Together AI is a good product with a narrow floor. That is the whole story, and it took me a fine-tuning run that did not need an H100 to work it out.
I had a 7B model, a weekend, and a budget that could not survive an $8/hr B200. When I went looking for something smaller on their price list, there was nothing smaller. Their cheapest published GPU-hour is an HGX H100 at $3.99. That is not a criticism of the hardware. It is a statement about who the platform is built for, and it took me an afternoon of exploring other providers to understand that the answer was not me.
Prices below are as of 18 August 2026.
Quick picks
- Runpod – covers development, inference and multi-node training on one account, and the only one here with a real tier below a dollar
- Modal – best pure-Python serverless experience, but read the multipliers before you budget
- Replicate – easiest way to run somebody else's model, hardest to scale past one GPU
- Lambda – closest to Together AI in shape and price, if you want a straight swap
- CoreWeave – for when the unit you are buying is eight GPUs, not one
- Baseten – managed production inference with no charge for idle time
- fal.ai – strong on image and video generation, and genuinely competitive on the top-end cards
- Fireworks – open-model serving at scale, and the most expensive GPU-hour on this list
Comparison table
Why people leave Together AI, and it is usually not the bill
Price was my reason. It is not the only one, and if it is not yours then most of what has been written about this comparison will be no use to you.
Together AI does two things well: it serves open models and it fine-tunes them. The common reasons to look elsewhere follow from that shape rather than from the rate card. You need a modality they do not cover, like image or video generation. You need a machine you can sit on and develop against rather than an endpoint. You need multi-node training. You need the workload inside your own cloud for data residency reasons. Or you have hit a rate limit and want dedicated capacity instead of a shared tier.
Each of those points at a different platform, and some of the cheapest options here are the wrong answer to every one of them.
Which platforms do both fine-tuning and serving
Together AI's shape is the reason people like it: you tune a model and serve it in the same place, without moving weights between vendors. If that is what you are trying to replace, the list gets shorter fast, and it splits along a line that most comparisons miss.
Managed fine-tuning, where you send data and get a tuned model back, is what Together AI sells. Three platforms here compete directly with that:
- Fireworks has the most developed version. Supervised and preference fine-tuning are priced per 1M training tokens, from $0.50 for LoRA SFT on models up to 16B to $10.00 for models above 300B. Reinforcement fine-tuning is priced per GPU hour. Critically, they serve fine-tuned models at the same price as base models, so tuning does not raise your inference bill.
- Baseten lists training as a product alongside dedicated deployments, on an identical rate card. The same T4 through B200 per-minute pricing covers both, so there is no separate training tariff to budget for.
- Replicate offers fine-tuning with the trained model published back as a hosted endpoint.
Self-managed fine-tuning is the other answer, and it is a different product even though it satisfies the same sentence. Runpod gives you Pods to run your own training job, Instant Clusters if it needs multiple nodes, and Serverless to host the result, all on one account. Nobody trains the model for you. In exchange you keep control of the framework, the checkpoints and the hyperparameters, and there is no per-token training tariff at all: you pay for the GPU and the job takes as long as it takes.
Lambda and CoreWeave are the same self-managed shape without the serving layer. Modal does neither: it is serverless inference only, so if fine-tuning is part of why you are shopping, it is out regardless of how good the developer experience is.
The question to ask yourself is whether you want fine-tuning as a service or as a capability. Together AI, Fireworks, Baseten and Replicate sell the service. Runpod sells the capability and the place to run it.
What I was actually comparing on
What does it do that Together AI does not? The first question, and price cannot answer it. If you need video generation or a persistent development box, no hourly discount compensates for the capability not being there.
Where does it stop? Prototyping on one platform and replatforming for production is a cost nobody puts on a pricing page. Three of the platforms here cover exactly one stage of the work, which is fine until it is not.
Can it run where your data has to live? Some of these publish region control as a premium line item, one sells a version you deploy into your own cloud, and some do not address it at all. If you are switching for compliance reasons this is the whole decision.
