Together AI hosts open models behind an OpenAI-compatible API, rents dedicated GPUs and clusters, and runs fine-tuning jobs. It is a capable, broad platform. This page is for people who have a specific reason to look at other options, and it is honest about when you should stay.
What Together AI does well
All figures from Together’s own pages, checked 2026-11-14.
- Open models, pay per token. Serverless examples: gpt-oss-120B at USD 0.15 input and USD 0.60 output per 1M tokens, DeepSeek V4.1 Flash at USD 0.30 and USD 1.20 (cached input USD 0.006), and Kimi K3 at USD 3.00 and USD 15.00.
- Range. The serverless list covers chat, vision, image, video and audio models; we counted about two dozen chat models.
- GPUs when you outgrow serverless. Dedicated inference on an H100 is USD 5.49 per GPU-hour; on-demand GPU clusters are USD 3.99 per H100-hour, and preemptible ones USD 1.99.
- Fine-tuning, function calling and JSON mode, each with its own docs.
- OpenAI compatibility. “The Together API is compatible with the OpenAI REST API and SDKs.” The base URL is
https://api.together.ai/v1. - Privacy controls. Its policy says “We do not use any data collected from you to train our models without your explicit opt-in and consent”, and you can turn on zero data retention by choosing “No” when asked whether to store prompts.
- Credits don’t expire. “Prepaid balance credits… do not currently have an expiration date.”
Why people look for a Together AI alternative
Each reason below comes from Together’s own docs.
- No free trial. “Together AI doesn’t currently offer free trials.” The first purchase is at least USD 15. Fine for a company; a hurdle for someone testing one prompt.
- Unpublished rate limits. “Serverless performance is best-effort”, and “you may receive a
429 Too Many Requestsor503 Service Unavailableresponse.” No per-tier numbers are listed. If you need to plan capacity, that is hard. Our 429 handling guide covers the client side. - Model churn. Standard serverless models get “2 or 3 weeks” of notice; preview models get under 24 hours. Llama-3.3-70B-Instruct-Turbo and DeepSeek-V4-Flash-0731 were removed on 2026-10-22, and gpt-oss-20b on 2026-09-15.
- Narrow batch discount. The batch docs say “Selected serverless models run at 50% off batch rates”, but only one model (whisper-large-v3) is listed; others run at standard rates.
- Per-token only. There is no flat or unlimited plan. An agent that runs all day has an open-ended bill.
- Zero balance stops everything. “If your balance reaches zero, API access is suspended until you add credits.”
Together AI alternatives compared
| Together AI | Fireworks AI | DeepInfra | Groq | BroadNet Inference | |
|---|---|---|---|---|---|
| Pricing model | Per token; GPUs per hour | Per token; GPUs per second | Per token or execution time; GPUs per hour | Per token | Per token, per audio minute, flat plans |
| DeepSeek V4.1 Flash, in / out per 1M | USD 0.30 / 1.20 | USD 0.30 / 1.20 | USD 0.20 / 0.60 list; discounted USD 0.14 / 0.42 (1017 build USD 0.098 / 0.294) | Not offered | $0.175 / $0.52 |
| Flat plan | No | No (Nexus plans are capped) | No | No | Yes: unlimited DeepSeek V4.1 Flash, $60/month per session |
| OpenAI-compatible | Yes | Yes | Yes (“may not be 100% compatible”) | Yes (“mostly compatible”) | Yes |
| Data retention (cited) | No prompt or response retention once you choose ZDR | Doesn’t log prompts for open models without opt-in; ZDR excludes Batch | “we do not store the data you submit to our APIs on disk” | Not retained by default; logs up to 30 days for abuse unless ZDR | “keeps no prompts, messages, images, audio, responses or transcripts” |
| Model range | Chat, image, video, audio | 300+ listed | Very wide catalogue | 14 models | Six: four chat, two speech-to-text |
| Free credit | None; USD 15 minimum | Not stated on pricing page | None; card or prepay required | Free plan with rate limits | $5, no card |
Sources for each cell are linked below. For a wider price table, see LLM inference provider pricing compared.
8 alternatives to Together AI at a glance
Prices per 1M tokens (input / output) from each provider’s own pages, checked 2026-11-14.
- DeepInfra: wide catalogue and 200 concurrent requests per model. DeepSeek V4.1 Flash lists at USD 0.20 and USD 0.60, currently discounted to USD 0.14 and USD 0.42.
- Fireworks AI: open models with fine-tuning, per-second GPUs and 50% batch and cached-input discounts. DeepSeek V4.1 Flash is USD 0.30 and USD 1.20.
- Groq: very fast LPU inference on 14 models. gpt-oss-120b is USD 0.15 and USD 0.60 at about 500 tokens a second, with a free plan.
