The Together AI alternative for open-model inference
Kimi K3 at $10 per million output tokens instead of $15. Gemma 4 31B at a third of the price. Plus 40,000+ models Together doesn't carry, on the same OpenAI-compatible API.
Curated catalogues decide for you
Together AI is a capable platform with strong engineering. But a curated catalogue means someone else decides which models are worth hosting — and you find out at the moment you need one that isn't on the list.
Waiting on availability
A model trends on Hugging Face. You wait for it to be added or you deploy it yourself.
Setup per model
Dedicated endpoints mean provisioning, warm-up and idle cost for every model you want to run.
Output tokens cost more
On Kimi K3, output runs 5x the input rate. Generation-heavy workloads pay the difference.
A hardware cost floor
Your rate per token is bounded by the silicon underneath it, and most providers run the same silicon.
Together AI pricing vs ours, model by model
Together AI's published serverless rates against ours, for the models both platforms serve. Cached input rates shown where Together publishes them.
| Model | Together in / out | Featherless in / out | Difference |
|---|---|---|---|
| Kimi K32780B MoE · 32K ctx | $3.00 / $15.00$0.30 cached in | $2.00 / $10.00 | −33% |
| DeepSeek V4 Pro862B · 256K ctx | $1.74 / $3.48$0.20 cached in | $1.60 / $3.20 | −8% |
| GLM-5.2753B · 256K ctx | $1.40 / $4.40$0.26 cached in | $1.00 / $4.00 | −14% |
| Kimi K2.6256K ctx | $1.20 / $4.50$0.20 cached in | $0.77 / $3.50 | −25% |
| Llama 3.3 70B | $1.04 / $1.04 | $0.65 / $0.75 | −33% |
| Qwen3.5 397B A17B256K ctx | $0.60 / $3.60$0.35 cached in | $0.55 / $3.50 | −4% |
| Gemma 4 31B | $0.39 / $0.97 | $0.14 / $0.40 | −60% |
| MiniMax M3427B · 256K ctx | $0.30 / $1.20$0.06 cached in | $0.55 / $2.20 | +83% |
| Qwen3.5 9B | $0.17 / $0.25 | $0.10 / $0.15 | −40% |
| gpt-oss-120B128K ctx | $0.15 / $0.60 | $0.10 / $0.55 | −13% |
| DeepSeek V4 Flash 0731284B · 256K ctx | $0.14 / $0.28$0.03 cached in | $0.14 / $0.28 | level |
| gpt-oss-20B128K ctx | $0.05 / $0.20 | $0.04 / $0.15 | −24% |
An agent workload on Kimi K3
100M input and 100M output tokens a month — a mid-size agentic product. Same model, same API, list rates on both sides.
100M input at $3.00 = $300. 100M output at $15.00 = $1,500. Output is where the cost sits on generation-heavy work.
100M input at $2.00 = $200. 100M output at $10.00 = $1,000. The $50 Developer plan includes $50 of credits against that.
Predictable pricing that scales with you
- Context size up to 256K
- 1 agent environment included
- Fastest response times
- Unused credits roll over
- Billed per token
- Dedicated H100, MI325, B200 & B300 GPUs
- Engineering team included
- Gets cheaper over time with fine-tuning
- Burst & failover to Public Cloud
A managed node for less than half a node elsewhere
Above a certain sustained volume, reserved capacity beats metered pricing on any platform. This is where the difference is largest. Together publishes dedicated inference at a per-GPU hourly rate, so the comparison is arithmetic.
4 × HGX H100 at $5.49 per GPU per hour, running continuously (730 hrs). A full 8-GPU node is $32,062. On B200 at $8.99/hr, the same half node is $26,251.
A dedicated AMD MI325X node, reserved and isolated. Inference stack tuning, quantisation, batching and on-call support are included rather than billed separately.
$3.99/GPU/hr works out at $11,651/mo for four H100s, which is still more than a fully managed node here,
and comes with no inference stack and no team.What the hardware sustains
| GPU | Memory | GPT-OSS 120B | Gemma 4 31B |
|---|---|---|---|
| AMD MI325X | 256 GB HBM3 | 500M tok/mo | 675M tok/mo |
| NVIDIA H100 | 80 GB HBM2e | 200M tok/mo | 70M tok/mo |
| NVIDIA B200 | 180 GB HBM3e | 1B tok/mo | 700M tok/mo |
| NVIDIA B300 | 288 GB HBM3e | 1.5B tok/mo | 1B tok/mo |
Switching from Together AI: a one-line change
Both APIs are OpenAI-compatible. If you're calling Together AI through the OpenAI SDK today, swap the base URL and key.
Get your API key# Before — Together AIclient = OpenAI( base_url="https://api.together.xyz/v1", api_key=os.environ["TOGETHER_API_KEY"] )# After — Featherlessclient = OpenAI( base_url="https://api.featherless.ai/v1", api_key=os.environ["FEATHERLESS_API_KEY"] )# Model names use the full Hugging Face
repo pathr = client.chat.completions.create( model="zai-org/GLM-5.2", messages=[{"role":"user","content":"Hello"}] )
Frequently asked questions
Featherless is the closest like-for-like Together AI alternative for open-model inference: the same OpenAI-compatible API and the same flagship models, with 40,000+ open models served on AMD hardware at lower per-token list rates on most flagships (Kimi K3 $12.00 vs $18.00 per 1M in + 1M out, August 2026). Together remains the better choice for managed fine-tuning, batch inference and raw GPU clusters.
No. Chat is sold for interactive, human-driven use by the subscriber — not for API traffic, background automation, reselling or benchmarking. If an application is making the calls, you want the Developer plan. We're explicit about this because signing up for the wrong plan is the single most common source of frustration when people switch.
Mostly hardware. We run on AMD MI-series accelerators rather than Nvidia, which costs less per unit of throughput, and we've done the kernel and serving work needed to actually use it. That's a structural cost difference rather than a promotional rate, which is why it doesn't expire.
For inference serving the gap with Nvidia is much narrower than it is for training, and on memory-bound workloads the MI325X's 256GB is an advantage — it sustains roughly 2.5x the GPT-OSS 120B throughput of an H100 on our infrastructure. Where it matters is model compatibility: most of the catalogue runs on AMD, but not every model does. If you have a specific model in production, ask before you migrate and we'll tell you straight. We also run Nvidia B200, B300 and H100 if that's the better fit.
Reserved, isolated GPU capacity plus the team to operate it, inference stack tuning, quantisation, batching, fine-tuning on your traffic, and on-call support. Not a separate professional services engagement. We benchmark your workload before you commit, then guarantee that performance level on the reserved capacity.
Not yet. Together publishes cached input rates and we don't currently match that, so on workloads with a high cache-hit rate, long shared system prompts, repeated context, the gap between us narrows. It's on the roadmap. If your workload is cache-heavy, factor that in when you compare.
Work out your monthly token volume and multiply by the per-token rate. Once that exceeds the monthly node cost, dedicated wins, and it also removes the variance from your bill. The cost calculator will do it for you, or talk to an engineer and we'll model it against your real traffic.
Fine-tuning is available on the Business tier, on your dedicated capacity. We don't rent raw GPU clusters for training, if that's your main requirement, Together AI is the better fit.
Run the numbers on your own workload
Estimate your cost in a minute, or get an API key and test on real traffic.
