-16 Groq row (0.52s, 193 tok/s, 100% uptime) and the 2026-10-14 row (0.40s, 233 tok/s). Same source, same model, one day later, slightly faster than yesterday and still ~15% slower than the Oct-14 snapshot. Both earlier pairs stay in this note so the drift is visible. measurement_source 'self' means measured-by-the-service-itself over production traffic, a different claim from both a vendor marketing number and the Artificial Analysis medians stored on the individual provider records (Groq's own record carries the AA figure of 305.7 tok/s, not this one). context_window is per-model-per-provider; 131072 is the llama-3.3-70b figure. Balanced/Nitro/Floor/Exacto routing modes pick different providers for the same request, so 'where does it run' is a routing decision, not a property of the endpoint. Other rows that moved on the same page, not promoted into the scalar fields: DeepInfra Turbo 0.62s / 19 tps / 99.94% (was 0.48s / 16 tps / 98.80% on 2026-10-16, and 0.51s / 14 tps on 2026-10-14); Novita 1.17s / 45 tps / 99.15% (was 1.05s / 44 tps); SambaNova Turbo 1.01s / 93 tps / 99.90% with the list price struck through at 25% off $0.45/$0.90 (was 0.81s / 104 tps); CoreWeave 0.13s / 86 tps / $0.71/$0.71 (was 0.12s / 85 tps). Vertex now has a latency cell on the US row: 1.19s / 92 tps at $0.72/$0.72 (the 2026-10-16 read had no cell). Together now has one too: 0.94s / 52 tps at $1.04/$1.04, 95.75% uptime. Billing goes through OpenRouter regardless of which upstream serves the request.

