Short version: Kimi K3 and DeepSeek V4 shipped open weights, but almost nobody wants to babysit a 32-GPU cluster to serve a 2.8-trillion-parameter model. So the choice became who serves it for you. Together AI and Fireworks AI both sell you tokens — pay per million, nothing when idle. Baseten sells you GPU-hours — a dedicated box you pay for around the clock. That single billing difference, not the price on the pricing page, is what decides your bill. Here's how to pick.
The three providers are not the same product#
It's tempting to line these up and compare price-per-token. Two of them don't even bill that way.
| Together AI | Fireworks AI | Baseten | |
|---|---|---|---|
| Billing | Serverless per-token | Serverless per-token | GPU-hour (dedicated + endpoints) |
| Pay when idle? | No | No | Yes (unless scaled to zero) |
| Headline rate | ~$0.03–$4.50 / M by model | Often lowest per-token on popular models | H100 ~$6.50/hr · B200 ~$9.98/hr |
| Batch discount | 50% async (24h window) | 50% async, flat off in + out | — |
| Kimi K3 day-zero | Yes | Yes (~200 tok/s) | Via dedicated deploy |
| Leans toward | Model breadth, long context | Raw throughput, low latency | Predictable p95 at steady load |
Together and Fireworks are serverless token vendors: you send a request, you pay for the tokens in and out, and when no requests arrive you pay nothing. Baseten is closer to a managed GPU vendor: you rent a box (an H100 at roughly $6.50/hr, a B200 at roughly $9.98/hr), and it's yours — running, and billing — whether or not traffic shows up. It offers scale-to-zero to blunt that, but scale-to-zero brings cold starts, which is exactly the latency you were paying dedicated hardware to avoid.
The question isn't "what's the cheapest token." It's "how busy will I keep the hardware." Those have opposite answers.
The number that actually decides it: your traffic shape#
Here is the break-even, and you can run it on the back of a napkin.
A single dedicated H100 at ~$6.50/hr runs about $4,700 a month if you hold it 24/7. At a serverless blended rate of roughly $2–3 per million tokens for an open model in Kimi K3's class, that same $4,700 buys you on the order of 1.5–2 billion tokens of on-demand inference.
So the rule is:
- Under ~1.5–2B tokens/month, or bursty traffic → serverless tokens (Together/Fireworks) win. You're not keeping a GPU busy enough to justify renting it whole.
- **Well above that, and steady enough to keep the GPU near full utilization** → dedicated GPU-hours (Baseten) win, because you stop paying the serverless markup on every token.
The trap is the word steady. A dedicated H100 that sits 30% utilized is worse than serverless on every axis — you're paying for 100% of a box to use a third of it. Serverless is the platform equivalent of only paying for the work you actually do; dedicated only pays off once "the work you actually do" fills the box.
Plug in your own numbers: **estimate monthly output tokens → multiply by your serverless per-token rate → compare to one dedicated GPU's monthly cost at the utilization you can realistically sustain.** If the serverless number is bigger and your traffic is steady, move to dedicated. Otherwise stay serverless.
Together vs Fireworks, if you've already ruled out dedicated#
Most teams start under the break-even, so the real fight is Together vs Fireworks. They're close, and both had Kimi K3 live on day zero, so pick on the axis that binds you:
- Fireworks leans speed. Built by former Meta PyTorch engineers, it competes on raw throughput and time-to-first-token — reportedly ~200 tokens/sec on Kimi K3 and P50 TTFT around 150ms on Llama 3.3 70B. If you're latency-bound on one hot model, benchmark Fireworks first.
- Together leans breadth. A wider catalog and strong long-context support make it the safer default when you're routing across several open models or leaning on a million-token window.
One honest caveat: the same open model can be priced ~20% apart between them (Llama 4 Maverick's input tokens have run about 20% higher on Together than Fireworks), purely from serving efficiency. Don't assume "same weights = same bill." And if you're only serving one small, popular model, check a specialist like Groq before you commit — on Llama 3.3 70B it has undercut both first-party.
The one lever almost nobody turns on#
If any part of your workload is asynchronous — nightly summarization, bulk classification, eval runs, embedding backfills — route it through the Batch API. Both Together and Fireworks apply an automatic 50% discount for async jobs (Together with a 24-hour processing window; Fireworks flat off both input and output). That's not a coupon; it's half your bill on every token that doesn't need to be real-time. We wrote the copy-paste setup in a companion piece: Batch inference and the 50% discount most teams never turn on.
The decision, in one line#
You are not choosing a price. You are choosing a billing model that matches your traffic: spiky or small → serverless tokens (Fireworks for speed, Together for breadth); big and steady → dedicated GPU-hours (Baseten), but only after you've measured that you'll keep the box busy. Everything else — the per-token sticker, the benchmark tweet — is downstream of that.
Adjacent reads: Kimi K3's Open Weights Are Public. Should You Self-Host? · Rent a GPU or Call an API? The Break-Even Math
