Together AI

Together AI hosts a wide catalogue of open-weight models behind one OpenAI-style API, billed per token, and sells dedicated GPUs and fine-tuning too. On 2026-10-31 gpt-oss-120B costs $0.15 in and $0.60 out per million tokens, and DeepSeek V4.1 Flash costs $0.30 and $1.20. Its policy says it will not train on your data without opt-in.

Facts at a glance

Prices checked 2026-10-31. Together does not make the models. It serves other labs' open weights, which makes it the host to look at when you want an open model with data terms that are not the maker's.

Models and prices

The serverless table lists many models. These are representative rows, per million tokens, with the cached input price where the page shows one.

Together also lists small and experimental entries, for example a 4B model at $0.04 and one free listing, and a long tail of older models. The page notes that displayed prices refer to the lowest resolution or duration setting for image, audio and video models, so those rows move with settings. The table also carries a Batch API price column.

Compare Together's DeepSeek rows against DeepSeek's own. V4.1 Flash costs $0.30 and $1.20 here, which matches DeepSeek's peak-hour price and is double DeepSeek's off-peak $0.15 and $0.60. V4 Pro 0813 at $1.32 and $3.96 also matches DeepSeek's peak price. You pay that premium for a US company's terms in place of those of DeepSeek, and for no peak and off-peak schedule to manage. Check the cached rate too: $0.006 for Flash is the same as DeepSeek's peak cache-hit rate.

For the labs behind these models see Qwen models, GLM models, Kimi models and open source LLM.

Rate limits and capacity

Together's rate-limits page makes an unusual statement for a vendor: in most cases you can use serverless inference without encountering rate limits, and limiting happens with high request volumes or large bursts. It lists no per-tier request tables. Instead it explains how it protects performance.

• Best-effort serverless. Response times vary by model and demand. Together aims to keep tokens per second and time to first token high, and when demand is high it may limit requests.
• Errors you should handle. 429 Too Many Requests means reduce your rate and use exponential backoff. 503 Service Unavailable means wait and retry with backoff.
• Committed capacity. Provisioned throughput reserves capacity for a selected model or model family under a defined service level agreement. You contact sales for it.
• Contract limits. Enterprise and Scale contracts keep their purchased limits until the contract expires.
• Hot models. When a model is in especially high demand, Together may apply custom rate limits or access restrictions to it.

Read the third bullet as the real position. Serverless on Together is the cheap lane, and a production service with a latency promise belongs on provisioned or dedicated capacity. We did not find published prices for provisioned throughput.

Data retention and training

The privacy policy states: "Data that we collect, including your Personal Data, will not be used to train the Company's models without your explicit opt-in and consent." That is a clear default for training, and it is stronger than the opt-out posture of several vendors.

What we could not confirm is how long prompts and completions are kept. The policy describes the company's collection practices for the website, the programmatic APIs and professional support. We did not find a statement of API prompt retention in days on the pages we read. Ask for this in writing if your data is sensitive.

Two background points matter. First, Together serves open weights, so the model maker has no access to your prompts through Together. A DeepSeek model on Together is covered by Together's terms, not DeepSeek's. Second, a few models come with their own licences. Check the licence of the weights you choose, particularly for the ones listed under a custom licence name on Hugging Face. See Mistral AI for one example of a licence clause that matters when you resell inference.

Beyond serverless

The pricing navigation lists the rest of the platform, which matters if your needs grow.

Together's own advice on the pricing page is that most teams start with serverless and move to dedicated endpoints at scale. For GPU rental from other suppliers, see Runpod, Modal and Lambda Cloud.

Which agents can use it, and how

The API is OpenAI-style, so any agent with a custom endpoint option can use it. The base URL is https://api.together.ai/v1, and the key goes in the usual bearer header.

curl https://api.together.ai/v1/chat/completions \
  -H "Authorization: Bearer $TOGETHER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"openai/gpt-oss-120b","messages":[{"role":"user","content":"Summarize this diff."}]}'

Together lists a "Coding agent setup" page in its docs and a product called Together Link, described on its site as a way to run open models inside the harness you already use. We did not read the setup steps. OpenCode, Cline and Roo-style agents accept an OpenAI-compatible provider with a base URL and key. See OpenCode and Cline.

