inference/research
contestedmedium · 0.78updated 2026-10-17

Novita AI serverless endpoints

other

Fields

id
services/svc-novita
Provider
other
Endpoint
https://api.novita.ai/v3/openai
Input $/Mtok
0.135
Output $/Mtok
0.4
Price basis
USD per 1M TOKENS, representative model meta-llama/llama-3.3-70b-instruct, re-verified 2026-10-17 on Novita's own model page and pricing table: $0.135 in / $0.40 out, Context Length 12K, Max Output 12K, bf16, serverless. The same split was the 2026-10-14 OpenRouter Novita row. Artificial Analysis's FAQ the week of 2026-10-16 states $0.14/$0.40; output matches and input differs by half a cent, so the price stays contested (see notes) rather than being overwritten. Novita's 2026-10-14 pricing page puts the cheap end lower still on other families (GLM-4.7-Flash $0.07/$0.40, ERNIE-4.5-21B-A3B $0.07/$0.28). Batch API 50% off both directions on supported models. provider='other' per the enum; the record name carries the provider.
Context window
12,288
TTFT (ms)
1,770
Output tok/s
45
Measured by
artificial-analysis
Regions
usapac
notes
2026-10-17 fifth re-read of OpenRouter's NovitaAI row for llama-3.3-70b-instruct: 1.30s P50, 35 tok/s, 99.35% uptime, still $0.135/$0.400. The fourth 2026-10-17 read of the same row was 1.05s / 41 tps / 99.29%, the third was 1.08s / 41 tps / 99.25%, the second was 1.04s / 48 tps / 99.16%, and the earlier one was 1.17s / 45 tps / 99.15%. measured_* stays the Artificial Analysis long-prompt median (45 tok/s, 1.77s first chunk), re-confirmed unchanged this pass (FAQ still $0.14 in / $0.40 out, so the half-cent contest with Novita's own $0.135 stays); the OpenRouter P50 is a different workload and is not promoted. 2026-10-17 fourth re-read of OpenRouter's NovitaAI row for llama-3.3-70b-instruct: 1.05s P50, 41 tok/s, 99.29% uptime, still $0.135/$0.400. The third 2026-10-17 read of the same row was 1.08s / 41 tps / 99.25%, the second was 1.04s / 48 tps / 99.16%, and the earlier one was 1.17s / 45 tps / 99.15%. measured_* stays the Artificial Analysis long-prompt median (45 tok/s, 1.77s first chunk), which was not re-fetched this pass; the OpenRouter P50 is a different workload and is not promoted. 2026-10-17 third re-read of OpenRouter's NovitaAI row for llama-3.3-70b-instruct: 1.08s P50, 41 tok/s, 99.25% uptime, still $0.135/$0.400. The second 2026-10-17 read of the same row was 1.04s / 48 tps / 99.16%, and the earlier one was 1.17s / 45 tps / 99.15%. measured_* stays the Artificial Analysis long-prompt median (45 tok/s, 1.77s first chunk), re-confirmed unchanged later the same day; the OpenRouter P50 is a different workload and is not promoted. 2026-10-17 later re-read of OpenRouter's NovitaAI row for llama-3.3-70b-instruct: 1.04s P50, 48 tok/s, 99.16% uptime, still $0.135/$0.400. The earlier 2026-10-17 read of the same row was 1.17s / 45 tps / 99.15%. measured_* stays the Artificial Analysis long-prompt median (45.3 tok/s, 1.77s first chunk, re-confirmed unchanged later the same day); the OpenRouter P50 is a different workload and is not promoted. 2026-10-17 OpenRouter re-read agrees with Novita's own price: $0.135/$0.40, P50 1.17s, 45 tok/s, 99.15% uptime (was 1.05s / 44 tok/s / 98.40% on 2026-10-16). The half-cent disagreement with Artificial Analysis ($0.14/$0.40) is unchanged, so status stays contested and the price fields stay on Novita's own page. context_window stays 12288, the cap Novita exposes, not OpenRouter's native window. 2026-10-17 context correction, price still contested. Novita's model page and pricing table both print Context Length 12K (12288) for meta-llama/llama-3.3-70b-instruct, and the sample request sets max_tokens=12288. The previous context_window of 131072 was Artificial Analysis's comparison-table cell and OpenRouter's native-window figure, not the cap Novita exposes on this serverless endpoint. 12288 is what a caller can actually send. Price fields stay $0.135/$0.40 because that is Novita's own page, agreeing with the 2026-10-14 OpenRouter row. Artificial Analysis's FAQ states $0.14/$0.40: output matches, input differs by half a cent, so status stays contested with both values recorded. AA speed (45 tok/s median, 1.77s first chunk) is kept; the OpenRouter 2026-10-14 P50 was 0.58s / 41 tok/s. Novita undercuts Together about 7.7x on this model ($0.135 vs $1.04 input) at mid-pack speed, which is the cheap-API outlier this record type exists to surface. The 12K window is the trade next to the price: DeepInfra serves the same model at $0.10/$0.32 with a 131072 context. 200+ models, plus GPU-instance rental that belongs in supply records. OpenRouter's table had Novita at 97.77% 3d uptime versus Groq at 100%.