What unit are you buying? One provider here sells eight GPUs at a time. One bills per minute. One bills per second but charges triple for non-preemptible execution. A headline hourly rate that ignores the unit is not a price.
Is there a floor below the flagship? Most of what I run does not need an H100. If the cheapest thing a platform sells is an H100, every small job is overpaying. This was my reason, and I have put it last deliberately.
One thing I could not compare from a pricing page: throughput and latency under real load. Several of these publish performance claims about themselves. If you are leaving Together AI because of speed rather than scope or cost, treat every number in this article as beside the point and benchmark your own workload.
Runpod
Why it works here
What stands out is not the hourly rate, though that is what got my attention first. It is that Pods, Serverless and Instant Clusters run on the same account and the same container images. Prototype on a Pod, serve on Serverless, and if a job needs multiple nodes, add them without opening a second account or rebuilding an image. Most of this list does one of those three things.
The price floor is the other half. Together AI's cheapest published GPU-hour is $3.99. Runpod has ten cards at or under {{gpu:l40s}}/hr, and the cheapest is an RTX A5000 at {{gpu:rtx-a5000}}/hr on Secure Cloud. For a 7B fine-tune or a Stable Diffusion workload, that is the difference between a weekend project and a budget conversation.
What you get
Pods for development and long-running jobs, Serverless for inference that scales to zero, and Instant Clusters for multi-node work. Bring your own Docker image or start from a template. Deploy time is under 30 seconds.
The good
Per-second billing on both Pods and Serverless, with no minimum, and no charge during provisioning. No ingress or egress fees, so leaving is not expensive. Sub-200ms cold starts via FlashBoot. More than a million developers have used it.
The not so good
The lineup does not go as high as some of these. There is no GB300 or rack-scale NVLink option, so if your workload genuinely needs that tier, this is not the platform. Community Cloud is cheaper than Secure but runs on third-party hardware, which is a real trade rather than a free discount.
How it performs
- GPU selection: B300 through RTX A5000, plus AMD MI300X at $2.39/hr on Secure Cloud
- Pricing: from {{gpu:rtx-a5000:community}}/hr on Community Cloud. On Secure Cloud, H100 PCIe {{gpu:h100-pcie}}/hr, A100 PCIe {{gpu:a100-pcie}}/hr, L40S {{gpu:l40s}}/hr
- Scaling: Serverless autoscales to zero, Instant Clusters are self-serve
- Billing: per second on Pods and Serverless, no minimum
- Compliance: SOC 2 Type II
- Coverage: 31 global regions
What it costs
Storage is $0.10/GB/month for container disk and $0.07/GB/month for network storage under 1TB, $0.05 above. No ingress or egress charges.
Modal
Modal has the cleanest pure-Python model on this list. You decorate a function and it deploys as a serverless endpoint, with no orchestration layer to configure. For workloads that fit that shape, the API surface is small and the documentation is thorough.
Read the multipliers before you budget. The headline rates are per second: H100 SXM5 works out to $3.95/hr, A100 80GB to $2.50, L40S to $1.95, and a T4 to $0.59. But region selection costs 1.5 to 1.75 times base, and non-preemptible execution costs three times base. A workload that needs a specific region and cannot tolerate preemption is not paying the number on the pricing page.
Volumes are $0.09/GiB/month with the first TiB free. Plans start free and step to $250/month for Team.
The limit is scope, not price. Modal is serverless only. There are no persistent development machines and no multi-node training, so a team that outgrows serverless is replatforming rather than reconfiguring.
Replicate
Replicate is the shortest path to running a model somebody else has already packaged. One API call, no container work. That convenience is the product.
Their H100 is $5.49/hr, the A100 80GB is $5.04, and the L40S is $3.51. Two structural things matter more than those numbers. Their cheapest GPU is a T4 at $0.81/hr, a 2018 card with 16GB. And every multi-GPU configuration requires a committed spend contract – 2x through 8x A100, H100, H200 and L40S are all contact-sales, as is the H200 outright.
