inference/research
verifiedhigh · 0.90updated 2026-10-17

DeepInfra serverless inference

deepinfra

Fields

id
services/svc-deepinfra
Provider
deepinfra
Endpoint
https://api.deepinfra.com/v1/openai
Input $/Mtok
0.1
Output $/Mtok
0.32
Price basis
USD per 1M TOKENS, representative model meta-llama/Llama-3.3-70B-Instruct-Turbo (FP8), re-pointed 2026-10-17. DeepInfra's public models list that day returns input_tokens 0.10, output_tokens 0.32, context_length 131072; the model page prints the same $0.10/$0.32 and a Flex tier at 0.8x ($0.08/$0.256). The pricing page prints the window as 128k, which is the same cap rounded. The previous representative row, deepseek-ai/DeepSeek-V3.1 at $0.25 in / $0.13 cached / $0.95 out, is still the live catalog price (models list context_length 163840; pricing page 160k) and is kept as a second data point rather than the headline, because the shared comparison model across this record set is Llama 3.3 70B. DeepInfra remains the low end: same-day pricing page has Qwen3-235B at $0.09/$0.55 and DeepSeek-V4-Flash-0731 at $0.06/$0.18. Some models bill per inference-execution second instead of per token; the token-priced subset is what the per-mtok fields describe. Account cap: 200 concurrent requests.
Context window
131,072
TTFT (ms)
2,140
Output tok/s
14.9
Measured by
artificial-analysis
Regions
us
notes
2026-10-17 fifth re-read of OpenRouter's DeepInfra (Turbo) row for llama-3.3-70b-instruct: 0.65s P50, 16 tok/s, 99.85% uptime, still $0.10/$0.32. The fourth 2026-10-17 read of the same row was 1.71s / 6 tps / 99.92%, the third was 0.55s / 16 tps / 99.94%, the second was 0.46s / 26 tps / 99.93%, and the earlier one was 0.62s / 19 tps / 99.94%. measured_* stays the Artificial Analysis long-prompt median (14.9 tok/s, 2.15s first chunk), re-confirmed unchanged this pass from the page summary (14.9 t/s, 20.5x slower than Groq at 305.7); the OpenRouter P50 is a different workload and is not promoted. 2026-10-17 fourth re-read of OpenRouter's DeepInfra (Turbo) row for llama-3.3-70b-instruct: 1.71s P50, 6 tok/s, 99.92% uptime, still $0.10/$0.32. The third 2026-10-17 read of the same row was 0.55s / 16 tps / 99.94%, the second was 0.46s / 26 tps / 99.93%, and the earlier one was 0.62s / 19 tps / 99.94%. measured_* stays the Artificial Analysis long-prompt median (14.9 tok/s, 2.15s 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 DeepInfra (Turbo) row for llama-3.3-70b-instruct: 0.55s P50, 16 tok/s, 99.94% uptime, still $0.10/$0.32. The second 2026-10-17 read of the same row was 0.46s / 26 tps / 99.93%, and the earlier one was 0.62s / 19 tps / 99.94%. measured_* stays the Artificial Analysis long-prompt median (14.9 tok/s, 2.15s first chunk), re-confirmed unchanged later the same day from the page summary (14.9 t/s; the comparison table still prints 15 / 2.14); the OpenRouter P50 is a different workload and is not promoted. 2026-10-17 later re-read of OpenRouter's DeepInfra (Turbo) row for llama-3.3-70b-instruct: 0.46s P50, 26 tok/s, 99.93% uptime, still $0.10/$0.32. The earlier 2026-10-17 read of the same row was 0.62s / 19 tps / 99.94%. measured_* stays the Artificial Analysis long-prompt median (14.9 tok/s, 2.15s first chunk, re-confirmed unchanged later the same day from the page's embedded byPromptType.long JSON); the OpenRouter P50 is a different workload and is not promoted. 