Deepak Gupta markTools
AI Tools · Fine-Tuning and Model Customization

Top 8 Fine-Tuning and Model Customization Platforms 2026: What It Costs, When It Wins, and Why OpenAI Is Shutting Its Own Down

Together, Fireworks, Unsloth, Axolotl, AWS Bedrock, Google Vertex AI, OpenAI and Predibase compared on training price, inference cost after tuning, and the lock-in each path creates. Prices taken from each provider's own pricing page on 2026-09-18.

By ·Sep 18, 2026·15 min·8 tools compared
Fine-TuningLoRAModel CustomizationTogether AIFireworks AIUnslothAxolotlAWS BedrockVertex AIAI Tools

The honest answer first

Most teams asking "which fine-tuning platform should we buy" should not buy one yet. The question underneath it is almost always "why is the model getting this wrong", and fine-tuning only fixes one of the three possible causes.

  • The model lacks facts. Fix with retrieval, not tuning. A fine-tuned model that does not know your product catalogue will state the wrong price with more confidence, not less. Start at Top 5 RAG-as-a-Service Platforms 2026.
  • The model is inconsistent in form. Wrong format on 2% of calls, drifting tone, ignoring a required schema at the tail of the distribution. This is the case fine-tuning genuinely solves, and prompting will keep failing at scale.
  • The model is correct but too slow or too expensive. That is a distillation problem: train a small model on a large one's outputs and serve the small one.

OpenAI made the same argument itself. On 7 May 2026 it began winding down its fine-tuning API and platform, stating that its newer base models follow instructions and formats well enough to make much of the process unnecessary. New organisations were blocked that day, organisations with no recent fine-tuned inference were blocked on 2 July 2026, and every remaining customer loses the ability to create new fine-tuning jobs on 6 January 2027. Inference on existing fine-tuned models continues only until the base models are deprecated.

That is worth pausing on, because it is the single most expensive lock-in event this category has produced. Teams that fine-tuned a proprietary model hold no weights, no adapter and no migration path. The ones who kept their training data and evaluation set face a few weeks of rebuilding. The ones who did not face a rewrite.

So when does fine-tuning win

Run this test before you spend anything:

  1. Re-run your evaluation set against the current generation of base models with a well-structured prompt. It is free, and in 2026 it frequently passes where it failed a year ago.
  2. If it still fails, ask whether the failure is missing knowledge. If yes, build retrieval.
  3. If the failure is form, behaviour or consistency, and you have at least a few thousand correctly labelled examples, fine-tune.
  4. If the output is right but the model is too expensive, distil into a smaller one.

If you cannot produce a labelled evaluation set, stop. Every platform below is machinery for a dataset you do not have yet, and assembling it is the actual project.

Where the money actually goes

The training run is the cheapest part and it is the part everyone budgets for. Take 20,000 examples averaging 500 tokens, trained for three epochs: 30M training tokens.

Route Training cost for that run What serving it then costs
Fireworks LoRA SFT, 70B model About $90 at $3.00 per 1M tokens Base-model inference rates, no premium
Together SFT, Llama 3.3 70B About $61 at $2.03 per 1M tokens Serverless, or dedicated HGX H100 at $3.99/GPU-hour promotional through 09/30/26
OpenAI, gpt-4.1 About $750 at $25.00 per 1M tokens $3.00 in / $12.00 out per 1M tokens, until the platform closes
Unsloth or Axolotl on rented GPUs The GPU hours, often under $20 Whatever you build
AWS Bedrock Per token, published by model provider Provisioned Throughput required: Llama 2 13B is published at $23.50 per model-unit hour with no commitment, roughly $17,000 a month

The last row is the one that ends projects. Every price above was read from the provider's own pricing page on 18 September 2026.

Quick Comparison

PlatformBest ForWhat You Can CustomisePublished Training Price (checked 2026-09-18)What Inference Costs After TuningLock-In
Together AIOpen-weight SFT and DPO with a price you can compute in advanceOpen-weight models, LoRA and full fine-tuningPer 1M training tokens: Qwen3.5 9B $0.34, Gemma 4 31B $1.05, Llama 3.3 70B $2.03, DeepSeek-V3.1 $7.00. DPO 2-4x those rates. Per-job minimum $4 to $100Serverless or dedicated endpoints; dedicated HGX H100 $3.99/GPU-hour (promotional through 09/30/26, list $5.49), HGX B200 $8.99/GPU-hourLow. Open weights, and the adapter concept is portable
Fireworks AICheapest published LoRA training, and no inference premium for the resultOpen-weight models, LoRA and full-parameter SFT and DPOLoRA SFT per 1M training tokens: $0.50 (up to 16B), $3.00 (16.1-80B), $6.00 (80-300B), $10.00 (>300B). Full-parameter SFT is 2x those. DPO is 2x SFT"Serve fine-tuned models for the same price as base models." On-demand GPUs also published: H100 $8.00/hr, H200 $8.00/hr, B200 $13.00/hr, B300 $15.00/hr, GB300 $20.00/hrLow. Open weights, LoRA adapters
UnslothTraining on your own GPU, for free, with the smallest memory footprintOpen-weight LLMs, embedding models and MoE models; LoRA, QLoRA and full fine-tuning$0. Core package is Apache 2.0; Unsloth Studio UI is AGPL-3.0. Claims 2x faster training with 70% less VRAM, and 12x faster MoE training with 35% less VRAMWhatever you pay for your own GPU, or any serving platform that accepts the weightsNone. You hold the weights
AxolotlReproducible multi-GPU training runs defined in a config fileOpen-weight models; LoRA, QLoRA, full fine-tuning, DPO, IPO, KTO, ORPO, GRPO, GDPO, with FSDP1, FSDP2 and DeepSpeed$0, Apache 2.0. Dedicated support available by enquiry, no published priceWhatever you pay for your own GPU or rented computeNone. You hold the weights
AWS BedrockRegulated enterprises that need customisation inside an existing AWS boundarySelected hosted models including Claude 3 Haiku, Llama and Cohere Command; provider-specific methodsPer token trained, varying by provider. AWS publishes Meta Llama 2 at $1.49 per million tokens trained and Cohere Command at $0.004 per 1,000 tokens. Custom model storage $1.95 per model per monthCustom models require Provisioned Throughput. AWS publishes Llama 2 13B at $23.50/hour per model unit with no commitment, $21.18/hour on a 1-month commitment and $13.08/hour on a 6-month commitmentHigh. Custom weights stay inside Bedrock
Google Vertex AIGemini customisation for teams already standardised on Google CloudGemini supervised fine-tuning, preference tuning, plus supervised and distillation tuning for open modelsGoogle publishes supervised tuning rates per 1M training tokens on its Vertex AI generative AI pricing page. The pricing table did not render a quotable figure when we checked on 2026-09-18, so no rate is stated hereTuned Gemini models are served through Vertex AI; confirm the serving rate and any Provisioned Throughput requirement in your quoteHigh. A tuned Gemini model exists only inside Vertex AI
OpenAINobody new. The platform is being withdrawnGPT-4.1, 4.1-mini, 4.1-nano, GPT-4o, GPT-4o-mini supervised; o4-mini reinforcement fine-tuningPer 1M training tokens: gpt-4.1 $25.00, gpt-4o $25.00, gpt-4.1-mini $5.00, gpt-4o-mini $3.00, gpt-4.1-nano $1.50. o4-mini reinforcement fine-tuning $100.00 per hourPer 1M tokens, input/output: gpt-4.1 $3.00/$12.00, gpt-4o $3.75/$15.00, gpt-4.1-mini $0.80/$3.20, gpt-4o-mini $0.30/$1.20, gpt-4.1-nano $0.20/$0.80Total, and now terminal. New job creation ends 6 January 2027
Predibase by RubrikExisting customers; new buyers face a sales process with no published priceOpen-weight models, LoRA and Turbo LoRA supervised tuning and reinforcement fine-tuning (GRPO)No published price. predibase.com and the docs pricing page both redirect to rubrik.com as of 2026-09-18Not publishedHigh, and now tied to an enterprise data-security vendor's roadmap