Notes

Fifth re-read later on 2026-10-17, Groq row on llama-3.3-70b-instruct: 0.52s P50 latency, 191 tok/s, 100.00% uptime. That replaces the fourth 2026-10-17 Groq row (0.51s, 198 tok/s), the third (0.53s, 201 tok/s), the second (0.49s, 209 tok/s), the earlier 2026-10-17 row (0.50s, 198 tok/s), the 2026-10-16 row (0.52s, 193 tok/s) and the 2026-10-14 row (0.40s, 233 tok/s). Same source, same model, same day: slightly slower and lower throughput than the previous capture. List prices did not move, including the Groq row this record's price fields are anchored to ($0.59/$0.79). Other rows that moved, not promoted into the scalar fields: DeepInfra Turbo 0.65s / 16 tps / 99.85% (was 1.71s / 6 tps / 99.92% on the fourth 2026-10-17 read); Novita 1.30s / 35 tps / 99.35% (was 1.05s / 41 tps / 99.29%); SambaNova Turbo 0.89s / 86 tps / 99.87% with the list price still struck through at 25% off $0.45/$0.90 (was 2.58s / 162 tps); CoreWeave 0.13s / 85 tps / $0.71/$0.71 (was 0.11s / 88 tps); Vertex US row 2.26s / 103 tps at $0.72/$0.72, and the second Vertex row has no latency cell again (was 1.22s / 96 tps, second row 1.40s / 96 tps); Together 0.94s / 28 tps / 96.72% at $1.04/$1.04 (was 0.55s / 41 tps / 96.68%). Best-across-providers widget: 191 tok/s (Groq), 0.13s (CoreWeave). OpenRouter's own routing telemetry is the measurement: its model pages publish per-provider P50 latency, throughput and uptime measured over real routed traffic. Fourth re-read later on 2026-10-17, Groq row on llama-3.3-70b-instruct: 0.51s P50 latency, 198 tok/s, 100.00% uptime. That replaces the third 2026-10-17 Groq row (0.53s, 201 tok/s), the second (0.49s, 209 tok/s), the earlier 2026-10-17 row (0.50s, 198 tok/s), the 2026-10-16 row (0.52s, 193 tok/s) and the 2026-10-14 row (0.40s, 233 tok/s). Same source, same model, same day: slightly faster and slightly lower throughput than the previous capture, back at the morning's 198 tok/s. List prices did not move, including the Groq row this record's price fields are anchored to ($0.59/$0.79). Other rows that moved, not promoted into the scalar fields: DeepInfra Turbo 1.71s / 6 tps / 99.92% (was 0.55s / 16 tps / 99.94% on the third 2026-10-17 read); Novita 1.05s / 41 tps / 99.29% (was 1.08s / 41 tps / 99.25%); SambaNova Turbo 2.58s / 162 tps / 99.86% with the list price still struck through at 25% off $0.45/$0.90 (was 0.78s / 110 tps); CoreWeave 0.11s / 88 tps / $0.71/$0.71 (was 0.16s / 86 tps); Vertex US row 1.22s / 96 tps at $0.72/$0.72, and the second Vertex row now has a cell at 1.40s / 96 tps (was 2.32s / 92 tps, second row empty); Together 0.55s / 41 tps / 96.68% at $1.04/$1.04 (was 0.71s / 39 tps / 96.30%). Best-across-providers widget: 198 tok/s (Groq), 0.11s (CoreWeave). OpenRouter's own routing telemetry is the measurement: its model pages publish per-provider P50 latency, throughput and uptime measured over real routed traffic. Third re-read later on 2026-10-17, Groq row on llama-3.3-70b-instruct: 0.53s P50 latency, 201 tok/s, 100.00% uptime. That replaces the second 2026-10-17 Groq row (0.49s, 209 tok/s), the earlier 2026-10-17 row (0.50s, 198 tok/s), the 2026-10-16 row (0.52s, 193 tok/s) and the 2026-10-14 row (0.40s, 233 tok/s). Same source, same model, same day: slower and lower throughput than the previous capture the same day, and still below the Oct-14 snapshot. The earlier pairs stay in this note so the drift is visible. measurement_source 'self' means measured-by-the-service-itself over production traffic, a different claim from both a vendor marketing number and the Artificial Analysis medians stored on the individual provider records (Groq's own record carries the AA figure of 305.7 tok/s, not this one). context_window is per-model-per-provider; 131072 is the llama-3.3-70b figure. Balanced/Nitro/Floor/Exacto routing modes pick different providers for the same request, so 'where does it run' is a routing decision, not a property of the endpoint. Other rows that moved on the same page, not promoted into the scalar fields: DeepInfra Turbo 0.55s / 16 tps / 99.94% (was 0.46s / 26 tps / 99.93% on the second 2026-10-17 read); Novita 1.08s / 41 tps / 99.25% (was 1.04s / 48 tps); SambaNova Turbo 0.78s / 110 tps / 99.88% with the list price still struck through at 25% off $0.45/$0.90 (was 0.96s / 117 tps); CoreWeave 0.16s / 86 tps / $0.71/$0.71 (was 0.13s / 85 tps). Vertex's US row: 2.32s / 92 tps at $0.72/$0.72 (was 1.24s / 89 tps). Together: 0.71s / 39 tps at $1.04/$1.04, 96.30% uptime (was 0.91s / 53 tps / 95.86%). List prices did not move, including the Groq row this record's price fields are anchored to ($0.59/$0.79). Billing goes through OpenRouter regardless of which upstream serves the request. OpenRouter's own routing telemetry is the measurement: its model pages publish per-provider P50 latency, throughput and uptime measured over real routed traffic. Second re-read later on 2026-10-17, Groq row on llama-3.3-70b-instruct: 0.49s P50 latency, 209 tok/s, 100.00% uptime. That replaces the earlier 2026-10-17 Groq row (0.50s, 198 tok/s, 100% uptime), the 2026-10-16 row (0.52s, 193 tok/s, 100% uptime) and the 2026-10-14 row (0.40s, 233 tok/s). Same source, same model, same day, faster than the morning capture and still ~10% slower than the Oct-14 snapshot. The earlier pairs stay in this note so the drift is visible. measurement_source 'self' means measured-by-the-service-itself over production traffic, a different claim from both a vendor marketing number and the Artificial Analysis medians stored on the individual provider records (Groq's own record carries the AA figure of 305.7 tok/s, not this one). context_window is per-model-per-provider; 131072 is the llama-3.3-70b figure. Balanced/Nitro/Floor/Exacto routing modes pick different providers for the same request, so 'where does it run' is a routing decision, not a property of the endpoint. Other rows that moved on the same page, not promoted into the scalar fields: DeepInfra Turbo 0.46s / 26 tps / 99.93% (was 0.62s / 19 tps / 99.94% earlier on 2026-10-17, 0.48s / 16 tps / 98.80% on 2026-10-16, and 0.51s / 14 tps on 2026-10-14); Novita 1.04s / 48 tps / 99.16% (was 1.17s / 45 tps); SambaNova Turbo 0.96s / 117 tps / 99.90% with the list price struck through at 25% off $0.45/$0.90 (was 1.01s / 93 tps); CoreWeave 0.13s / 85 tps / $0.71/$0.71 (was 0.13s / 86 tps). Vertex's US row: 1.24s / 89 tps at $0.72/$0.72 (was 1.19s / 92 tps). Together: 0.91s / 53 tps at $1.04/$1.04, 95.86% uptime (was 0.94s / 52 tps / 95.75%). Billing goes through OpenRouter regardless of which upstream serves the request. Earlier 2026-10-17 note, kept: OpenRouter's own routing telemetry is the measurement: its model pages publish per-provider P50 latency, throughput and uptime measured over real routed traffic. Re-read 2026-10-17, Groq row on llama-3.3-70b-instruct: 0.50s P50 latency, 198 tok/s, 100.00% uptime. That replaces the 2026-10-16 Groq row (0.52s, 193 tok/s, 100% uptime) and the 2026-10-14 row (0.40s, 233 tok/s). Same source, same model, one day later, slightly faster than yesterday and still ~15% slower than the Oct-14 snapshot. Both earlier pairs stay in this note so the drift is visible. measurement_source 'self' means measured-by-the-service-itself over production traffic, a different claim from both a vendor marketing number and the Artificial Analysis medians stored on the individual provider records (Groq's own record carries the AA figure of 305.7 tok/s, not this one). context_window is per-model-per-provider; 131072 is the llama-3.3-70b figure. Balanced/Nitro/Floor/Exacto routing modes pick different providers for the same request, so 'where does it run' is a routing decision, not a property of the endpoint. Other rows that moved on the same page, not promoted into the scalar fields: DeepInfra Turbo 0.62s / 19 tps / 99.94% (was 0.48s / 16 tps / 98.80% on 2026-10-16, and 0.51s / 14 tps on 2026-10-14); Novita 1.17s / 45 tps / 99.15% (was 1.05s / 44 tps); SambaNova Turbo 1.01s / 93 tps / 99.90% with the list price struck through at 25% off $0.45/$0.90 (was 0.81s / 104 tps); CoreWeave 0.13s / 86 tps / $0.71/$0.71 (was 0.12s / 85 tps). Vertex now has a latency cell on the US row: 1.19s / 92 tps at $0.72/$0.72 (the 2026-10-16 read had no cell). Together now has one too: 0.94s / 52 tps at $1.04/$1.04, 95.75% uptime. Billing goes through OpenRouter regardless of which upstream serves the request.

Cross-references

Models (8)