Notes

2026-10-17 fifth re-read of OpenRouter's NovitaAI row for llama-3.3-70b-instruct: 1.30s P50, 35 tok/s, 99.35% uptime, still $0.135/$0.400. The fourth 2026-10-17 read of the same row was 1.05s / 41 tps / 99.29%, the third was 1.08s / 41 tps / 99.25%, the second was 1.04s / 48 tps / 99.16%, and the earlier one was 1.17s / 45 tps / 99.15%. measured_* stays the Artificial Analysis long-prompt median (45 tok/s, 1.77s first chunk), re-confirmed unchanged this pass (FAQ still $0.14 in / $0.40 out, so the half-cent contest with Novita's own $0.135 stays); the OpenRouter P50 is a different workload and is not promoted. 2026-10-17 fourth re-read of OpenRouter's NovitaAI row for llama-3.3-70b-instruct: 1.05s P50, 41 tok/s, 99.29% uptime, still $0.135/$0.400. The third 2026-10-17 read of the same row was 1.08s / 41 tps / 99.25%, the second was 1.04s / 48 tps / 99.16%, and the earlier one was 1.17s / 45 tps / 99.15%. measured_* stays the Artificial Analysis long-prompt median (45 tok/s, 1.77s first chunk), which was not re-fetched this pass; the OpenRouter P50 is a different workload and is not promoted. 2026-10-17 third re-read of OpenRouter's NovitaAI row for llama-3.3-70b-instruct: 1.08s P50, 41 tok/s, 99.25% uptime, still $0.135/$0.400. The second 2026-10-17 read of the same row was 1.04s / 48 tps / 99.16%, and the earlier one was 1.17s / 45 tps / 99.15%. measured_* stays the Artificial Analysis long-prompt median (45 tok/s, 1.77s first chunk), re-confirmed unchanged later the same day; the OpenRouter P50 is a different workload and is not promoted. 2026-10-17 later re-read of OpenRouter's NovitaAI row for llama-3.3-70b-instruct: 1.04s P50, 48 tok/s, 99.16% uptime, still $0.135/$0.400. The earlier 2026-10-17 read of the same row was 1.17s / 45 tps / 99.15%. measured_* stays the Artificial Analysis long-prompt median (45.3 tok/s, 1.77s first chunk, re-confirmed unchanged later the same day); the OpenRouter P50 is a different workload and is not promoted. 2026-10-17 OpenRouter re-read agrees with Novita's own price: $0.135/$0.40, P50 1.17s, 45 tok/s, 99.15% uptime (was 1.05s / 44 tok/s / 98.40% on 2026-10-16). The half-cent disagreement with Artificial Analysis ($0.14/$0.40) is unchanged, so status stays contested and the price fields stay on Novita's own page. context_window stays 12288, the cap Novita exposes, not OpenRouter's native window. 2026-10-17 context correction, price still contested. Novita's model page and pricing table both print Context Length 12K (12288) for meta-llama/llama-3.3-70b-instruct, and the sample request sets max_tokens=12288. The previous context_window of 131072 was Artificial Analysis's comparison-table cell and OpenRouter's native-window figure, not the cap Novita exposes on this serverless endpoint. 12288 is what a caller can actually send. Price fields stay $0.135/$0.40 because that is Novita's own page, agreeing with the 2026-10-14 OpenRouter row. Artificial Analysis's FAQ states $0.14/$0.40: output matches, input differs by half a cent, so status stays contested with both values recorded. AA speed (45 tok/s median, 1.77s first chunk) is kept; the OpenRouter 2026-10-14 P50 was 0.58s / 41 tok/s. Novita undercuts Together about 7.7x on this model ($0.135 vs $1.04 input) at mid-pack speed, which is the cheap-API outlier this record type exists to surface. The 12K window is the trade next to the price: DeepInfra serves the same model at $0.10/$0.32 with a 131072 context. 200+ models, plus GPU-instance rental that belongs in supply records. OpenRouter's table had Novita at 97.77% 3d uptime versus Groq at 100%.

Cross-references

Models (7)