Together AI

Best For
Open-weight SFT and DPO with a price you can compute in advance
What You Can Customise
Open-weight models, LoRA and full fine-tuning
Published Training Price (checked 2026-09-18)
Per 1M training tokens: Qwen3.5 9B $0.34, Gemma 4 31B $1.05, Llama 3.3 70B $2.03, DeepSeek-V3.1 $7.00. DPO 2-4x those rates. Per-job minimum $4 to $100
What Inference Costs After Tuning
Serverless or dedicated endpoints; dedicated HGX H100 $3.99/GPU-hour (promotional through 09/30/26, list $5.49), HGX B200 $8.99/GPU-hour
Lock-In
Low. Open weights, and the adapter concept is portable

Fireworks AI

Best For
Cheapest published LoRA training, and no inference premium for the result
What You Can Customise
Open-weight models, LoRA and full-parameter SFT and DPO
Published Training Price (checked 2026-09-18)
LoRA SFT per 1M training tokens: $0.50 (up to 16B), $3.00 (16.1-80B), $6.00 (80-300B), $10.00 (>300B). Full-parameter SFT is 2x those. DPO is 2x SFT
What Inference Costs After Tuning
"Serve fine-tuned models for the same price as base models." On-demand GPUs also published: H100 $8.00/hr, H200 $8.00/hr, B200 $13.00/hr, B300 $15.00/hr, GB300 $20.00/hr
Lock-In
Low. Open weights, LoRA adapters

Unsloth

Best For
Training on your own GPU, for free, with the smallest memory footprint
What You Can Customise
Open-weight LLMs, embedding models and MoE models; LoRA, QLoRA and full fine-tuning
Published Training Price (checked 2026-09-18)
$0. Core package is Apache 2.0; Unsloth Studio UI is AGPL-3.0. Claims 2x faster training with 70% less VRAM, and 12x faster MoE training with 35% less VRAM
What Inference Costs After Tuning
Whatever you pay for your own GPU, or any serving platform that accepts the weights
Lock-In
None. You hold the weights

Axolotl

Best For
Reproducible multi-GPU training runs defined in a config file
What You Can Customise
Open-weight models; LoRA, QLoRA, full fine-tuning, DPO, IPO, KTO, ORPO, GRPO, GDPO, with FSDP1, FSDP2 and DeepSpeed
Published Training Price (checked 2026-09-18)
$0, Apache 2.0. Dedicated support available by enquiry, no published price
What Inference Costs After Tuning
Whatever you pay for your own GPU or rented compute
Lock-In
None. You hold the weights

AWS Bedrock

Best For
Regulated enterprises that need customisation inside an existing AWS boundary
What You Can Customise
Selected hosted models including Claude 3 Haiku, Llama and Cohere Command; provider-specific methods
Published Training Price (checked 2026-09-18)
Per token trained, varying by provider. AWS publishes Meta Llama 2 at $1.49 per million tokens trained and Cohere Command at $0.004 per 1,000 tokens. Custom model storage $1.95 per model per month
What Inference Costs After Tuning
Custom models require Provisioned Throughput. AWS publishes Llama 2 13B at $23.50/hour per model unit with no commitment, $21.18/hour on a 1-month commitment and $13.08/hour on a 6-month commitment
Lock-In
High. Custom weights stay inside Bedrock

Google Vertex AI

Best For
Gemini customisation for teams already standardised on Google Cloud
What You Can Customise
Gemini supervised fine-tuning, preference tuning, plus supervised and distillation tuning for open models
Published Training Price (checked 2026-09-18)
Google publishes supervised tuning rates per 1M training tokens on its Vertex AI generative AI pricing page. The pricing table did not render a quotable figure when we checked on 2026-09-18, so no rate is stated here
What Inference Costs After Tuning
Tuned Gemini models are served through Vertex AI; confirm the serving rate and any Provisioned Throughput requirement in your quote
Lock-In
High. A tuned Gemini model exists only inside Vertex AI