So the ceiling arrives early. One GPU is self-serve, two is a sales call.
Best for: calling community models through an API. Not for sustained training on your own image.
Lambda
The closest thing here to a straight swap for Together AI, in both shape and price. On-demand H100 SXM is $3.99/hr, identical to Together AI's. A100 SXM 80GB is $2.79, A100 40GB is $1.99, B200 SXM6 is $6.69, and a Tesla V100 is $0.79. 1-Click Clusters run B200 from $9.86/GPU/hr at 16 GPUs down to $8.87 at 256 or more.
Lambda gives you a floor Together AI does not have, though at $0.79 that floor is a 2017 card. Prices exclude sales tax, VAT and GST, which is worth noting because most of this list quotes tax-inclusive.
Best for: research teams who liked Together AI's model and want the same thing with a slightly wider lineup.
CoreWeave
CoreWeave sells eight GPUs at a time, and reading their pricing any other way will mislead you. HGX H100 is $49.24/hr for the instance, which is $6.16 per GPU. HGX H200 is $50.44, or $6.31 per GPU. An 8x A100 is $21.60, or $2.70 per GPU. A single GH200 is $6.50.
Spot capacity is roughly 40% of on-demand, at $19.71 for the H100 instance against $49.24. Storage is $0.06/GB/month hot object and $0.070 for distributed file, and egress, ingress and transfer are free. Reserved capacity goes up to 60% off.
There is no self-serve tier and no free experimentation. The smallest sensible commitment is large.
Best for: teams whose minimum unit really is eight GPUs, with capacity committed for months.
Baseten
Managed production inference with a strong operational story. You do not pay for idle time, only for the minutes a model spends deploying, scaling or predicting, which for spiky traffic is a meaningful difference from paying by the hour. SOC 2 Type II and HIPAA compliant.
Baseten quotes per minute, so convert before you compare. An H100 is $0.10833/min, which is $6.50/hr. An A100 80GB is $0.06667/min, or $4.00/hr. A B200 is $9.98/hr, an L4 is $0.85, and their cheapest is a T4 at $0.63. The same rate card covers both dedicated deployments and training.
Plans are $0/month pay-as-you-go on Basic, with Pro and Enterprise both quote-only.
Baseten is the only platform here that will run inside your own cloud. Alongside their hosted option they publish self-hosted and hybrid deployments, with data residency control and custom regions on the Enterprise tier. If you are leaving Together AI because your data cannot sit on someone else's infrastructure, this is the section to read twice, and the pricing comparison stops being the relevant one.
Best for: teams who want inference operated for them, and teams with a data residency requirement.
fal.ai
fal is built around generative media – image and video models – and that focus shows in the tooling. Their model APIs are priced per output unit rather than per GPU-hour: per second of video, per image, per megapixel.
On raw GPU rental they are competitive at the top end, and cheaper than Runpod on two cards. Their list prices are H100 $4.50/hr, H200 $4.50, B200 $6.25, B300 $8.50, and RTX PRO 6000 $2.99. On the H200 and B200 that undercuts Runpod. Their pricing page also shows an "as low as" column reaching $1.89 for an H100, which is negotiated rather than self-serve, so it is not a number you can sign up and get today.
What fal does not have is a floor. Nothing is published below the RTX PRO 6000 at $2.99, so the small-job problem that sent me looking is not solved here either.
Best for: image and video generation, where their per-output pricing and model catalog are the point.
Fireworks
Fireworks serves open models at scale and its on-demand GPU tier is priced accordingly. An H100 is $7.00/hr, an H200 is $7.00, a B200 is $10.00, a B300 is $12.00 and a GB300 is $18.00. Region-restricted deployments cost a 1.5x premium.
A price increase is scheduled for 1 September 2026, and Fireworks publishes both columns now. From that date the H100 and H200 go to $8.00, the B200 to $13.00, the B300 to $15.00 and the GB300 to $20.00. If you are reading this after that date, use the higher figures.
