Together AI review: the widest catalog, and what it costs you

Together AI is an inference and compute platform for open models. It sells four things that most competitors sell one or two of: per-token serverless inference, reserved GPU clusters by the hour, managed fine-tuning, and code sandboxes. That breadth is genuinely useful if you are doing more than calling an endpoint, and mostly irrelevant if you are not.

By the Continuum team. We build a workbench that runs Claude Code, Codex, and their peers, so the model rates quoted here are the ones our own cost analytics ship with.

The short version

Together AI serves the current open-model catalog (DeepSeek V4, Kimi K3, GLM 5.2, Qwen 3.8, MiniMax M3, Llama, Gemma) at prices essentially identical to Fireworks and Baseten, plus image, video, audio, and embedding models that its rivals do not all carry. Underneath that it rents H100, H200, and B200 clusters from $3.19 to $8.19 per GPU hour, runs managed fine-tuning from $0.48 per million training tokens, and sells a code sandbox by the vCPU hour. Its research lineage (FlashAttention, Medusa, Mamba variants, the Together Kernel Collection) is real and shows up in the serving stack.

What you need to know
  • Widest surface in the category: tokens, GPU clusters, fine-tuning, sandboxes, and storage on one bill.
  • Token prices match the market. GLM 5.2 at $1.40 / $4.40, the same as Fireworks and Baseten.
  • Genuinely broad media catalog: FLUX image models, Veo and Seedance video, Whisper and Nemotron ASR.
  • GPU clusters are competitively priced: H100 at $3.99 on demand, $3.19 reserved.
  • Batch inference is up to 50 percent cheaper with a 24-hour window, and scales to 30 billion tokens.
  • No published free tier as of August 2026. Budget real money for evaluation.

What Together AI is

Together started as a research lab and became an infrastructure company, and both halves are still visible. The research half produced FlashAttention (now at version 4), Medusa and Sequoia for speculative decoding, the BASED and Mamba architecture work, SpecExec, and the Together Kernel Collection. The infrastructure half sells the result.

Together product surfaces, August 2026.
SurfaceBilling unitWhat it is for
Serverless inferencePer input and output tokenCall an open model from an endpoint. Rate-limited shared capacity.
Batch inferencePer token, up to 50% offAsynchronous jobs on a 24-hour window. Scales to 30 billion tokens.
Provisioned throughputReserved, token-pricedGuaranteed capacity with a 99% uptime SLA.
Dedicated inferencePer GPU hourYour model on your own instances, from $5.49/hour on an H100.
GPU clustersPer GPU hourRaw compute, self-serve up to thousands of GPUs.
Fine-tuningPer million training tokensLoRA and full-parameter runs from $0.48/M.
SandboxPer vCPU hour and GiB hourCode execution environments for agents.

The headline marketing claims are 2x faster inference, 60 percent lower cost through workload-specific optimization, and 90 percent faster pre-training with the kernel collection. Treat all three the way you should treat any vendor multiplier: they are against an unnamed baseline. The verifiable part is the published rate card, which is where the rest of this page lives.

The catalog, which is the actual differentiator

Text models are the same set everyone serves, at the same prices. Where Together separates is everything that is not text.

Non-text catalog with published rates, August 2026.
CategoryExamplesPrice
ImageFLUX.2 [pro]$0.03 per image
ImageFLUX.1 [schnell]$0.0027 per image at 4 steps
ImageIdeogram 4.0$0.06 per image
ImageFLUX.2 [max]$0.070 per megapixel
VideoByteDance Seedance 2.5$0.115 per video
VideoFLUX 3$0.17 per video
VideoGoogle Veo 3.0$1.60 per video
AudioWhisper Large v3$0.0015 per audio minute
AudioNVIDIA Nemotron 3.5 ASR$0.0045 per audio minute
EmbeddingsMultilingual e5 large instruct$0.02 per million tokens

On the text side the catalog runs MiniMax M3, GLM 5.2, Kimi K3, DeepSeek V4 in Flash and Pro variants, Qwen 3.8, Llama 3.3 70B, and Gemma 4 31B, among more than thirty others. Full rates are on the Together AI pricing page; if you are choosing between them on quality rather than price, the cards for Kimi K3 and GLM 5.2 carry the independently published coding scores.

GPU clusters and dedicated inference

This is where Together diverges most sharply from Fireworks and Baseten, and where its pricing is genuinely competitive rather than at parity.

GPU rates per GPU hour.checked aug 2026
GPUCluster, on demandCluster, reserved 91 to 180 daysDedicated inference
NVIDIA HGX H100$3.99$3.19$5.49
NVIDIA HGX H200$5.99$3.99Not listed
NVIDIA HGX B200$8.19$6.79$8.99

Two things worth noticing. Reserving for a 91 to 180 day reservation cuts an H200 from $5.99 to $3.99, a 33 percent discount, which is the single largest lever on the page for anyone with steady load. And dedicated inference costs more per GPU hour than a raw cluster ($5.49 against $3.99 for an H100) because Together is running the serving stack for you. That premium is the honest price of not operating vLLM yourself.

