Together AI
by Together AI
The AI-native cloud for open-source models — inference, fine-tuning and GPU clusters on one platform.
Together AI is a full-stack cloud built specifically for open-source and open-weight AI models. It sells token-billed serverless inference, batch and provisioned-throughput endpoints, fine-tuning, managed storage and raw GPU clusters, so engineering teams can run Llama, DeepSeek, Qwen, FLUX and Whisper-class models in production without operating their own accelerator fleet or signing a long-term datacentre contract.
Together AI operates a full-stack cloud built specifically for open-source and open-weight AI models, spanning serverless token-billed inference, asynchronous batch inference, provisioned throughput with a 99% uptime SLA, dedicated single-tenant model and container endpoints, fine-tuning, managed storage, secure code sandboxes, and raw GPU clusters that scale from an instant handful of cards to thousands. Founded in 2022 by Vipul Ved Prakash, Ce Zhang, Chris Re, Tri Dao and Percy Liang, the company positions itself as 'the AI-native cloud' and sells research-derived speedups rather than commodity capacity, claiming roughly 2x faster inference than competing providers, a 60% cost reduction from workload-specific optimisation, and 90% faster pre-training via the Together Kernel Collection. Unusually for this market, list pricing is fully published: H100 GPUs run $3.99 per GPU-hour on demand and $3.19 reserved, H200 $5.99/$3.99, B200 $8.19/$6.79, dedicated H100 inference starts at $5.49/hour, and serverless tokens range from $0.14 per million input tokens on DeepSeek V4 Flash to $1.04 on Llama 3.3 70B. On 1 July 2026 the company closed an $800 million Series C at an $8.3 billion valuation led by Aramco Ventures, with NVIDIA, Vista Equity Partners, General Catalyst, Salesforce Ventures, Emergence Capital, March Capital, Pegatron and S Ventures participating, plus investor commitments to independently capitalise more than 500 MW of compute. Together describes itself as one of the largest producers of AI tokens in the world and names Cursor, Cognition, Decagon, ElevenLabs, Suno, Salesforce, Zoom and Mozilla among thousands of customers; Cursor's deployment runs on GB200 NVL72 topology. It reached SOC 2 Type 2 and HIPAA compliance on 8 July 2025.
The platform or ML infrastructure lead standardising on open-weight models who needs published GPU and token pricing, fine-tuning and dedicated capacity from one vendor instead of stitching together a GPU cloud, an inference vendor and a training stack.
Run and fine-tune open-source models in production at transparent, per-token or per-GPU-hour rates without buying, scheduling or operating your own accelerator fleet.
At a Glance
- Category
- Infrastructure & Cloud
- Pricing
- Usage-based, Subscription, Contact for pricing
- Target Market
- CTOs, VPs of Engineering, ML Infrastructure Engineers, Data Scientists, Enterprise Developers
- Deployment
- Cloud-first, API-based
- Founded
- 2022
- Headquarters
- San Francisco, United States
- Team Size
- 201-500
- Customers
- Thousands of customers, including Cursor, Cognition, Decagon, ElevenLabs, Suno, Salesforce, Zoom and Mozilla
Key Features
- ✓Serverless inference
Per-token access to 200+ open chat, image, video, audio and embedding models with no capacity to provision or idle GPUs to pay for.
- ✓GPU clusters and accelerated compute
On-demand or reserved H100, H200 and B200 capacity billed by the GPU-hour, scaling from single cards to thousands without a long-term datacentre contract.
- ✓Provisioned throughput and dedicated endpoints
Reserved single-tenant capacity with a 99% uptime SLA, so production traffic is insulated from noisy-neighbour variance on shared serverless pools.
- ✓Fine-tuning service
Managed fine-tuning of open models from $0.48 per million tokens on sub-16B models, with a $4 minimum per job and no infrastructure to stand up.
- ✓Together Kernel Collection and ATLAS speculator
In-house CUDA kernels and adaptive speculative decoding that the company credits for 90% faster pre-training and materially faster serving on its own stack.
- ✓Batch inference and code sandboxes
Asynchronous bulk processing plus isolated vCPU sandboxes at $0.0446 per vCPU-hour for agent workloads that must execute untrusted generated code.
Capabilities
Use Cases
- •Replacing a proprietary API with an open model
Swap a frontier-model API call for a hosted open-weight equivalent to cut per-token cost while keeping an OpenAI-compatible interface.
- •Domain fine-tuning on private data
Adapt a 7B–70B open model to internal terminology and formats, then serve the result on dedicated endpoints under one vendor contract.
- •Burst training and evaluation runs
Rent B200 or H200 clusters for the weeks a training or evaluation campaign actually runs, then release the capacity instead of holding reservations.
- •Coding-assistant inference at editor latency
Serve autocomplete and agent traffic on GB200 NVL72 topology, the architecture Cursor uses to keep responses inside the editor feedback loop.
- •Regulated-industry AI workloads
Run clinical decision support or biomedical analysis on open models under Together's SOC 2 Type 2 and HIPAA posture with a signed BAA.