OpenAI

Best For
Nobody new. The platform is being withdrawn
What You Can Customise
GPT-4.1, 4.1-mini, 4.1-nano, GPT-4o, GPT-4o-mini supervised; o4-mini reinforcement fine-tuning
Published Training Price (checked 2026-09-18)
Per 1M training tokens: gpt-4.1 $25.00, gpt-4o $25.00, gpt-4.1-mini $5.00, gpt-4o-mini $3.00, gpt-4.1-nano $1.50. o4-mini reinforcement fine-tuning $100.00 per hour
What Inference Costs After Tuning
Per 1M tokens, input/output: gpt-4.1 $3.00/$12.00, gpt-4o $3.75/$15.00, gpt-4.1-mini $0.80/$3.20, gpt-4o-mini $0.30/$1.20, gpt-4.1-nano $0.20/$0.80
Lock-In
Total, and now terminal. New job creation ends 6 January 2027

Predibase by Rubrik

Best For
Existing customers; new buyers face a sales process with no published price
What You Can Customise
Open-weight models, LoRA and Turbo LoRA supervised tuning and reinforcement fine-tuning (GRPO)
Published Training Price (checked 2026-09-18)
No published price. predibase.com and the docs pricing page both redirect to rubrik.com as of 2026-09-18
What Inference Costs After Tuning
Not published
Lock-In
High, and now tied to an enterprise data-security vendor's roadmap
1

Together AI

Best Overall

Best for: Teams that want supervised or preference fine-tuning on open weights with a cost they can calculate before they start

“Together publishes a per-model, per-1M-token price for supervised fine-tuning and direct preference optimisation across its open-weight catalogue, which makes it the easiest platform here to budget against. Llama 3.3 70B supervised tuning is $2.03 per million training tokens, so a 30M-token run costs about $61. Because the models are open weights, nothing about the decision traps you: you can move the workload, and in most cases the artefact, elsewhere.”

Pros

  • Per-model training prices are published for both supervised fine-tuning and DPO, so a run can be costed before it is launched
  • DPO is available as a first-class option at 2 to 4 times the SFT rate, which matters when you are tuning for preference rather than format
  • Both LoRA and full fine-tuning are offered
  • The same account covers serverless inference, dedicated endpoints and GPU clusters, so training and serving do not require two vendors
  • Dedicated inference on HGX H100 is listed at $3.99 per GPU-hour under a promotion running through 09/30/26, against a $5.49 list price

Cons

  • Every job carries a model-specific minimum charge of $4 to $100, which makes small experimental runs disproportionately expensive
  • The pricing page does not state whether LoRA and full fine-tuning are billed differently; both are listed as available under one price table
  • The pricing page does not document how a fine-tuned model is deployed or what a LoRA adapter costs to serve, so that has to be confirmed separately
  • The $3.99 H100 dedicated inference rate is promotional and dated, so the post-promotion price is the one to plan against
Honest Weakness: The per-job minimum charge is the trap for exactly the team that should be using this platform. Fine-tuning is iterative, and the honest workflow is ten small cheap runs before one real one. On a model with a $100 minimum, those ten experiments cost $1,000 regardless of how few tokens they consume. Check the minimum for your specific model before you design your experiment plan, and do early iteration on a small model or on your own GPU with Unsloth.

What a real run costs

Take 20,000 training examples averaging 500 tokens each. That is 10M tokens per epoch, or 30M tokens across three epochs. On Llama 3.3 70B at $2.03 per million, the supervised run costs roughly $61. The same run as DPO at $5.08 per million costs roughly $152. Those are small numbers, and they are the reason people underestimate fine-tuning: the training bill is almost never what makes the project expensive.

What happens at inference time

This is where the cost actually lands. A fine-tuned open-weight model has to be served, either serverlessly, on a dedicated endpoint, or on your own hardware. Together publishes dedicated endpoint pricing per GPU-hour, so a single always-on HGX H100 endpoint at the promotional $3.99 per GPU-hour is roughly $2,900 a month for eight GPUs at full utilisation. Compare that against your current per-token API spend before you commit.

Lock-in

Low, and this is the strongest argument for the open-weight path generally. The base models are open weights and the tuning artefact is a checkpoint or an adapter. Moving to Fireworks, to your own GPUs, or to a self-hosted vLLM deployment is an engineering task rather than a retraining project.

Supervised fine-tuning per 1M training tokens: Qwen3.5 9B $0.34, Gemma 4 31B $1.05, Llama 3.3 70B $2.03, DeepSeek-V3.1 $7.00. Direct preference optimisation is 2 to 4 times higher: Qwen3.5 9B $0.84, Llama 3.3 70B $5.08, DeepSeek-V3.1 $17.50. All jobs carry a model-specific minimum charge of $4 to $100. Dedicated inference: HGX H100 $3.99/GPU-hour promotional through 09/30/26 against a $5.49 list price, HGX B200 $8.99/GPU-hour. Checked on together.ai/pricing, 2026-09-18.

Visit Together AI
2

Fireworks AI

Best Value

Best for: LoRA fine-tuning where the point is a cheap adapter that costs nothing extra to serve

“Fireworks publishes the clearest fine-tuning price structure in this comparison: a flat per-1M-token rate banded by model size, doubled for full-parameter training and doubled again for DPO. The decisive line is on serving. Fireworks states that fine-tuned models are served at the same price as base models, which removes the single largest hidden cost of customisation. If the plan is a LoRA adapter on an open-weight model, start here.”

