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

Together AI serverless inference

together

Fields

id
services/svc-together-serverless
Provider
together
Endpoint
https://api.together.xyz/v1
Input $/Mtok
1.04
Output $/Mtok
1.04
Price basis
USD per 1M TOKENS, representative model meta-llama/Llama-3.3-70B serverless list price, observed 2026-10-14 ($1.04 input = $1.04 output, symmetric). Not the cheapest rate on the page: gpt-oss-120b is $0.15/$0.60, Qwen3-235B FP8 is $0.20/$0.60, GLM-5.3-Flash $0.15/$0.50, Llama-3-8B-Instruct-Lite $0.14/$0.14. Cached input is a separate lower price where offered (e.g. MiniMax M3 $0.30/$0.06-cached/$1.20). Dedicated endpoints and GPU clusters are priced per-hour separately; batch API has its own column on the same page.
Context window
131,072
TTFT (ms)
1,500
Output tok/s
79
Measured by
artificial-analysis
Regions
us
notes
2026-10-17 fifth re-read: OpenRouter's Together row for llama-3.3-70b-instruct is now 0.94s P50, 28 tok/s, 96.72% uptime, still $1.04/$1.04. The fourth 2026-10-17 read of the same row was 0.55s / 41 tps / 96.68%, the third was 0.71s / 39 tps / 96.30%, the second was 0.91s / 53 tps / 95.86%, and the earlier one was 0.94s / 52 tps / 95.75%. measured_* stays the 2026-10-16 Artificial Analysis median (79 tok/s, 1.50s first chunk), re-confirmed unchanged this pass from the page summary (Together AI Turbo first token 1.53s); the OpenRouter P50 is a different workload and is not promoted. 2026-10-17 fourth re-read: OpenRouter's Together row for llama-3.3-70b-instruct is now 0.55s P50, 41 tok/s, 96.68% uptime, still $1.04/$1.04. The third 2026-10-17 read of the same row was 0.71s / 39 tps / 96.30%, the second was 0.91s / 53 tps / 95.86%, and the earlier one was 0.94s / 52 tps / 95.75%. measured_* stays the 2026-10-16 Artificial Analysis median (79 tok/s, 1.50s first chunk), re-confirmed unchanged later on 2026-10-17 and not re-fetched this pass; the OpenRouter P50 is a different workload and is not promoted. 2026-10-17 third re-read: OpenRouter's Together row for llama-3.3-70b-instruct is now 0.71s P50, 39 tok/s, 96.30% uptime, still $1.04/$1.04. The second 2026-10-17 read of the same row was 0.91s / 53 tps / 95.86%, and the earlier one was 0.94s / 52 tps / 95.75%. measured_* stays the 2026-10-16 Artificial Analysis median (79 tok/s, 1.50s first chunk), re-confirmed unchanged later on 2026-10-17 (summary first-token 1.53s for Together AI Turbo; the comparison table still prints 79 / 1.50, the same recorded pair); the OpenRouter P50 is a different workload and is not promoted. 2026-10-17 later re-read: OpenRouter's Together row for llama-3.3-70b-instruct is now 0.91s P50, 53 tok/s, 95.86% uptime, still $1.04/$1.04. The earlier 2026-10-17 read of the same row was 0.94s / 52 tps / 95.75%. measured_* stays the 2026-10-16 Artificial Analysis median (79 tok/s, 1.50s first chunk), re-confirmed unchanged later on 2026-10-17 from the page's embedded long-prompt JSON (80.37 tok/s, 1.53s, which rounds to the same recorded pair); the OpenRouter P50 is a different workload and is not promoted. 2026-10-17: OpenRouter's Together row for llama-3.3-70b-instruct now has a latency cell — 0.94s P50, 52 tok/s, 95.75% uptime, still $1.04/$1.04. The 2026-10-16 read of the same table had no latency cell. measured_* stays the 2026-10-16 Artificial Analysis median (79 tok/s, 1.50s first chunk); the OpenRouter P50 is a different workload and is not promoted. Price fields stay $1.04/$1.04, the 2026-10-14 Together pricing-table figure that OpenRouter still matches. 2026-10-17: Together's Llama 3.3 70B model page confirms the endpoint id meta-llama/Llama-3.3-70B-Instruct-Turbo but the extracted page carries no per-1M price, so $1.04/$1.04 is not moved. It remains the 2026-10-14 Together pricing-table and OpenRouter Together-row figure. Speed stays the 2026-10-16 Artificial Analysis median (79 tok/s, 1.50s first chunk). 2026-10-16 measurement refresh. Artificial Analysis lists Together AI (Turbo) on Llama 3.3 70B at 79 output tok/s median and 1.50s median first chunk, context 131k. The previous measured pair (0.74s, 43 tok/s) was OpenRouter's 2026-10-14 P50 for the Together row and is a different traffic mix. The $1.04/$1.04 price is still the 2026-10-14 OpenRouter Together row; Together's own pricing page fetched 2026-10-16 is a long per-model catalog that did not surface a Llama 3.3 70B row in the extracted text, so the price was not moved. Measured figures are the OpenRouter routing table's 2026-10-14 P50 for Together's llama-3.3-70b endpoint (0.74s latency, 43 tok/s, 95.60% uptime) - third-party-measured over live traffic, matching the Artificial Analysis provider set which also benchmarks Together Turbo on this model. 43 tps is a SHARED-infrastructure number: Together's pitch is dedicated endpoints for higher single-stream speed, so the serverless figure is the floor. Models listed are a sample of a ~100-model catalog spelled as the repo's model slugs; provider-side ids are namespaced (meta-llama/Llama-3.3-70B-Instruct-Turbo).