Ideal For
Best For
- ✓Serving open-weight LLMs (Llama, DeepSeek, Qwen, MiniMax) behind an OpenAI-compatible API without managing GPUs
- ✓Fine-tuning open models on proprietary data at published per-million-token rates rather than a negotiated enterprise contract
- ✓Bursty or short-lived GPU cluster needs where a 12-month reserved datacentre commitment is the wrong shape
- ✓Large asynchronous batch jobs — document extraction, embedding backfills, synthetic data generation — priced below real-time inference
- ✓Cost-reduction migrations off proprietary frontier APIs where an open model is good enough for the task
Not Ideal For
- ✗Latency-critical, user-facing token streaming where every millisecond counts — independent benchmarks and Hacker News practitioners put Groq and Cerebras ahead of Together on time-to-first-token and observed throughput
- ✗Teams that need proprietary frontier models (GPT, Claude, Gemini) as their primary engine; Together's catalogue is open-weight and hosted community models
- ✗Organisations with hard on-premise or air-gapped requirements — everything is delivered as a hosted cloud service
- ✗Cost-averse teams without spend guardrails; G2 reviewers report unexpectedly large monthly bills when usage is not tested and capped first
Integrations
Deployment
Market & Ratings
Thousands of customers, including Cursor, Cognition, Decagon, ElevenLabs, Suno, Salesforce, Zoom and Mozilla
Market Analysis
Pros
- ✓Breadth no single competitor matches: 200+ open models, fine-tuning, dedicated endpoints and bare GPU clusters under one account and one bill
- ✓Transparent published pricing for tokens, GPU-hours, fine-tuning, storage and sandboxes, which makes cost modelling possible before a sales call
- ✓SOC 2 Type 2 and HIPAA compliance (achieved July 2025) with BAAs, which unblocks regulated healthcare and life-sciences workloads on open models
- ✓Removes GPU cluster operations entirely — G2 reviewers cite not having to manage accelerators as the main reason they stay
Cons
- ✗Practitioners on Hacker News dispute the speed claims: one cites OpenRouter Kimi-K2 numbers of 'Groq averaging 1,086tps vs Together doing 59tps', calling the marketed parity into question
- ✗Reliability under agentic load has been questioned — the same thread reports Together with 'one of the highest tool call failure rates (>300 failures over the benchmark, compared to 0-2 for the official API, groq, SiliconFlow)' on the Kimi vendor verifier
- ✗Time-to-first-token trails the LPU/wafer-scale specialists; independent comparisons put Together around 1.1s TTFT against roughly 0.5s for Fireworks, so it is not the pick for real-time voice or typing-latency UX
- ✗G2 reviewers report thin documentation in places and unexpectedly high month-end bills when usage is not capped, a common failure mode of pure usage-based pricing
- ✗No on-premise or self-hosted option, and the published security posture covers SOC 2 and HIPAA but not ISO 27001 or documented data residency
Pricing
Serverless inference
From $0.14 per 1M input tokens
- ✓Pay-per-token across 200+ open models
- ✓OpenAI-compatible API
- ✓No capacity to reserve
- ✓Image from $0.0027, video from $0.115, transcription from $0.0015/min
GPU clusters (on-demand)
From $3.99/GPU-hour
- ✓H100 $3.99, H200 $5.99, B200 $8.19 per GPU-hour
- ✓Instant clusters to thousands of GPUs
- ✓Managed storage at $0.16/GiB-month
GPU clusters (reserved, 181+ days)
From $3.19/GPU-hour
- ✓H100 $3.19, H200 $3.99, B200 $6.79 per GPU-hour
- ✓Committed-term discount versus on-demand
Dedicated inference
From $5.49/hour
- ✓Single-tenant GPU instances
- ✓Guaranteed performance
- ✓Provisioned throughput with 99% uptime SLA
Fine-tuning
From $0.48 per 1M tokens
- ✓Up to 16B: $0.48–$1.35
- ✓17B–69B: $1.50–$4.12
- ✓70–100B: $2.90–$8.00
- ✓$4.00 minimum per job
Together publishes full list pricing, which is rare in this category — inference is metered per million tokens and varies roughly 10x by model, GPUs are billed per GPU-hour with a reserved tier for 181+ day commitments, and fine-tuning is per training token with a $4 job minimum. There is no free tier and no named Build/Scale/Enterprise subscription on the public pricing page; the enterprise motion is reserved capacity and provisioned throughput rather than a seat plan. Storage ($0.16/GiB-month), sandboxes and code interpreter sessions are metered separately, and G2 reviewers specifically warn that usage-based billing produces surprise invoices without testing and spend limits first.
Security & Compliance
Connect
Sources
This page was written from 8 sources, 2 on domains other than together.ai.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
LiteLLM
Open-source AI gateway: 100+ LLM APIs behind one OpenAI-compatible endpoint, with cost tracking and guardrails
CIQ Fuzzball
Sovereign AI and HPC orchestration: train, fine-tune and serve models on infrastructure you control
Crusoe Cloud
Vertically integrated AI cloud — own the power, own the datacentre, rent the GPUs by the minute
Nexthop AI
Ethernet switches and network software co-designed for AI clusters, built for hyperscalers and NeoClouds