Pros

  • LoRA supervised fine-tuning starts at $0.50 per 1M training tokens for models up to 16B, the cheapest published rate here
  • Fine-tuned models are served at base-model inference rates, with no adapter premium
  • The price bands are simple and model-agnostic, so you can budget without looking up individual models
  • Full-parameter training and DPO are available with published multipliers rather than a sales call
  • On-demand GPU pricing is also published (H100 $8.00/hr, H200 $8.00/hr, B200 $13.00/hr, B300 $15.00/hr, GB300 $20.00/hr) for teams that want to run their own deployment

Cons

  • Published GPU rates are high against dedicated GPU clouds; $8.00 per H100-hour is roughly double Crusoe's $3.90 and Lambda's $3.99
  • Rates are banded by parameter count, so a model at the top of a band is priced the same as one at the bottom
  • Full-parameter training at 2x the LoRA rate climbs quickly: a >300B model is $20.00 per 1M training tokens
  • The platform's breadth means fine-tuning shares a roadmap with inference serving rather than owning one
Honest Weakness: The generous serving policy is subsidised by the serving rates. Fireworks charges $8.00 per H100-hour for on-demand GPUs when Crusoe publishes $3.90 and Lambda publishes $3.99 for the same chip. If you fine-tune cheaply here and then serve heavily here, the saving on training is recovered several times over on compute. The platform is excellent value for low-to-moderate serverless volume and poor value as a place to park a permanently busy dedicated deployment.

Why serving at base-model price matters more than training price

Training is a one-off cost measured in tens or hundreds of dollars. Serving is a recurring cost measured in thousands a month. Every platform that charges a premium to serve a customised model, or that requires dedicated capacity to serve one at all, has moved the real cost to the line item people forget to model. Fireworks removing that premium is a bigger commercial fact than any of its training rates.

What a real run costs

The same 30M-token run described above, on a 70B model, costs $90 as LoRA SFT and $180 as full-parameter SFT. Doubling again for DPO gives $180 and $360. Those are rounding errors against a team's monthly inference bill, which is the point: choose the method for quality reasons, not to save $90.

Fine-tuning per 1M training tokens. LoRA SFT: $0.50 (up to 16B), $3.00 (16.1-80B), $6.00 (80-300B), $10.00 (over 300B). LoRA DPO: $1.00, $6.00, $12.00, $20.00. Full-parameter SFT: $1.00, $6.00, $12.00, $20.00. Full-parameter DPO: $2.00, $12.00, $24.00, $40.00. Serving: fine-tuned models are charged at base-model rates. On-demand GPUs per hour: H100 80GB $8.00, H200 141GB $8.00, B200 180GB $13.00, B300 288GB $15.00, GB300 288GB $20.00. Checked on fireworks.ai/pricing, 2026-09-18.

Visit Fireworks AI
3

Unsloth

Best Open Source

Best for: Running the first twenty experiments on hardware you already have, at zero licence cost

“Unsloth is an open-source training library that cuts the memory and time cost of fine-tuning enough to move a class of runs from rented multi-GPU nodes onto a single card. The project claims 2x faster training with 70% less VRAM and no accuracy loss, and 12x faster MoE training with 35% less VRAM. The core package is Apache 2.0, multi-GPU is supported in the free version, and there is nothing to buy. This is where iteration should happen before you spend money anywhere else.”

Pros

  • Free, with the core package under Apache 2.0 and no paid gate on the training path
  • Claims 2x faster training with 70% less VRAM and no accuracy loss versus baseline
  • Claims 12x faster MoE training with 35% less VRAM, which is the harder case
  • Multi-GPU setups are supported in the free version across NVIDIA, AMD and Intel hardware
  • A native desktop application shipped on 10 August 2026 for Mac, Windows and Linux, lowering the entry barrier considerably
  • You hold the resulting weights, so there is no platform lock-in of any kind

Cons

  • Dual licensing means the optional Unsloth Studio UI is AGPL-3.0, which some legal teams will not accept alongside proprietary code
  • It is a library, not a platform: no hosted jobs, no experiment tracking, no dataset management, no serving
  • You supply, configure and pay for the GPU, and you own every failure mode of the training run
  • Performance claims are the project's own and are not independently verified here
Honest Weakness: Unsloth solves the compute-efficiency problem and none of the operational ones. There is no job queue, no artefact registry, no lineage between a dataset version and a checkpoint, and no serving path. Teams that adopt it for cost reasons often rebuild a worse version of a managed platform around it over the following six months. It is the right tool for experimentation and for organisations with real ML infrastructure, and the wrong one for a product team that wants to fine-tune once and move on.

Where the savings come from

The memory reduction is what changes the decision, more than the speed. Cutting VRAM by 70% can move a run that needed an 80GB card onto a 24GB one, which changes the hourly rate from roughly $2 to roughly $0.30 on a GPU marketplace. Across the twenty throwaway experiments a real fine-tuning project actually needs, that is the difference between a meaningful budget line and a rounding error.

How to use it alongside a platform

The practical pattern is iterate locally with Unsloth on a small model, prove the dataset and the objective are sound, then run the production training job on Together or Fireworks against the model you actually intend to serve.

Free. Core package licensed Apache 2.0; optional components including Unsloth Studio UI are AGPL-3.0. No Pro or Enterprise tier is described in the project README. You pay only for the compute you run it on. Checked on github.com/unslothai/unsloth, 2026-09-18.

Visit Unsloth
4

Axolotl

Best Free Option

Best for: Reproducible, config-driven training runs across multiple GPUs and multiple post-training methods

“Axolotl is a free and open-source fine-tuning framework under the Apache 2.0 licence, driven by a YAML config rather than bespoke training code. It covers the widest method surface here: LoRA, QLoRA and full fine-tuning, preference optimisation via DPO, IPO, KTO and ORPO, reinforcement methods including GRPO and GDPO, and distributed training with FSDP1, FSDP2 and DeepSpeed. If you need a method a managed platform does not offer, this is usually where it lives first.”

Pros

  • Free and Apache 2.0 licensed, with no paid tier required for any training method
  • Broadest method coverage in this comparison, including GRPO and GDPO reinforcement training
  • Config-file driven, so a training run is a reviewable artefact in your repository rather than a notebook someone ran once
  • First-class distributed training via FSDP1, FSDP2 and DeepSpeed for multi-node runs
  • Dedicated support is available by enquiry for teams that need a commercial relationship

Cons

  • No published price for the commercial support option, so that path requires a conversation
  • No hosted compute, no job scheduler, no artefact registry and no serving
  • The configuration surface is large, and a misconfigured run is an expensive way to learn that
  • Requires real ML engineering capability on the team, which most product teams do not have
Honest Weakness: Axolotl's flexibility is its cost. Because it exposes every post-training method and every distribution strategy, there is no guard rail telling you that GRPO is the wrong choice for your problem or that your learning rate will destroy the base model's instruction following. The managed platforms constrain you into the small number of configurations that usually work. If nobody on the team can read a loss curve and say why it looks wrong, that constraint is worth more than the flexibility.