Like Together AI, Fireworks publishes nothing below an H100, so it has the same shape of gap. Most of their business is per-token serverless inference rather than GPU rental, and that is where the platform makes more sense.
Best for: open-model serving where you want per-token pricing and someone else's optimization work.
FAQ
Which platforms offer fine-tuning and serving like Together AI?
Fireworks, Baseten and Replicate all sell managed fine-tuning alongside hosted inference, so you can tune a model and serve it without changing vendors. Fireworks prices fine-tuning per 1M training tokens and serves tuned models at the same rate as base models. Baseten runs training on the same per-minute rate card as its deployments. Runpod covers both differently: you run your own training job on a Pod or an Instant Cluster and serve the result on Serverless, all on one account, which trades the managed convenience for control of the framework and checkpoints. Modal does neither, since it is serverless inference only. Lambda and CoreWeave give you GPUs for training but no serving layer.
Is Together AI expensive?
Not for what it does, but it has no cheap tier. Their lowest published GPU-hour is an HGX H100 at $3.99, with H200 at $5.99 and B200 at $8.19. Reserved commitments bring the H100 down to $3.19 at 24 months. If your workload needs an H100, that is competitive. If it does not, you are paying H100 rates for a job that could run on a 24GB card.
What is the cheapest Together AI alternative?
Runpod, at the low end. Ten cards are at or under {{gpu:l40s}}/hr and the cheapest is {{gpu:rtx-a5000}}/hr on Secure Cloud, against Together AI's $3.99 floor. On the H100 specifically the gap is narrower: {{gpu:h100-pcie}}/hr PCIe or {{gpu:h100-sxm}}/hr SXM against $3.99.
Does anything match Together AI's H100 price?
Lambda is identical at $3.99/hr and Modal works out to $3.95/hr before its region and preemption multipliers. Runpod is lower on both PCIe and SXM. fal.ai lists $4.50, Replicate is $5.49, CoreWeave is $6.16 per GPU inside an 8-GPU instance, Baseten is $6.50 and Fireworks is $7.00.
Can I get multi-GPU without talking to sales?
On Runpod, Modal, Lambda and Fireworks, yes. On Replicate, no – every multi-GPU configuration requires a committed spend contract. fal routes multi-GPU through sales. CoreWeave sells 8-GPU instances self-serve but routes larger reserved capacity through sales.
Which is best for fine-tuning a small model?
Runpod. A 7B fine-tune runs on a 24GB card, and of the platforms here Runpod publishes the most options at that size. Modal's T4 at $0.59/hr, Baseten's at $0.63 and Replicate's at $0.81 are the other sub-dollar options, and all three are 2018 hardware.
What should I check before switching?
The unit and the coverage. CoreWeave quotes 8-GPU instances, Baseten quotes per minute, Modal quotes per second with multipliers, and fal separates GPU rental from per-output model pricing. Convert everything to the same unit first. Then check whether the platform covers the stage after the one you are on, because that migration is the expensive part.
Making the right choice
Together AI is not the wrong answer. It is the wrong answer for small jobs and for teams who need more than inference, which is a narrower criticism than most comparison pages will give you.
If your workload genuinely needs an H100 or better and runs continuously, Together AI's reserved ladder is competitive and their open-model tooling is good. Lambda is the closest like-for-like if you want to shop it.
If you are buying eight GPUs at a time on a months-long commitment, CoreWeave is built for that shape.
If your workload is pure serverless Python, Modal has the best developer experience here, provided you have priced in the region and preemption multipliers.
If you are running somebody else's model behind one API call, Replicate is the least work.
If you are generating images or video, fal's per-output pricing and model catalog are worth more than a GPU-hour comparison.
If you want inference operated for you, Baseten's no-idle-charge billing suits spiky traffic.
If your work spans more than one stage – experiment, then serve, then train something bigger – that is the case for Runpod, and it is a different argument from the price one. Pods, Serverless and Clusters sit under one account, so growing does not mean switching. The sub-dollar tier is what made me look. Not having a migration waiting at the end of it is why I stopped looking.