Notes

2026-10-17 fifth re-read: OpenRouter's Together row for llama-3.3-70b-instruct is now 0.94s P50, 28 tok/s, 96.72% uptime, still $1.04/$1.04. The fourth 2026-10-17 read of the same row was 0.55s / 41 tps / 96.68%, the third was 0.71s / 39 tps / 96.30%, the second was 0.91s / 53 tps / 95.86%, and the earlier one was 0.94s / 52 tps / 95.75%. measured_* stays the 2026-10-16 Artificial Analysis median (79 tok/s, 1.50s first chunk), re-confirmed unchanged this pass from the page summary (Together AI Turbo first token 1.53s); the OpenRouter P50 is a different workload and is not promoted. 2026-10-17 fourth re-read: OpenRouter's Together row for llama-3.3-70b-instruct is now 0.55s P50, 41 tok/s, 96.68% uptime, still $1.04/$1.04. The third 2026-10-17 read of the same row was 0.71s / 39 tps / 96.30%, the second was 0.91s / 53 tps / 95.86%, and the earlier one was 0.94s / 52 tps / 95.75%. measured_* stays the 2026-10-16 Artificial Analysis median (79 tok/s, 1.50s first chunk), re-confirmed unchanged later on 2026-10-17 and not re-fetched this pass; the OpenRouter P50 is a different workload and is not promoted. 2026-10-17 third re-read: OpenRouter's Together row for llama-3.3-70b-instruct is now 0.71s P50, 39 tok/s, 96.30% uptime, still $1.04/$1.04. The second 2026-10-17 read of the same row was 0.91s / 53 tps / 95.86%, and the earlier one was 0.94s / 52 tps / 95.75%. measured_* stays the 2026-10-16 Artificial Analysis median (79 tok/s, 1.50s first chunk), re-confirmed unchanged later on 2026-10-17 (summary first-token 1.53s for Together AI Turbo; the comparison table still prints 79 / 1.50, the same recorded pair); the OpenRouter P50 is a different workload and is not promoted. 2026-10-17 later re-read: OpenRouter's Together row for llama-3.3-70b-instruct is now 0.91s P50, 53 tok/s, 95.86% uptime, still $1.04/$1.04. The earlier 2026-10-17 read of the same row was 0.94s / 52 tps / 95.75%. measured_* stays the 2026-10-16 Artificial Analysis median (79 tok/s, 1.50s first chunk), re-confirmed unchanged later on 2026-10-17 from the page's embedded long-prompt JSON (80.37 tok/s, 1.53s, which rounds to the same recorded pair); the OpenRouter P50 is a different workload and is not promoted. 2026-10-17: OpenRouter's Together row for llama-3.3-70b-instruct now has a latency cell — 0.94s P50, 52 tok/s, 95.75% uptime, still $1.04/$1.04. The 2026-10-16 read of the same table had no latency cell. measured_* stays the 2026-10-16 Artificial Analysis median (79 tok/s, 1.50s first chunk); the OpenRouter P50 is a different workload and is not promoted. Price fields stay $1.04/$1.04, the 2026-10-14 Together pricing-table figure that OpenRouter still matches. 2026-10-17: Together's Llama 3.3 70B model page confirms the endpoint id meta-llama/Llama-3.3-70B-Instruct-Turbo but the extracted page carries no per-1M price, so $1.04/$1.04 is not moved. It remains the 2026-10-14 Together pricing-table and OpenRouter Together-row figure. Speed stays the 2026-10-16 Artificial Analysis median (79 tok/s, 1.50s first chunk). 2026-10-16 measurement refresh. Artificial Analysis lists Together AI (Turbo) on Llama 3.3 70B at 79 output tok/s median and 1.50s median first chunk, context 131k. The previous measured pair (0.74s, 43 tok/s) was OpenRouter's 2026-10-14 P50 for the Together row and is a different traffic mix. The $1.04/$1.04 price is still the 2026-10-14 OpenRouter Together row; Together's own pricing page fetched 2026-10-16 is a long per-model catalog that did not surface a Llama 3.3 70B row in the extracted text, so the price was not moved. Measured figures are the OpenRouter routing table's 2026-10-14 P50 for Together's llama-3.3-70b endpoint (0.74s latency, 43 tok/s, 95.60% uptime) - third-party-measured over live traffic, matching the Artificial Analysis provider set which also benchmarks Together Turbo on this model. 43 tps is a SHARED-infrastructure number: Together's pitch is dedicated endpoints for higher single-stream speed, so the serverless figure is the floor. Models listed are a sample of a ~100-model catalog spelled as the repo's model slugs; provider-side ids are namespaced (meta-llama/Llama-3.3-70B-Instruct-Turbo).

Cross-references

Models (10)