Method coverage

Supervised fine-tuning in LoRA, QLoRA and full-parameter forms; preference optimisation via DPO, IPO, KTO and ORPO; reinforcement training via GRPO and GDPO; distributed training via FSDP1, FSDP2 and DeepSpeed. That is a superset of what any managed platform on this page exposes self-serve.

Where to run it

Axolotl needs GPUs and does not supply them. Pair it with a GPU cloud rather than a hyperscaler: on-demand H100 capacity is published at $3.90 per GPU-hour on Crusoe and $3.99 on Lambda, and preemptible H100 on Together is $1.99, which suits fault-tolerant training well.

Free, Apache 2.0 licence. No hosted offering or published pricing. The documentation directs teams needing dedicated support to contact the project by email, with no rates published. You pay only for the compute you run it on. Checked on docs.axolotl.ai, 2026-09-18.

Visit Axolotl
5

AWS Bedrock model customization

Best for Enterprise

Best for: Regulated enterprises that need a customised model to stay inside an existing AWS account boundary

“Bedrock is the only managed route to a fine-tuned Claude model: Anthropic does not expose fine-tuning through its own API, and Claude 3 Haiku supervised fine-tuning has been generally available on Bedrock in US West (Oregon) since November 2024. Training is billed per token, custom model storage is $1.95 per model per month, and then the real number arrives: a custom model requires Provisioned Throughput to serve. Buy it for the compliance boundary, not for the economics.”

Pros

  • The only managed path to fine-tuning a Claude model, via Claude 3 Haiku supervised fine-tuning
  • Runs inside an existing AWS account, with the IAM, VPC, logging and procurement posture a regulated buyer already has approved
  • Training billed per token, with published example rates (Meta Llama 2 at $1.49 per million tokens trained, Cohere Command at $0.004 per 1,000 tokens)
  • Custom model storage is a flat $1.95 per model per month, which makes holding several variants cheap
  • Provisioned Throughput commitments cut the hourly rate substantially: AWS publishes Llama 2 13B at $23.50/hour with no commitment, $21.18/hour at 1 month and $13.08/hour at 6 months

Cons

  • Custom models cannot use on-demand inference; Provisioned Throughput is required, and it is billed by the hour whether or not you send traffic
  • At the published Llama 2 13B example rate, one always-on model unit is roughly $17,000 a month with no commitment and roughly $9,500 a month on a 6-month commitment
  • Claude fine-tuning is limited to Claude 3 Haiku and to one region, US West (Oregon)
  • Training rates vary by model provider and are published only as examples, so your model's rate needs checking directly
  • Custom weights stay inside Bedrock, so the artefact is not portable
Honest Weakness: Provisioned Throughput turns fine-tuning from a variable cost into a fixed one, and almost nobody models it before the project starts. On AWS's own published example, a single Llama 2 13B model unit is $23.50 an hour with no commitment, which is roughly $17,000 a month for a model that may be handling a few thousand requests a day. Teams arrive at Bedrock customisation to save money on tokens and leave with a six-month capacity commitment. Compute the monthly Provisioned Throughput bill in your first meeting, not your last.

The Claude question

Anthropic does not offer fine-tuning through its own API. Claude 3 Haiku supervised fine-tuning became generally available on Amazon Bedrock in US West (Oregon) in November 2024 and remains the only managed route to a customised Claude model. If your requirement is specifically a fine-tuned Claude, Bedrock is not one option among several, it is the option, and the Provisioned Throughput cost comes with it.

Why it still wins for some buyers

For an organisation whose security review already covers AWS, the alternative is not a cheaper platform, it is a nine-month procurement exercise for a new vendor. That is a real cost and it is frequently larger than the compute premium. The mistake is paying the Bedrock premium without that constraint actually applying.

Training charged per token, varying by model provider. AWS publishes Meta Llama 2 at $1.49 per million tokens trained and Cohere Command at $0.004 per 1,000 tokens of training data as examples. Custom model storage $1.95 per model per month. Inference requires Provisioned Throughput: AWS publishes Llama 2 13B at $23.50 per hour per model unit with no commitment, $21.18 per hour on a 1-month commitment and $13.08 per hour on a 6-month commitment. Checked on aws.amazon.com/bedrock/pricing, 2026-09-18.

Visit AWS Bedrock model customization
6

Google Vertex AI model tuning

Runner Up

Best for: Teams standardised on Google Cloud that want to customise Gemini rather than an open-weight model

“Vertex AI offers supervised fine-tuning and preference tuning for Gemini models, plus supervised and distillation tuning for open models, inside the Google Cloud boundary. It is the natural choice if your data already lives in BigQuery and your security review already covers Google Cloud. The caveat is commercial transparency: we could not extract a quotable tuning rate from Google's pricing page on the date of checking, so treat the price as something to confirm in writing.”

Pros

  • Supervised fine-tuning and preference tuning for Gemini models, with documented support across text, document, image, audio, video and function-calling modalities
  • Also supports supervised and distillation tuning for open models, so both paths live in one platform
  • Tuning checkpoints and continuous tuning are documented features, which suits an iterative programme
  • Runs inside an existing Google Cloud project with the IAM and data-residency posture already approved
  • Natural fit when the training data already lives in BigQuery or Cloud Storage

Cons

  • Google's generative AI pricing page did not render a quotable supervised tuning rate when checked on 2026-09-18, so the cost per run has to be confirmed directly
  • A tuned Gemini model exists only inside Vertex AI; there is no portable artefact
  • The serving arrangement for a tuned model, including whether Provisioned Throughput is required, needs confirming in your quote
  • Product naming has shifted toward the Gemini Enterprise Agent Platform, which makes documentation harder to navigate than it should be
Honest Weakness: The pricing is the weak point, and not because it is high. Google publishes supervised tuning rates, but they sit inside a large dynamic pricing table that resists a simple reading, and the naming of the surrounding product has changed. A buyer cannot do what they can do on Together or Fireworks in thirty seconds: read a number, multiply by their token count, and decide. Insist on a written per-1M-training-token rate and a written serving rate before you commit, and do not accept a third-party site's figure as a substitute.