Storage is $0.16 per GiB-month on the shared filesystem, and Together advertises zero egress fees on its managed storage, which is worth checking against your cloud bill if you are moving large datasets.

Fine-tuning, batch, and sandboxes

Fine-tuning

Priced per million training tokens, like Fireworks, and cheaper at the small end. Supervised fine-tuning on a model up to 16B is $0.48 per million training tokens, DPO is $0.54, and full-parameter runs are $1.20 to $1.35. Specialized base models cost considerably more: LoRA on a DeepSeek V4 Flash class model runs $6.00 to $40, and DPO $15.00 to $100.

Batch inference

Up to 50 percent off serverless rates on a 24-hour window, and Together advertises scaling to 30 billion tokens in a single job. For evaluation runs, dataset labelling, or embedding backfills, this is the correct surface and most people never use it.

Sandbox

A code execution environment for agents, billed at $0.0446 per vCPU hour and $0.0149 per GiB hour, with a code interpreter session (60 minutes) at $0.03. This is directly relevant if you are building an agent that runs generated code and do not want to operate the container fleet.

Where Together is the wrong choice

  • You just want one text endpoint. The breadth is the product. If you need one model behind one URL, a narrower provider is simpler and identically priced.
  • You want to try before you buy. Together publishes no ongoing free credit on its pricing page as of August 2026. The application-only startup program is not an evaluation path for a weekend prototype.
  • You need frontier models. Like every open-model shop, Together does not sell Claude Opus 5 or gpt-5.6-sol. Those come from the labs.
  • You are optimizing purely on token price. DeepInfra undercut Together on every model we compared in August 2026.

Questions people ask

What is Together AI?

Together AI is an inference and compute platform for open models. It sells per-token serverless inference, batch inference at up to 50 percent off, provisioned throughput with a 99 percent uptime SLA, dedicated inference instances, raw GPU clusters by the hour, managed fine-tuning, code sandboxes, and managed storage. It also publishes research including FlashAttention, Medusa, and the Together Kernel Collection.

What models does Together AI support?

On the text side, more than thirty open models including MiniMax M3, GLM 5.2, Kimi K3, DeepSeek V4 Flash and Pro, Qwen 3.8, Llama 3.3 70B, and Gemma 4 31B. It also serves image models (FLUX.2 pro and max, FLUX.1 schnell, Ideogram 4.0), video models (Seedance 2.5, FLUX 3, Google Veo 3.0), transcription (Whisper Large v3, Nemotron 3.5 ASR), embeddings, and rerankers.

Is Together AI cheaper than Fireworks?

No. As of August 2026 both list DeepSeek V4 Flash at $0.14 / $0.28, GLM 5.2 at $1.40 / $4.40, and Kimi K3 at $3.00 / $15.00 per million tokens. Together is cheaper on GPU clusters, where an H100 is $3.99 per hour on demand against $7.00 at Fireworks, but that is a different product from serverless tokens.

Does Together AI have a free tier?

No ongoing free credit is published on its pricing page as of August 2026. There is a separate application-only startup program offering credits, but that is not a self-serve trial. Budget real spend for an evaluation, or start on a provider that grants credits (Modal renews $30 per month on its free Starter plan).

How much are Together AI GPUs per hour?

On demand: HGX H100 at $3.99, H200 at $5.99, B200 at $8.19 per GPU hour. Reserved for 91 to 180 days: $3.19, $3.99, and $6.79 respectively. Dedicated inference, where Together runs the serving stack for you, is $5.49 for an H100 and $8.99 for a B200.

Can Together AI fine-tune a model for me?

Yes, billed per million training tokens rather than per GPU hour. Supervised fine-tuning on models up to 16B is $0.48 per million training tokens, DPO is $0.54, and full-parameter runs are $1.20 to $1.35. Specialized base models cost substantially more, with LoRA at $6.00 to $40 and DPO at $15.00 to $100 per million training tokens.

Sources

Every figure above was read from these pages on August 2026. Vendors reprice without notice; if you find a stale number, tell us.

  1. Together AI product surfaces, research lineage, customers, performance claims
  2. Together AI pricing per-model rates, GPU cluster and dedicated rates, fine-tuning, sandbox, storage
  3. Together AI pricing docs billing units and the batch discount
  4. Fireworks serverless pricing comparison rates
Try it

Run every agent
from one place.

Continuum drives Claude Code, Codex, and peers under your own subscriptions, with live quota gauges and spend by repo. The app is free. Mac is stable; Windows and Linux desktop are beta.

free app · your subscriptions · local-first