Together AI vs Fireworks AI: Which Open-Model API Fits?
November 19, 2026

Together AI vs Fireworks AI is closer than most comparisons make it look. Both serve the same popular open models behind an OpenAI-compatible API, and on several models both list, the standard per-token rates were identical when we checked in November 2026. So the per-token price won't decide it for you.
What separates them is everything around the API:
- Pick Together AI if you also want raw GPU clusters, code sandboxes, or a training service where you write the loop and can download the weights.
- Pick Fireworks AI if you want more serverless modes (batch, spot, priority, fast), the newest speed tier, and model-routing tools like Fireworks Nexus.
- Know this before you fine-tune on Fireworks: its managed SFT and DPO stopped accepting new jobs. Training now runs through its Training API, where a Python loop on your side drives the job.
This guide uses only each vendor's own pricing pages, docs and announcements. Speed figures are vendor claims and are labelled that way.
Quick Comparison
| Together AI | Fireworks AI | |
|---|---|---|
| Core product | Open-model inference plus GPU compute | Open-model inference and training |
| API | OpenAI-compatible | OpenAI-compatible |
| Serverless options | Serverless, Batch Inference, Provisioned Throughput | Standard, Priority, Fast, Batch, Spot, Ultra-fast (preview) |
| Dedicated capacity | Dedicated Model Inference, Dedicated Container Inference | On-demand deployments per GPU-hour, reserved GPUs |
| Raw GPU clusters | Yes, self-serve instant clusters | GPUs sold as deployments and training |
| Managed fine-tuning | Yes: LoRA or full, SFT or DPO | Closed to new jobs; use the Training API |
| Write-your-own training loop | Together Post-Training (SFT, GRPO, PPO) | Training API (SFT, DPO, RL) |
| Serving LoRA adapters | Dedicated endpoints | Dedicated deployments; serverless adapters in preview |
| Extras | Code Sandbox, Code Interpreter, managed storage | Nexus and Nexus Desktop, No Eval (private preview), Fireworks Sovereign (private preview) |
Serverless Inference: Same Models, Different Modes
Switching between the two takes minutes. Together's docs say its API "is compatible with the OpenAI REST API and SDKs", so you change the API key and base URL. Fireworks works the same way with the OpenAI client. If your code already talks to OpenAI, it talks to both.
The catalogs overlap heavily. Both list GLM 5.3, DeepSeek V4.1 Flash, Kimi K3 and gpt-oss-120B, and for those models the input, cached-input and output rates on Together's pricing page matched the standard rates on Fireworks' pricing page when we compared them. Run your own list through both pages, because each vendor also carries models the other doesn't, and rates move.
The real difference is how many ways you can buy those tokens.
Fireworks sells several serverless modes:
- Standard, the default.
- Priority, at a higher rate, for traffic that can't be slowed at peak load.
- Fast, aimed at 100+ tokens per second.
- Batch, at 50% off input and output, for work that doesn't need an immediate answer.
- Spot, 30% off Standard on select models, with throughput that can drop to zero.
- Ultra-fast, in preview on NVIDIA Blackwell GPUs.
Cached input is billed at half price or less by default, with no write fee.
Together offers Serverless Inference, Batch Inference and Provisioned Throughput, which reserves capacity for steady traffic. Its pricing page also shows cached-input rates per model.
On speed, both make claims you should test. Fireworks says its inference delivers up to 4x the throughput of providers wrapping vLLM, and its Forge announcement cites up to 1,500 tokens per second on DeepSeek V4.1 Flash in Ultra-fast mode. Together's research team builds FlashAttention and the ATLAS speculator system, and it recently introduced Together Compiler. None of that tells you how your prompts will perform. Send the same prompts to both and measure time to first token and tokens per second.
Dedicated GPUs and Raw Compute
Once traffic is steady, or you want a model neither catalog carries, you move to dedicated capacity.