- Novita AI: OpenAI-compatible open-model host. DeepSeek V4.1 Flash is USD 0.30 input and USD 1.20 output per 1M tokens, gpt-oss-120b USD 0.05 and USD 0.25, and batch has an introductory 50% discount.
- Baseten: Model APIs plus dedicated GPUs. DeepSeek V4.1 Flash is USD 0.30 and USD 1.20 (cached input USD 0.007); an H100 is about USD 6.50 an hour.
- SambaNova: gpt-oss-120b at USD 0.22 input and USD 0.59 output, and GLM 5.3 at USD 2.80 and USD 8.80. Its pricing page lists no DeepSeek model.
- Cerebras: built for speed. Its docs list GPT OSS 120B at about 3,000 tokens a second for USD 0.35 and USD 0.75, on two shared models, with a USD 5 trial that needs a card.
- BroadNet Inference: six models, flat unlimited DeepSeek V4.1 Flash plans ($60 a month per session), zero retention, and $0.175 / $0.52 per token. This is BroadNet’s own product.
When to stay with Together AI
- You fine-tune. Together runs training jobs and serves the result on the same platform. None of the flat-plan options do that.
- You need dedicated GPUs or clusters from one vendor, with an uptime SLA on provisioned throughput.
- You use image, video or audio generation alongside chat.
- Your traffic is bursty and modest. Per-token billing is cheapest when you don’t use much, and credits that never expire suit irregular use.
When BroadNet Inference fits
BroadNet Inference serves six models under zero data retention; BroadNet’s privacy statement says they run on BroadNet-owned or leased US hardware or, for partner-served models such as DeepSeek V4.1 Flash, with US inference partners. It is narrower than Together. It fits when:
- An agent runs all day on DeepSeek V4.1 Flash. A flat-rate plan costs $60 a month per session with no token meter. Take a full-time coding agent using 1 billion input tokens a month (80% cached) and 50 million output tokens. At Together’s prices: 200M uncached x USD 0.30 + 800M cached x USD 0.006 + 50M output x USD 1.20 = about USD 124.80. At DeepInfra’s list price the same month is about USD 74.80, at its current 30% discount about USD 52, and on the discounted 1017 build about USD 37. So the plan beats Together and DeepInfra’s list price here, but not DeepInfra’s current discounts. Check your own numbers with the flat-rate vs per-token calculator.
- You want zero retention without a setting. The gateway “keeps no prompts, messages, images, audio, responses or transcripts” on every request. See zero data retention.
- You want to test before paying. New accounts get $5 of credit with no card. It covers Gemma 4 VL 26B and speech to text; DeepSeek, GLM and Qwen open after the first top-up.
- You need a 1M-token context window on DeepSeek V4.1 Flash or GLM 5.3.
Per token on DeepSeek V4.1 Flash, DeepInfra’s discounted price is lower than BroadNet’s. The case for BroadNet is the flat plan, not the meter.
When another alternative fits better
- Fireworks AI if you want Together’s feature set somewhere else: full-parameter training up to 1T-parameter models (in preview), a 50% batch discount, cached input “at 50% off by default”, and on-demand H100s at USD 8.00 an hour billed by the second.
- DeepInfra if per-token price is the main goal. Its list prices are lower on the open models we checked, batch is billed 20% below real-time, and it allows 200 concurrent requests per model. You need a card or prepayment to start.
- Groq if latency matters most and its 14 models cover your needs. Its free plan lets you test before paying, within published rate limits.
How to switch from Together AI
Change three things: base URL, key and model ID.
from openai import OpenAI
import os
client = OpenAI(
base_url="https://inference.broadnet.ai/v1", # was https://api.together.ai/v1
api_key=os.environ["BROADNET_API_KEY"],
)
resp = client.chat.completions.create(
model="broadnet/deepseek-v4.1-flash", # was deepseek-ai/DeepSeek-V4.1-Flash on Together
messages=[{"role": "user", "content": "Summarise this changelog in five bullets."}],
max_tokens=2000,
)
Then:
- Log a week of Together usage per model: input, cached input and output tokens.
- Price it per token and on a plan with the LLM API cost calculator. Include caching; the prompt caching calculator shows what a cache hit rate saves.
- Run your test prompts on both for a day. Model builds and quantisation differ between hosts, so outputs can differ.
- Move one workload at a time. The Python SDK guide and migration checklist cover the details.
The short version
Together AI is a good default for open models, especially with fine-tuning or dedicated GPUs. If price per token is your problem, look at DeepInfra first. If a growing per-token bill for one always-on agent is the problem, a flat plan is the cleaner fix, and that is where BroadNet Inference fits. Read more on DeepSeek API alternatives.