Together AI API Pricing & Model Rates
Together AI is the premier open-source AI platform, hosting over 200 open models, serverless LoRA fine-tuning, and dedicated GPU clusters. Compare token rates, fine-tuning costs, and enterprise SLAs.
Together AI Serverless Inference Simulator
Calculate monthly bills across Together AI serverless models and dedicated clusters.
Production Pricing & Model Tier Rates
Verified rates, context capacities, typical inference latency, and recommended enterprise workloads.
| Model / Tier | Input Rate (1M) | Output Rate (1M) | Context Window | Latency / Throughput | Recommended Workload |
|---|---|---|---|---|---|
| Llama 3.3 70B Instruct Turbo | $0.880 | $0.880 | 128,000 | 110–140 tok/s | Flagship open weights reasoning, agent loops, summarization |
| Llama 3.1 8B Instruct Turbo | $0.180 | $0.180 | 128,000 | 180–240 tok/s | Fast classification, intent routing, synthetic data |
| DeepSeek-V3 (671B MoE) | $1.250 | $1.250 | 64,000 | 50–80 tok/s | High-volume coding, analysis, and developer tooling |
| DeepSeek-R1 (Reasoning) | $1.500 | $3.000 | 64,000 | 35–55 tok/s | Complex math, competitive programming, formal logic |
| Qwen 2.5 Coder 32B | $0.800 | $0.800 | 32,768 | 100–130 tok/s | Specialized coding completion, bug fixing, test writing |
| Flux.1 Schnell (Image) | $0.003 / img | N/A | N/A | Sub-1s | High-speed diffusion image generation |
⚡ Architectural Deep Dive & Unit Economics Analysis
Together AI has established itself as the comprehensive research-to-production cloud for the open-source AI ecosystem. Unlike single-purpose inference hosts, Together AI covers the entire model lifecycle: from pre-training and fine-tuning to high-speed serverless serving and dedicated reserved GPU clusters.
Their proprietary Together Turbo inference engine delivers 110–140 tokens per second on Llama 3.3 70B while maintaining strict FP8 and BF16 numeric precision. Furthermore, their self-service fine-tuning API allows teams to upload domain datasets, execute LoRA or full-parameter tuning for nominal fees ($0.008/1k tokens), and immediately host the resulting adapter without provisioning dedicated infrastructure.
Production Workloads & Monthly Cost Modeling
Evaluating practical unit economics across real-world application architectures.
Custom LoRA Domain Adapter for Legal Contracts
Fine-tuning a 70B model on 5M tokens of legal clauses ($40 one-time) and serving queries at standard serverless token rates without GPU idle fees.
High-Volume Customer Support Agent (100M tokens)
Processing 80M prompt tokens and 20M output tokens monthly. Together AI costs $88.00/month versus $540.00/month on Claude Sonnet.
Dedicated 8x H100 Reserved Cluster
Running continuous private model inference with zero noisy-neighbor degradation and custom weights for $3.50/hour (~$2,520/month).
import os
from together import Together
client = Together(api_key=os.environ.get("TOGETHER_API_KEY"))
def stream_together_llama(prompt: str):
"""Stream completions using Together AI Python client."""
response = client.chat.completions.create(
model="meta-llama/Llama-3.3-70B-Instruct-Turbo",
messages=[{"role": "user", "content": prompt}],
temperature=0.3,
stream=True
)
for chunk in response:
print(chunk.choices[0].delta.content or "", end="", flush=True)Operational Gotchas & Failure Modes
Engineering realities, throttling thresholds, and hidden surcharge traps to monitor.
Related Models, Showdowns & Benchmarks
Cross-reference provider rates against hosted models, direct head-to-head battles, and calculators.
Together AI vs Fireworks AI Showdown
Compare serverless GPU engines, latency, and LoRA support.
Groq vs Together AI Comparison
Evaluate LPU speed against serverless GPU flexibility.
Cheapest LLM APIs of 2026
Rankings of all low-cost token providers on the market.
LLM Fine-Tuning ROI Calculator
Calculate when fine-tuning a small model beats prompting frontier APIs.
Frequently Asked Questions
What is Together AI's token pricing for Llama 3.3 70B?
Together AI charges $0.88 per million input tokens and $0.88 per million output tokens for Llama 3.3 70B Instruct Turbo.
How does Together AI fine-tuning work?
Together AI offers a self-service Fine-Tuning API where you upload JSONL training data. LoRA fine-tuning costs approximately $0.008 per 1,000 tokens, and the resulting model can be served immediately via serverless API endpoints.
Is Together AI compatible with the OpenAI Python SDK?
Yes! Together AI provides an OpenAI-compatible endpoint. Set base_url to 'https://api.together.xyz/v1' and supply your TOGETHER_API_KEY.
What is Together Turbo?
Together Turbo is Together AI's custom-engineered inference stack that delivers 110–140 tokens per second on 70B models, providing low latency without compromising output precision.
What is the cost of dedicated GPU clusters on Together AI?
Together AI offers dedicated H100 instances starting at approximately $3.50 per GPU-hour with private VPC endpoints and guaranteed throughput.
Does Together AI host DeepSeek models?
Yes! Together AI hosts both DeepSeek-V3 ($1.25/M tokens) and DeepSeek-R1 ($1.50 in / $3.00 out) with high-availability enterprise SLAs.
Together AI Serverless & Dedicated Clusters: Turbo Inference Economics
Together AI provides a full-lifecycle cloud platform for open-source AI: serverless pay-per-token inference (Llama 3.3, DeepSeek-R1, Qwen 2.5), fine-tuning pipelines, and dedicated GPU instance reservations. On serverless endpoints, Together charges $0.18 per million input tokens and $0.59 per million output tokens on Llama 3.3 70B Turbo. For enterprise workloads requiring dedicated capacity, Together manages multi-GPU clusters (8x H100 SXM nodes at $2.95/hr per GPU) backed by high-speed InfiniBand networking.
Together AI Serverless Pricing Matrix
| Model Family | Input ($/M) | Output ($/M) | Throughput SLA | Context Length |
|---|---|---|---|---|
| Llama 3.3 70B Turbo | $0.18 | $0.59 | 115 tok/s | 128,000 tokens |
| DeepSeek-R1 (Serverless) | $0.80 | $2.40 | 70 tok/s | 64,000 tokens |
| Llama 3.1 8B Turbo | $0.06 | $0.18 | 180 tok/s | 128,000 tokens |
| Dedicated 8x H100 Node | $23.60 / hr | Unlimited tokens | Dedicated silicon | Full hardware control |
Together AI Streaming Client (TypeScript)
import Together from 'together-ai';
const together = new Together({ apiKey: process.env.TOGETHER_API_KEY });
export async function streamTogetherResponse(prompt: string) {
const runner = await together.chat.completions.create({
model: 'meta-llama/Llama-3.3-70B-Instruct-Turbo',
messages: [{ role: 'user', content: prompt }],
stream: true
});
for await (const chunk of runner) {
process.stdout.write(chunk.choices[0]?.delta?.content || '');
}
}Engineering Q&A: Together AI Optimization
What is the difference between Together Turbo and standard models?
Together Turbo models run on custom FlashAttention kernels with optimized FP8 quantization, delivering 2x to 3x higher tokens per second at 30% lower prices compared to standard FP16 endpoints, with imperceptible quality loss.
Does Together AI provide fine-tuning infrastructure?
Yes. Together provides full-parameter and LoRA fine-tuning APIs. You upload a JSONL dataset, and Together handles distributed cluster training on H100s, generating a private model endpoint ready for production inference.