2026-10-17 Artificial Analysis re-read. The provider page's embedded per-provider JSON (byPromptType.long) now gives DeepInfra (Turbo, FP8) 14.89 output tok/s at 2.146s median first chunk, down from the 15.2 tok/s recorded on 2026-10-16. TTFT is unchanged, so measured_ttft_ms stays 2140. The page disagrees with itself: the summary prose says 14.9 while the comparison table still prints 15, so the scalar takes the embedded JSON rounded to one decimal and 15.2 stays in this note. Price is untouched: the same JSON restates $0.10 in / $0.32 out, matching DeepInfra's own model page, models API and pricing table. DeepInfra stays the slowest of the providers on that page (Groq 305.7). OpenRouter's DeepInfra (Turbo) P50 the same day (0.62s, 19 tok/s) is a different quantity and is not promoted. 2026-10-17 OpenRouter telemetry only; price not moved. DeepInfra's own pages were not re-extracted this pass, so $0.10/$0.32 and context 131072 stay the 2026-10-17 first-party figures. OpenRouter's DeepInfra (Turbo) row the same day reads 0.62s P50, 19 tok/s, 99.94% uptime at the same $0.10/$0.32 (was 0.48s / 16 tok/s / 98.80% on 2026-10-16). measured_* stays the 2026-10-16 Artificial Analysis median (15.2 tok/s, 2.14s first chunk): the OpenRouter P50 is a different quantity and is recorded here rather than promoted. 2026-10-17: representative price moved off DeepSeek-V3.1 ($0.25/$0.95, still live) onto Llama-3.3-70B-Instruct-Turbo at $0.10/$0.32, context 131072. Three DeepInfra pages agree (models API, model page, pricing table), and the split matches the 2026-10-16 Artificial Analysis FAQ. The 2026-10-16 pass left the headline on V3.1 because the pricing page did not re-extract cleanly that day; it did on 2026-10-17. Speed is unchanged from the 2026-10-16 Artificial Analysis read: 15.2 output tok/s, 2.14s median first chunk, the slowest of the 11 providers on that page (Groq 305.7). The older OpenRouter P50 (0.51s, 14 tok/s, 99.52% uptime on 2026-10-14) is a different quantity and stays in the history below. 15.2 tps is shared-infrastructure Turbo, not a dedicated deployment. DeepInfra's own page states all models run on H100 or A100. Two tiers appear on provider tables (Turbo vs base). 2026-10-16: Artificial Analysis confirms the floor. Its Llama 3.3 70B FAQ states DeepInfra (Turbo, FP8) at $0.10 per 1M input and $0.32 per 1M output — the same split the 2026-10-14 OpenRouter table had — with a blended $0.12 per 1M, the lowest of 11 providers, and 15.2 output tok/s (20x slower than Groq's 305.7 on the same page) at 2.14s median first chunk. measured_tps moved from 14 (OpenRouter P50) to 15.2 (AA median); measured_ttft_ms moved from 510 to 2140 because the two measurements are not the same quantity (OpenRouter 'latency' vs AA 'time to first chunk'). The representative price fields stay on DeepSeek-V3.1 at $0.25/$0.95 as recorded 2026-10-14: that number was read off DeepInfra's own pricing page, which did not re-extract cleanly on 2026-10-16, so it is not silently overwritten. Measured figures are the OpenRouter 2026-10-14 P50 for DeepInfra (Turbo) on llama-3.3-70b-instruct: 0.51s latency, 14 tok/s, 99.52% uptime - the cheapest and the SLOWEST row on that table, which is exactly the price-vs-speed trade this record type exists to expose. 14 tps is shared-infrastructure Turbo, not a dedicated deployment. DeepInfra's own page states all models run on H100 or A100 GPUs. Two tiers appear on provider tables: standard and Turbo (cheaper, sometimes FP8); Artificial Analysis benchmarks them separately.

