SambaNova Cloud inference

2026-10-17 fifth re-read of OpenRouter's SambaNova Turbo row for llama-3.3-70b-instruct: 0.89s P50, 86 tok/s, 99.87% uptime, with the $0.60/$1.20 list still struck through at 25% off $0.45/$0.90. The fourth 2026-10-17 read of the same row was 2.58s / 162 tps / 99.86%, the third was 0.78s / 110 tps / 99.88%, the second was 0.96s / 117 tps / 99.90%, and the earlier one was 1.01s / 93 tps / 99.90%. The first-party list price is what the scalar fields keep; this row is the same disagreement, not a new one. measured_* stays the Artificial Analysis long-prompt median (286.3 tok/s, 2.06s first chunk), re-confirmed unchanged this pass from the page summary, a different workload from the OpenRouter P50. 2026-10-17 fourth re-read of OpenRouter's SambaNova Turbo row for llama-3.3-70b-instruct: 2.58s P50, 162 tok/s, 99.86% uptime, with the $0.60/$1.20 list still struck through at 25% off $0.45/$0.90. The third 2026-10-17 read of the same row was 0.78s / 110 tps / 99.88%, the second was 0.96s / 117 tps / 99.90%, and the earlier one was 1.01s / 93 tps / 99.90%. The first-party list price is what the scalar fields keep; this row is the same disagreement, not a new one. measured_* stays the Artificial Analysis long-prompt median (286.3 tok/s, 2.06s first chunk), which was not re-fetched this pass, a different workload from the OpenRouter P50. 2026-10-17 third re-read of OpenRouter's SambaNova Turbo row for llama-3.3-70b-instruct: 0.78s P50, 110 tok/s, 99.88% uptime, with the $0.60/$1.20 list still struck through at 25% off $0.45/$0.90. The second 2026-10-17 read of the same row was 0.96s / 117 tps / 99.90%, and the earlier one was 1.01s / 93 tps / 99.90%. The first-party list price is what the scalar fields keep; this row is the same disagreement, not a new one. measured_* stays the Artificial Analysis long-prompt median (286.3 tok/s, 2.06s first chunk), re-confirmed unchanged later the same day (summary 286.3; the comparison table still prints 288), a different workload from the OpenRouter P50. 2026-10-17 later re-read of OpenRouter's SambaNova Turbo row for llama-3.3-70b-instruct: 0.96s P50, 117 tok/s, 99.90% uptime, with the $0.60/$1.20 list still struck through at 25% off $0.45/$0.90. The earlier 2026-10-17 read of the same row was 1.01s / 93 tps / 99.90%. The first-party list price is what the scalar fields keep; this row is the same disagreement, not a new one. measured_* stays the Artificial Analysis long-prompt median (286.3 tok/s, 2.06s first chunk, re-confirmed unchanged later the same day), a different workload from the OpenRouter P50. 2026-10-17 Artificial Analysis re-read. The provider page's embedded per-provider JSON (byPromptType.long) now gives SambaNova 286.27 output tok/s at 2.055s median first chunk, down from the 288.5 tok/s recorded on 2026-10-16. TTFT is unchanged, so measured_ttft_ms stays 2060, and SambaNova stays second only to Groq (305.7) on that page. The page disagrees with itself: the summary prose says 286.3 while the comparison table still prints 288, so the scalar takes the embedded JSON rounded to one decimal and 288.5 stays in this note. The price contest is unchanged: the same JSON restates the first-party list of $0.60/$1.20, which cloud.sambanova.ai/plans/pricing still prints today, while OpenRouter's SambaNova Turbo row still strikes it through at $0.45/$0.90. Status stays contested and the scalar price fields stay on the first-party list. 2026-10-17 second read, contested. SambaNova's own pricing page still lists Meta-Llama-3.3-70B-Instruct at $0.60/$1.20 (identical to the earlier same-day capture and to 2026-10-16). OpenRouter's provider table the same day strikes that list through and shows $0.45/$0.90 under a 25% off badge, with P50 latency 1.01s, 93 tok/s and 99.90% uptime (was 0.81s / 104 tok/s / 99.83% on 2026-10-16). The discount is not written into the price fields: it is a router-side badge, not a number on SambaNova's rate card, and the two pages disagree. measured_tps stays the 2026-10-16 Artificial Analysis median of 288.5 tok/s (2.06s first chunk), not the OpenRouter P50 of 93 — those are different workloads. The OpenRouter 2026-10-14 P50 was 0.77s / 98 tok/s. RDU-based (SN40L reconfigurable dataflow), the non-GPU high-throughput option in this set. SambaNova also sells on-prem RDU systems; those are supply-type records. Prior note follows. 2026-10-17 re-read of cloud.sambanova.ai/plans/pricing: Meta-Llama-3.3-70B-Instruct still $0.60 in / $1.20 out, no cached-input rate. gpt-oss-120b $0.22/$0.59, gemma-4-31B-it $0.38/$1.15, DeepSeek-V3.1 and V3.2 $3.00/$4.50, MiniMax-M3 $0.60/$2.40 with cached input $0.06. Identical to the 2026-10-16 capture, so price fields are not moved. Speed stays the 2026-10-16 Artificial Analysis median (288.5 tok/s, 2.06s first chunk). 2026-10-16 refresh. Price held: SambaNova's own pricing page still lists Llama 3.3 70B at $0.60/$1.20, the same number the 2026-10-14 OpenRouter provider row carried. Speed did not hold. The previous measured_tps of 98 was OpenRouter's P50 over its own routed traffic on 2026-10-14; Artificial Analysis's Llama 3.3 70B provider page fetched 2026-10-16 measures SambaNova at 288.5 output tok/s median and 2.06s median time to first chunk, second only to Groq (305.7 tok/s) on that page. The two numbers describe different workloads (OpenRouter production mix vs AA's fixed benchmark), so the old 98 tok/s is recorded here rather than deleted: it is not a contradiction of the price, it is a different measurement. measured_ttft_ms was null before this refresh. Context window 128k is the AA comparison-table figure for SambaNova on this model. RDU-based (SN40L reconfigurable dataflow), the non-GPU high-throughput option in this set. Measured values are OpenRouter's 2026-10-14 P50 on SambaNova Turbo's llama-3.3-70b endpoint: 0.77s latency, 98 tok/s, 98.52% uptime - second-fastest on that table after Groq, roughly 4x DeepInfra-Turbo speed at ~6x DeepInfra-Turbo price. Catalog is small (6 models priced on the page) but covers the open-weight frontier. SambaNova also sells on-prem RDU systems; those are supply-type records.