Fireworks on-demand deployments bill per GPU-hour for H100, H200 and B200 GPUs. The pricing page says there's no charge for start-up time, idle deployments scale to zero after an hour, region-restricted deployments cost 1.5x, and reserved GPUs are cheaper per hour. Elastic Burst, now generally available, lets you go past your usual ceiling when spare capacity exists.
Together splits dedicated inference in two:
- Dedicated Model Inference for serving a model on GPUs reserved for you.
- Dedicated Container Inference for running your own container.
It lists on-demand per-GPU-hour rates and sends reserved pricing to sales. On the pages we checked, Together's on-demand H100 rate for dedicated inference was below Fireworks' on-demand H100 rate. Compare current numbers yourself.
Together also rents GPUs as GPUs. Its Accelerated Compute line offers self-serve instant clusters, with preemptible, on-demand and reserved H100 and B200 rates on the pricing page, plus managed storage with no egress fees. Fireworks' pricing page sells GPU time as model deployments and training, not as clusters you manage. Its closest move toward your own hardware is Fireworks Sovereign, a private preview that brings its inference engine into your own cloud, a neocloud or your data centers.
| You need | Better fit |
|---|---|
| Run your own container on dedicated GPUs | Together (Dedicated Container Inference) |
| A training cluster you control | Together GPU clusters |
| Dedicated deployments that scale to zero when idle | Fireworks on-demand |
| The inference engine inside your own environment | Fireworks Sovereign (private preview) |
Fine-Tuning and Training: A Recent Change at Fireworks
This section changed this month, so older comparisons are out of date.
Fireworks closed managed SFT and DPO to new jobs. Its migration guide says managed SFT and DPO "no longer accept new jobs" and tells you to train on the Training API instead. Jobs created before November 9, 2026 keep running, and models you already trained stay deployable.
The Training API works differently from a managed job. You run a Python training loop from Fireworks' cookbook, and the docs say that loop "runs in your Python process for the whole job, not on Fireworks." There are two flavours:
- Serverless Training: LoRA only, billed per token, on listed models. Supports SFT, DPO and RL.
- Dedicated Training: LoRA or full-parameter, billed per GPU-hour, on any model with a Training API shape. Supports SFT, DPO, ORPO and RL.
Fireworks also announced Specialized Model Inference in public preview. It serves LoRA adapters serverlessly, starting with GLM 5.3 Flash, with per-token pricing and no charge while idle.
Together keeps a managed service and adds a hands-on one. Together's fine-tuning service runs LoRA (the default) or full fine-tuning, with SFT or DPO, and handles data upload, training and hosting on a dedicated endpoint. It bills by tokens processed, with a minimum charge per job, and the docs say you can download the model for local use.
For reinforcement learning, Together points you to Together Post-Training. Its pitch is "You own the training loop; Together handles the hard infra." You write the recipe in Python, including data, rewards and loss. Together runs the trainer, the rollout engine and weight sync, and bills by the GPU-minute while your resource is ready. It supports SFT and RL with GRPO, PPO or custom losses, and you can serve the result on dedicated inference or download the weights.
| Together AI | Fireworks AI | |
|---|---|---|
| Upload data, get a model back | Yes (fine-tuning service) | Not for new jobs |
| You write the loop | Post-Training | Training API |
| RL | Post-Training (GRPO, PPO, custom) | Training API |
| Billing | Tokens (managed); GPU-minutes (Post-Training) | Tokens (Serverless Training); GPU-hours (Dedicated) |
| Weights download | Yes | Check the docs for your model |
If you want the simplest path, upload a dataset and get a fine-tuned model back, Together is the one that still offers it. If you want full control of an RL loop, both now let you write it.
What Each Sells Beyond the API
Together leans toward compute. Next to inference it sells GPU clusters, a Code Sandbox and a Code Interpreter for agents (billed by vCPU and memory, or per session), and managed storage. Its research output, from FlashAttention-4 to the Mamba-3 architecture, is part of the pitch.
Fireworks leans toward model selection and evaluation. The Forge announcement on November 3, 2026 added:
- Nexus Desktop, which joins the Nexus CLI. Nexus routes routine work to open models and keeps closed models for hard tasks. Fireworks says it now runs in the Codex app, Claude Desktop and OpenCode. Nexus has its own paid plans, separate from the API.