What you can tune

Gemini supervised fine-tuning and preference tuning, plus supervised and distillation fine-tuning for open models. Distillation is the underrated option: training a small model on a large model's outputs is often a better answer than fine-tuning the large one, and it directly attacks inference cost rather than output quality.

Lock-in

High. A tuned Gemini model is a Vertex AI object. Leaving means retraining on a different base model with a different tokenizer and different behaviour, which is a project rather than a migration. Weigh that against the convenience of staying in one cloud.

Google publishes supervised tuning rates per 1M training tokens on its Vertex AI generative AI pricing page. That table did not render a quotable figure in our check on 2026-09-18, so no rate is asserted here. Confirm the per-1M-training-token rate, the serving rate for a tuned model, and any Provisioned Throughput requirement directly with Google before budgeting. Checked on cloud.google.com/vertex-ai/generative-ai/pricing, 2026-09-18.

Visit Google Vertex AI model tuning
7

OpenAI fine-tuning

Honorable Mention

Best for: Nobody starting today. Existing customers should be planning their exit

“OpenAI announced on 7 May 2026 that it is winding down the fine-tuning API and platform, on the stated basis that newer base models follow instructions and formats well enough to make much of the process unnecessary. New organisations were cut off the same day. Organisations without recent fine-tuned inference were cut off on 2 July 2026. Every remaining customer loses the ability to create new fine-tuning jobs on 6 January 2027. Inference on existing fine-tuned models continues until the base models are deprecated.”

Pros

  • Inference on models you already fine-tuned continues to work until the underlying base model is deprecated, so there is no immediate outage
  • Published training and inference prices remain documented, which makes exit planning possible
  • The nano and mini tiers were genuinely cheap to train: gpt-4.1-nano at $1.50 and gpt-4o-mini at $3.00 per 1M training tokens
  • The stated reason, that current base models need less format-shaping, is a real change and applies to your decision regardless of vendor

Cons

  • Closed to new organisations since 7 May 2026
  • Closed since 2 July 2026 to organisations that had not run inference on a fine-tuned model in the preceding 60 days
  • All remaining customers lose new job creation on 6 January 2027, which means no retraining on fresh data after that date
  • There is no export path: an OpenAI fine-tuned model has never been a portable artefact
  • Training was the most expensive on this list at $25.00 per 1M tokens for gpt-4.1 and gpt-4o
Honest Weakness: This is the clearest lock-in lesson available anywhere in AI infrastructure, and it cost real teams real money. Organisations that fine-tuned a proprietary model have no weights, no adapter and no migration path. Their only options are to keep running inference on a model they can no longer retrain until it is deprecated, or to rebuild on an open-weight platform from their original dataset. If you kept your training data and your evaluation set, the rebuild is a few weeks. If you did not, it is a rewrite. Keep the dataset, whatever you choose.

What to do if you are an existing customer

Three steps, in order. Confirm you still hold the training dataset and the evaluation set that justified the fine-tune in the first place. Re-run that evaluation set against the current base model with a good prompt, because OpenAI's stated reason for the wind-down is that this now often passes. If it does not, rebuild on an open-weight platform where the artefact is yours, and use the remaining window to do it calmly.

What it means for everyone else

The wind-down is a data point about the whole category. The largest vendor in the market has concluded that for most customers, better base models plus better prompting beat customisation. That does not make fine-tuning wrong, it narrows where it wins: consistent output format at scale, a narrow domain with idiosyncratic language, latency and cost reduction by distilling into a smaller model, and behaviour that prompting cannot make reliable.

Training per 1M tokens: gpt-4.1 $25.00, gpt-4o $25.00, gpt-4.1-mini $5.00, gpt-4o-mini $3.00, gpt-4.1-nano $1.50. o4-mini reinforcement fine-tuning $100.00 per hour. Inference per 1M tokens, input/output: gpt-4.1 $3.00/$12.00, gpt-4o $3.75/$15.00, gpt-4.1-mini $0.80/$3.20, gpt-4o-mini $0.30/$1.20, gpt-4.1-nano $0.20/$0.80. Batch pricing is 50% of standard rates for most models. Wind-down dates: 7 May 2026 (new organisations blocked), 2 July 2026 (inactive organisations blocked), 6 January 2027 (all new job creation ends). Checked on developers.openai.com pricing and deprecations pages, 2026-09-18.

Visit OpenAI fine-tuning
8

Predibase by Rubrik

Honorable Mention

Best for: Existing customers. New buyers should expect a sales process and no published pricing

“Predibase built a strong reputation for efficient LoRA and Turbo LoRA fine-tuning and for shipping an end-to-end reinforcement fine-tuning workflow, served through the LoRAX multi-adapter stack. Rubrik, a data security and backup vendor, announced its acquisition on 25 June 2025 and closed it in August 2025. As of 2026-09-18 both predibase.com and the documentation pricing page redirect to rubrik.com, and the product is positioned as Predibase by Rubrik.”

Pros

  • Genuine technical depth in parameter-efficient fine-tuning and multi-adapter serving
  • One of the first platforms to offer an end-to-end managed reinforcement fine-tuning workflow
  • Now backed by a larger public company, which reduces the risk of an outright shutdown
  • Existing enterprise customers gain access to Rubrik's security and governance posture

Cons

  • No published pricing: both the marketing pricing page and the documentation pricing page redirect to rubrik.com as of 2026-09-18
  • No self-serve signup path visible from the redirected pages
  • Product roadmap is now set by a data-security company's agentic AI strategy rather than by a fine-tuning roadmap
  • Evaluating it requires a sales cycle, which rules it out for the exploratory phase where fine-tuning decisions are actually made
Honest Weakness: The problem is not the technology, it is that you can no longer price the technology without talking to someone. Every other platform in this comparison except Promptfoo's enterprise tier lets you read a number and model a run in a minute. A fine-tuning decision is made by an engineer running cheap experiments, and a platform that cannot be tried without a procurement conversation is effectively unavailable to that engineer. Treat it as an enterprise option to consider if you are already a Rubrik customer, and not as a candidate for your first fine-tune.