Picking a model from the shelf

A broad catalogue is only useful if you can narrow it. A simple order of work keeps the choice short.

1. Write ten real prompts from your own work, from a short question to a long code review.
2. Start with the cheapest plausible model, for example gpt-oss-120B or GLM-5.3-Flash, and score the outputs.
3. Step up one tier at a time only where the cheap model fails, for example to DeepSeek V4.1 Flash, then to GLM-5.3 or V4 Pro.
4. Keep Kimi K3 for the hard cases. At $3 and $15 it is the dearest row we list, so use it only where nothing cheaper passes.
5. Record the model id and snapshot you tested. Open models get new snapshots often, and a name such as "DeepSeek V4.1 Flash 1017" carries the date in the id for a reason.

Reading a licence before you ship

A host serving open weights does not remove the licence. Each model carries its own. DeepSeek V4 is MIT. gpt-oss is Apache-2.0. Others on the list carry custom licence names on Hugging Face, for example GLM-5.3 and Kimi K3, and Qwen3.8-2.4T-A95B is listed as a custom licence in our notes. Use the model card, not the host's page, as the source for what you may do with outputs and fine-tuned derivatives. Two clauses decide most disputes: whether you may resell inference to third parties, and whether a derived model inherits a display or attribution rule. Meta's Muse Spark Open licence, for instance, asks that products using it or a derivative show "Built with Muse Spark Open". See Meta Llama API for that licence.

Common errors and what to do

Together recommends that you limit retries to what your application's latency allows. For a user-facing product that means a fallback model on a second host, not an endless retry loop. A second host also protects you from the model-specific limits that Together says it can apply when a model is in especially high demand.

Moving to Together from another provider

1. Change the base URL and key. The request shape is the OpenAI one.
2. Change the model id to the Together name, which is usually the maker's organisation slash model name.
3. Re-run your evaluation set. A model hosted by a new provider can differ slightly in precision and sampling defaults.
4. Compare cost on your traffic, including cached tokens, because the cached rate differs far more across hosts than the headline input rate.
5. Get the data paperwork. Ask for the retention period and a data processing agreement before real data goes through.

Moving off is the same in reverse. Since the interface follows the OpenAI shape, a switch to Fireworks AI or Groq is a base URL change plus a model id check.

Alternatives

A worked cost example

Take a job that reads 10 million tokens and writes 2 million. The arithmetic is ours, from the table above, and uses cache-miss input prices.

The spread from the cheapest to the dearest row is more than twenty to one. As with any host, the saving comes from choosing the smallest model that passes your test. If a long prefix repeats, the cached price matters more than the headline: DeepSeek V4.1 Flash at a $0.006 cached rate turns a repeating 8 million token prefix into about five cents of input.

Who should pick Together AI

A pre-launch checklist

1. Get the API retention period in writing.
2. Pick the model by licence as well as price.
3. Add backoff for 429 and 503.
4. Move to provisioned throughput before a launch that needs latency promises.
5. Use cached input where your prompts repeat.
6. Compare the host price with the maker's own for the same model.
7. Check that the model you picked is still listed before each release.

Our take

Together is the broad shelf: more open models than most hosts, a clear non-training default, and a path from per-token to dedicated capacity. The honest costs are best-effort serverless performance, a retention period we could not find on the pages we read, and DeepSeek rows priced at the maker's peak rate. For a team that wants to try several open models, then settle on one under a US company's terms, it is a sensible first host. For one that needs a stated zero-retention promise today, Fireworks AI states it more plainly.

We have not run tests for this page. It reflects Together's published pages on 2026-10-31. Together does not pay for placement here.

Related pages

Open source LLM, LLM API prices, AI hosting, Fireworks AI, OpenRouter alternatives.