- Ember-1, a Fireworks model it says used about 40% fewer tokens than Kimi K3 at comparable quality in its own evaluations.
- No Eval (private preview), which turns production traces into task-specific evaluations.
- Enterprise controls, with model governance, spending controls and audit logs generally available.
Together AI vs Fireworks AI: Which Should You Pick
Choose Together AI if:
- You want inference and raw GPU clusters from one vendor.
- You want to upload a dataset and get a fine-tuned model back without writing a training loop.
- You need sandboxes for agent code next to your models.
- Downloading your trained weights matters to you.
Choose Fireworks AI if:
- You want the most serverless buying options, especially Batch and Spot for offline work.
- Latency is the product, and you want to test the Ultra-fast tier.
- You want dedicated deployments that scale to zero when idle.
- Routing across open and closed models with Nexus appeals to your team.
Either works for a first build on a popular open model. The APIs are interchangeable enough that you can keep both keys and route by cost or speed.
If neither fits, the wider field is in our Together AI alternatives and Fireworks AI alternatives guides. If you'd rather deploy models on your own GPUs, see Modal vs Runpod. For raw speed, Groq vs Cerebras covers two specialist chips.
Where Taku Fits
Together and Fireworks are APIs for developers who build on open models. If you'd rather use an AI tool someone has already built than wire up an inference provider, that's where Taku comes in. Taku Apps and Sites, built by the best AI minds. Find one, run it, and keep it in your Mac Dock. Taku is in Beta, and the Mac app is available now.
Key Points
- Both offer OpenAI-compatible APIs for the same popular open models, and their standard per-token rates matched on several shared models in November 2026.
- Fireworks has more serverless modes: Priority, Fast, Batch at 50% off, Spot at 30% off and an Ultra-fast preview.
- Together sells raw GPU clusters, sandboxes and managed storage alongside inference.
- Fireworks managed SFT and DPO stopped taking new jobs. Training now runs through its Training API, driven by your own Python loop.
- Together still runs a managed fine-tuning service, and Together Post-Training lets you write your own SFT or RL loop.
- Fireworks' Forge launches (Nexus Desktop, Ember-1, No Eval, Sovereign) point toward model routing and evaluation.
FAQ
Is Together AI or Fireworks AI cheaper?
On several models both carry, such as GLM 5.3 and gpt-oss-120B, the standard per-token rates were the same when we checked. Fireworks can come out cheaper for offline work through Batch and Spot. For dedicated GPUs, Together's listed on-demand H100 rate was lower. Check both pricing pages for your models.
Which is faster, Together AI or Fireworks AI?
Both claim speed leadership. Fireworks cites up to 1,500 tokens per second on DeepSeek V4.1 Flash in its Ultra-fast preview, and Together points to its FlashAttention and ATLAS research. Benchmark your own prompts on both, measuring time to first token and tokens per second.
Can I still fine-tune models on Fireworks?
Yes, but not through managed SFT or DPO jobs, which stopped accepting new jobs. You use the Training API, either Serverless Training (LoRA, per token) or Dedicated Training (LoRA or full-parameter, per GPU-hour), driven by a Python loop you run.
Does Together AI support reinforcement learning?
Yes, through Together Post-Training. You write the recipe in Python, and Together runs the trainer and rollout engine. It supports GRPO, PPO or custom losses and bills by the GPU-minute.
Can I switch from Together AI to Fireworks AI easily?
Usually. Both are OpenAI-compatible, so you change the base URL, the API key and the model name. Fine-tuned models take more work: you'd retrain on the other platform or bring the weights, so check each vendor's docs on importing a model.
What is Fireworks Nexus Desktop?
It's a Fireworks product announced at Forge on November 3, 2026. It connects everyday tools to Fireworks Nexus with one switch per tool. Nexus routes routine work to open models and keeps closed models for hard tasks, and Fireworks says it runs in the Codex app, Claude Desktop and OpenCode.