Notes

2026-10-17 fifth re-read of OpenRouter's DeepInfra (Turbo) row for llama-3.3-70b-instruct: 0.65s P50, 16 tok/s, 99.85% uptime, still $0.10/$0.32. The fourth 2026-10-17 read of the same row was 1.71s / 6 tps / 99.92%, the third was 0.55s / 16 tps / 99.94%, the second was 0.46s / 26 tps / 99.93%, and the earlier one was 0.62s / 19 tps / 99.94%. measured_* stays the Artificial Analysis long-prompt median (14.9 tok/s, 2.15s first chunk), re-confirmed unchanged this pass from the page summary (14.9 t/s, 20.5x slower than Groq at 305.7); the OpenRouter P50 is a different workload and is not promoted. 2026-10-17 fourth re-read of OpenRouter's DeepInfra (Turbo) row for llama-3.3-70b-instruct: 1.71s P50, 6 tok/s, 99.92% uptime, still $0.10/$0.32. The third 2026-10-17 read of the same row was 0.55s / 16 tps / 99.94%, the second was 0.46s / 26 tps / 99.93%, and the earlier one was 0.62s / 19 tps / 99.94%. measured_* stays the Artificial Analysis long-prompt median (14.9 tok/s, 2.15s 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 DeepInfra (Turbo) row for llama-3.3-70b-instruct: 0.55s P50, 16 tok/s, 99.94% uptime, still $0.10/$0.32. The second 2026-10-17 read of the same row was 0.46s / 26 tps / 99.93%, and the earlier one was 0.62s / 19 tps / 99.94%. measured_* stays the Artificial Analysis long-prompt median (14.9 tok/s, 2.15s first chunk), re-confirmed unchanged later the same day from the page summary (14.9 t/s; the comparison table still prints 15 / 2.14); the OpenRouter P50 is a different workload and is not promoted. 2026-10-17 later re-read of OpenRouter's DeepInfra (Turbo) row for llama-3.3-70b-instruct: 0.46s P50, 26 tok/s, 99.93% uptime, still $0.10/$0.32. The earlier 2026-10-17 read of the same row was 0.62s / 19 tps / 99.94%. measured_* stays the Artificial Analysis long-prompt median (14.9 tok/s, 2.15s first chunk, re-confirmed unchanged later the same day from the page's embedded byPromptType.long JSON); the OpenRouter P50 is a different workload and is not promoted. 2026-10-17 Artificial Analysis re-read. The provider page's embedded per-provider JSON (byPromptType.long) now gives DeepInfra (Turbo, FP8) 14.89 output tok/s at 2.146s median first chunk, down from the 15.2 tok/s recorded on 2026-10-16. TTFT is unchanged, so measured_ttft_ms stays 2140. The page disagrees with itself: the summary prose says 14.9 while the comparison table still prints 15, so the scalar takes the embedded JSON rounded to one decimal and 15.2 stays in this note. Price is untouched: the same JSON restates $0.10 in / $0.32 out, matching DeepInfra's own model page, models API and pricing table. DeepInfra stays the slowest of the providers on that page (Groq 305.7). OpenRouter's DeepInfra (Turbo) P50 the same day (0.62s, 19 tok/s) is a different quantity and is not promoted. 2026-10-17 OpenRouter telemetry only; price not moved. DeepInfra's own pages were not re-extracted this pass, so $0.10/$0.32 and context 131072 stay the 2026-10-17 first-party figures. OpenRouter's DeepInfra (Turbo) row the same day reads 0.62s P50, 19 tok/s, 99.94% uptime at the same $0.10/$0.32 (was 0.48s / 16 tok/s / 98.80% on 2026-10-16). measured_* stays the 2026-10-16 Artificial Analysis median (15.2 tok/s, 2.14s first chunk): the OpenRouter P50 is a different quantity and is recorded here rather than promoted. 2026-10-17: representative price moved off DeepSeek-V3.1 ($0.25/$0.95, still live) onto Llama-3.3-70B-Instruct-Turbo at $0.10/$0.32, context 131072. Three DeepInfra pages agree (models API, model page, pricing table), and the split matches the 2026-10-16 Artificial Analysis FAQ. The 2026-10-16 pass left the headline on V3.1 because the pricing page did not re-extract cleanly that day; it did on 2026-10-17. Speed is unchanged from the 2026-10-16 Artificial Analysis read: 15.2 output tok/s, 2.14s median first chunk, the slowest of the 11 providers on that page (Groq 305.7). The older OpenRouter P50 (0.51s, 14 tok/s, 99.52% uptime on 2026-10-14) is a different quantity and stays in the history below. 15.2 tps is shared-infrastructure Turbo, not a dedicated deployment. DeepInfra's own page states all models run on H100 or A100. Two tiers appear on provider tables (Turbo vs base). 2026-10-16: Artificial Analysis confirms the floor. Its Llama 3.3 70B FAQ states DeepInfra (Turbo, FP8) at $0.10 per 1M input and $0.32 per 1M output — the same split the 2026-10-14 OpenRouter table had — with a blended $0.12 per 1M, the lowest of 11 providers, and 15.2 output tok/s (20x slower than Groq's 305.7 on the same page) at 2.14s median first chunk. measured_tps moved from 14 (OpenRouter P50) to 15.2 (AA median); measured_ttft_ms moved from 510 to 2140 because the two measurements are not the same quantity (OpenRouter 'latency' vs AA 'time to first chunk'). The representative price fields stay on DeepSeek-V3.1 at $0.25/$0.95 as recorded 2026-10-14: that number was read off DeepInfra's own pricing page, which did not re-extract cleanly on 2026-10-16, so it is not silently overwritten. Measured figures are the OpenRouter 2026-10-14 P50 for DeepInfra (Turbo) on llama-3.3-70b-instruct: 0.51s latency, 14 tok/s, 99.52% uptime - the cheapest and the SLOWEST row on that table, which is exactly the price-vs-speed trade this record type exists to expose. 14 tps is shared-infrastructure Turbo, not a dedicated deployment. DeepInfra's own page states all models run on H100 or A100 GPUs. Two tiers appear on provider tables: standard and Turbo (cheaper, sometimes FP8); Artificial Analysis benchmarks them separately.

Cross-references

Models (11)