What changed

Predibase was an independent, self-serve fine-tuning platform with published pricing. It is now an enterprise product line inside a data-security company, reached through rubrik.com. The technology, LoRA and Turbo LoRA tuning, GRPO reinforcement fine-tuning and LoRAX multi-adapter serving, appears to be intact. The commercial model is not the one you may remember.

No published pricing as of 2026-09-18. predibase.com redirects to rubrik.com/products/rubrik-agent-cloud and docs.predibase.com/user-guide/pricing redirects to rubrik.com. Rubrik announced the acquisition on 25 June 2025 and it closed in August 2025. Contact sales for rates. Checked 2026-09-18.

Visit Predibase by Rubrik

Which One Should You Pick?

Use CaseOur Recommendation
You are not sure fine-tuning is the right answer at allIt probably is not, yet. Run your evaluation set against the current base model with a well-structured prompt and a retrieval step first. Fine-tuning teaches form and behaviour; retrieval supplies facts. If the failure is that the model does not know something, fine-tuning will not fix it and will make the model confidently wrong instead.
The model keeps breaking your required output format at scaleThis is the strongest fine-tuning case. Supervised fine-tuning on a few thousand correctly formatted examples fixes format reliability far better than prompt instructions do. Fireworks LoRA SFT at $0.50 to $3.00 per 1M training tokens, served at base-model rates, is the cheapest credible route.
You want the same quality at lower latency and lower costDistillation, not fine-tuning for quality. Train a small open-weight model on a larger model's outputs, then serve the small one. Google Vertex AI offers managed distillation tuning; Axolotl and Unsloth do it on your own hardware for free.
You need a fine-tuned Claude modelAWS Bedrock is the only managed path, via Claude 3 Haiku supervised fine-tuning in US West (Oregon). Anthropic does not expose fine-tuning through its own API. Budget for Provisioned Throughput, because custom models cannot use on-demand inference.
You are an existing OpenAI fine-tuning customerPlan the exit now. New job creation ends for everyone on 6 January 2027 and inference survives only until the base model is deprecated. Re-test your evaluation set against the current base model first, then rebuild on Together or Fireworks if it still fails.
You want to run twenty cheap experiments before spending real moneyUnsloth on a single GPU. The claimed 70% VRAM reduction moves many runs onto a 24GB card, which on a GPU marketplace is cents per hour. Prove the dataset and the objective there before paying per-token minimums on a managed platform.
You need GRPO or another reinforcement method a managed platform does not exposeAxolotl, which supports GRPO and GDPO alongside DPO, IPO, KTO and ORPO with FSDP and DeepSpeed distribution. Pair it with rented GPU capacity: preemptible H100 on Together is $1.99 per GPU-hour, suited to fault-tolerant training.
Your data cannot leave a specific cloud accountAWS Bedrock or Google Vertex AI, depending on which boundary your security review already covers. Both cost more than the open-weight platforms and both keep the artefact inside the cloud. That is the trade you are making, and it is sometimes correct.

How we evaluated

Last verified: 18 September 2026. Every price, limit and availability claim on this page was read from the provider's own pricing page, documentation or deprecation notice on that date. Nothing is carried from a third-party comparison site or a pricing aggregator. Where a provider does not publish a figure, the page says so explicitly rather than filling the gap.

Two gaps are stated openly. Google publishes supervised tuning rates on its Vertex AI generative AI pricing page, but that table did not render a quotable figure in our check, so no Vertex tuning rate is asserted here. Predibase publishes no pricing at all as of the check date: both predibase.com and its documentation pricing path redirect to rubrik.com.

The criteria, in the order that decides real outcomes:

  • Training price, expressed in a unit you can multiply. Per 1M training tokens where published, so a run can be costed before it is launched. Per-job minimum charges are treated as a first-class cost because they dominate the experimental phase.
  • What inference costs after tuning. The largest and most frequently ignored number. A platform that serves a fine-tuned model at base-model rates and a platform that requires hourly provisioned capacity are not in the same price bracket, whatever their training rates say.
  • Method coverage. LoRA, QLoRA, full-parameter, DPO and other preference methods, reinforcement methods, and distillation. Distillation is included because it is often the correct answer to the problem people bring to fine-tuning.
  • Lock-in. Whether the artefact is a portable checkpoint or adapter, a cloud-resident object, or a vendor-controlled model you cannot export. Graded explicitly on every entry.
  • Availability as a purchase. Whether a new buyer can sign up and run a job today. This removed OpenAI from consideration for new projects and reduced Predibase to an enterprise sales conversation.

Vendor selection and changes since the last market

Eight platforms are listed because each represents a distinct path: managed open-weight tuning (Together, Fireworks), self-run open-source training (Unsloth, Axolotl), cloud-platform customisation (AWS Bedrock, Google Vertex AI), first-party proprietary tuning (OpenAI), and the specialist platform tier (Predibase).

Three findings changed the shape of this comparison:

  • OpenAI is withdrawing. The fine-tuning API and platform wind-down was announced on 7 May 2026 with a final cutoff of 6 January 2027 for new job creation. It is listed for the benefit of existing customers, not as a recommendation.
  • Anthropic has no first-party fine-tuning. Claude 3 Haiku supervised fine-tuning is available only through Amazon Bedrock, generally available in US West (Oregon) since November 2024. There is no Anthropic-hosted alternative, so Anthropic does not appear as a standalone entry.
  • Predibase is no longer an independent purchase. Rubrik announced the acquisition on 25 June 2025 and closed it in August 2025. As of 2026-09-18 the domain and documentation pricing page redirect to rubrik.com and no self-serve pricing is published.

What we did not do

No platform on this page was benchmarked, and no fine-tuning run was performed. There are no hands-on testing claims here. Quality outcomes from fine-tuning depend almost entirely on your dataset, not on the platform, which is why this comparison is built on published cost structure, method coverage and portability rather than on quality claims nobody can generalise.

Note

Editorial independence: this is a vendor-neutral comparison with no paid placements, sponsorships or affiliate links. Rankings reflect fit for the stated use cases, not commercial relationships.

Sources

Frequently Asked Questions

Should we fine-tune at all?
Most teams should exhaust prompting and retrieval first, and a meaningful fraction never need to go further. The test is diagnostic rather than aspirational. If the model lacks facts, that is a retrieval problem and fine-tuning will make it confidently wrong. If the model is inconsistent in format, tone or task structure across thousands of calls, that is a fine-tuning problem and prompting will keep failing at the tail. If the model is correct but too slow or too expensive, that is a distillation problem. OpenAI's own stated reason for winding down its fine-tuning platform in 2026 is that current base models follow instructions and formats well enough that much of this work is no longer needed. That is a data point worth taking seriously before you start.
When does fine-tuning actually win?
Four situations, reliably. Consistent structured output at high volume, where prompt-based formatting fails on a small percentage of calls and that percentage is unacceptable. A narrow domain with idiosyncratic vocabulary or conventions that the base model was not trained on, such as an internal ticket taxonomy or a specialist legal register. Cost and latency reduction, by distilling a large model's behaviour into a small one you can serve cheaply. And behaviour the model resists, where prompting works in testing and drifts under real inputs. Everything else, especially knowledge, belongs in retrieval.
What does one fine-tuning run cost?
Less than people expect, and that is the misleading part. Take 20,000 examples averaging 500 tokens, trained for three epochs, which is 30M training tokens. On Fireworks LoRA for a 70B model at $3.00 per 1M, that is $90. On Together for Llama 3.3 70B supervised at $2.03 per 1M, about $61. On OpenAI for gpt-4.1 at $25.00 per 1M, $750. On your own GPU with Unsloth or Axolotl, the price of the GPU hours. The training bill is almost never what makes a fine-tuning project expensive. The dataset labour and the inference-time arrangement are.
What happens at inference time after you fine-tune?
This is where the real money is, and the platforms differ enormously. Fireworks states that fine-tuned models are served at base-model prices, so there is no premium at all. Together serves them serverlessly or on dedicated endpoints priced per GPU-hour, with dedicated HGX H100 listed at $3.99 promotionally through 09/30/26. AWS Bedrock requires Provisioned Throughput for custom models, billed hourly whether traffic arrives or not, and publishes Llama 2 13B at $23.50 per model unit hour with no commitment, which is roughly $17,000 a month. If you fine-tune and self-host, you pay for GPUs continuously. Model the monthly serving cost before the training cost, because it is one to two orders of magnitude larger.
What lock-in does each path create?
Three tiers. Open-weight platforms such as Together and Fireworks, and open-source libraries such as Unsloth and Axolotl, create little: the base model is open, the artefact is a checkpoint or a LoRA adapter, and moving is engineering work. Cloud platform tuning on AWS Bedrock or Google Vertex AI creates high lock-in: the customised model is an object inside that cloud and leaving means retraining. Proprietary first-party fine-tuning creates total lock-in, and OpenAI's wind-down is the proof. Customers who fine-tuned a proprietary model have no weights and no migration path, only a deadline. Whatever you choose, keep the training dataset and the evaluation set under your own control, because that is the only asset that survives a vendor decision.
Can you fine-tune Claude or GPT models in 2026?
Claude, only through Amazon Bedrock, and only Claude 3 Haiku supervised fine-tuning, generally available in US West (Oregon) since November 2024. Anthropic does not expose fine-tuning through its own API. GPT models, only if you are already an active OpenAI fine-tuning customer. New organisations were blocked on 7 May 2026, organisations without recent fine-tuned inference were blocked on 2 July 2026, and all new job creation ends on 6 January 2027. For any new project in 2026, the practical answer is that frontier proprietary models are not a fine-tuning target, and open-weight models are.
Who should not buy a fine-tuning platform at all?
Four groups. Teams without at least a few thousand consistent, correctly labelled examples, because every platform here is machinery for a dataset you do not have and collecting it is the actual project. Teams whose failure mode is missing knowledge, who need retrieval instead. Teams that have not first re-tested their prompt against the current generation of base models, since that test is free and frequently passes. And teams without the inference volume to justify a dedicated serving arrangement. A fine-tuned model needing $9,500 a month of provisioned capacity to answer 2,000 requests a day is a worse deal than paying per token for a frontier model.
LoRA or full fine-tuning?
Start with LoRA and almost always stay there. LoRA trains a small adapter on top of frozen base weights, which is cheaper to train, cheaper to store, faster to iterate on, and servable alongside other adapters on shared capacity. Fireworks prices full-parameter training at exactly twice its LoRA rate, and Together offers both under one price table. Full fine-tuning earns its cost when you are changing the model's behaviour broadly rather than adding a capability, which is rarer than it sounds. If your first LoRA run does not move your evaluation scores, the problem is almost always the dataset, not the method.

About the author

is the founder and creator of LoginRadius, a customer identity platform he built and scaled to over a billion users. He is now the founder of GrackerAI, a GEO platform for B2B SaaS and cybersecurity teams, and has spent more than 15 years building identity and security products.

Related Comparisons

GPU Cloud and AI Compute

Top 7 GPU Cloud and AI Compute Providers 2026: Price Per GPU-Hour by Chip Class, Verified From Each Provider's Own Pricing Page

7 tools compared

LLM Evaluation and Prompt Management

Top 7 LLM Evaluation and Prompt Management Platforms 2026: How to Stop Shipping Prompt Changes Blind

7 tools compared

AI Legal / Contract

Top 5 AI Legal and Contract Tools 2026: Harvey vs Spellbook vs Ironclad vs LegalOn vs Luminance

5 tools compared

AI Sales / SDR

Top 5 AI Sales / SDR Tools in 2026

5 tools compared