Amazon Bedrock, Microsoft Foundry, Google Vertex AI, OpenAI or Claude: which is best for classification?

Most companies now have at least one AI platform under contract: Amazon Bedrock in their AWS account, Microsoft Foundry beside Microsoft 365, Google's platform in Google Cloud, or an OpenAI or Anthropic API key. And most have a pile of work that is really one repeated decision: which queue a ticket goes to, which account an invoice line is coded to, which group gets access to a new folder, which clause type a contract paragraph is.

This guide compares the five for that job, classification: putting each new item into one of your own categories, the way your team already does. It's written to be useful whichever platform you use. At the end, we show how StayCharted AI Model Trainer works on top of any of them.

Before comparing platforms, it helps to be clear about the method, because the method can change both accuracy and cost substantially.

What the difference looks like. We scored the same held-out rows by asking a strong general assistant (given the category names and a one-line definition of each, method 1), and with models trained on about 10,000 labeled examples (methods 3 and 4):

Where the categories follow general knowledge (medical conditions), the assistant won. Where they follow a team's own conventions (which banking intent, which clause heading), training on examples won clearly. Method 2 adds your past decisions without training. We have not measured it in this comparison; whether it closes the gap needs a separate test. Test it on your own rows rather than trusting anyone's average. Full comparison and method →

So the useful question for each platform is: how well does it support methods 2 and 3, at what cost per thousand items, and where does your data go?

Claude appears in four columns: besides Anthropic's own API, it's offered through Amazon Bedrock, Google's platform and Microsoft Foundry (announced in November 2025). Choosing Claude doesn't force a choice of cloud.

Amazon Bedrock

Strengths. One AWS API for many model families, billed on your AWS bill and usable inside your existing AWS account, roles and network controls. Fine-tuning Amazon Nova is the most practical self-serve path we've found for classification: training is cheap, and a tuned Nova model answers on demand at the same per-token price as the base model, so there's no hourly endpoint to pay for while it's idle (there is a small monthly storage fee per tuned model). The trained model stays in your AWS account.

Watch for. Fine-tuning runs in a few US regions; plan for the training data and the model to live there. The tunable models are Amazon's own Nova family, not Claude: Claude 3 Haiku had a historical fine-tuning offering; availability and retirement dates can differ by hosting platform. Check Bedrock’s current model lifecycle before relying on it.

A data point. In our own test, a Nova Micro tuned in Bedrock on 726 labeled support tickets answered 127 it had never seen at 88% accuracy, using about 130 input tokens and 3 output tokens per ticket. That's a small sample, not a benchmark, but it shows the shape: once a model is tuned, the prompt is short and each answer is cheap.

Microsoft Foundry

Strengths. A broad menu of tunable models: OpenAI's GPT-4.1 family (supervised and preference tuning), o4-mini (reinforcement fine-tuning), and open models such as Llama 3.3 70B, Qwen3 and gpt-oss. It sits beside Microsoft 365 and Entra ID, with regional deployments for data residency, and now offers Claude as well as OpenAI models.

Watch for. Standard and Global Standard deployments of fine-tuned models have hourly hosting fees as well as token charges. Calculate the selected model’s hourly rate × deployed hours, then add usage. Idle hosting can matter for a classifier that fills only a few thousand rows a week. The Developer tier has no hosting fee but is removed after 24 hours, so it's for testing, not production. Provisioned throughput suits steady, large volumes.

Google Vertex AI (now Gemini Enterprise Agent Platform)

Google’s current documentation uses Gemini Enterprise Agent Platform; Vertex AI remains a familiar name in existing integrations and searches.

Strengths. Supervised tuning of Gemini models, with training priced per token. A tuned Gemini 2.x model answers at the same per-token price as its base model, with no hourly hosting fee listed. Gemini 3 and later tuned models cost 1.5 times their base price per token, still without an idle charge. Batch prediction is about half price. Claude is available here too.

Watch for. The tuned endpoint price differs by Gemini generation, so check the pricing page for the model you choose. Like the others, plan where the tuning data and the endpoint live.

OpenAI API

Strengths. Strong general models, structured outputs that keep answers to your category list, and a Batch API at half price. For methods 1 and 2 (categories, or categories plus past examples), it's a good choice.

Watch for. OpenAI is winding down self-serve fine-tuning. Its official guide says new users no longer have access and existing users can create jobs during the wind-down. Check the current deprecation schedule for your model. Models already tuned keep answering until their base model is deprecated. If you plan to fine-tune GPT models from here on, Microsoft Foundry is the place that still offers it.

Anthropic Claude API

Strengths. Claude models are strong at reading long, messy text and following detailed labeling guides. The Message Batches API halves the price for work that can wait. Claude is available through Anthropic directly, Amazon Bedrock, Google's platform and Microsoft Foundry, so a company can standardize on Claude whichever cloud it uses.

Watch for. There's no self-serve fine-tuning for current Claude models. For classification that means methods 1 and 2: give Claude your categories, your labeling guide and good past examples, and measure.

Whichever platform you choose, five things decide whether AI classification works in production.

1. Test on rows it has never seen. Hold back 10–20% of your labeled examples, never show them to the model as examples, and score it on them. Look at accuracy per category, not just the average, and at which categories it confuses.

2. Check it's consistent. A large language model can give a different answer to the same item on a second try. Ask a sample of items twice and count how often the answer changes. A classifier you'll rely on should answer the same thing every time.

3. Get a confidence with every answer. An LLM doesn't give a reliable probability on its own. Workable signals are how often repeated answers agree and whether the most similar past items had the same category. Set a cutoff, and send the answers below it to a person.

4. Work out the cost per 10,000 items. Cost is roughly tokens per item × items × price. Few-shot prompting puts examples in every request, so it can use ten times or more the input tokens of a short prompt to a tuned model. Asking each item three times for consistency triples it. Batch modes halve it. Do this sum before you compare price lists.

5. Know where the data goes. Check each platform’s processing locations, deployment type, retention settings and agreement. A global endpoint does not necessarily keep processing in a region you select. Review the provider’s business-data terms and any opt-in data sharing. With a tuned model, your examples are used once to train, and only new items are sent afterwards.

• You run on AWS: Amazon Bedrock. Use Claude or Nova with past examples, or tune Nova and keep the model in your account.
• You run on Microsoft: Microsoft Foundry. Use GPT or Claude with past examples; tune when the volume justifies the hosting fee.
• You run on Google Cloud: Google's platform. Use Gemini with past examples, or tune Gemini and pay per token.
• You have an OpenAI or Anthropic key and no cloud preference: use the model you trust most with past examples, and measure. Don't start a new fine-tuning project on OpenAI's platform now.
• The categories follow general knowledge (topics, sentiment, broad document types): a strong model with a category list may be enough.
• The categories are your team's conventions (queues, account codes, access groups): show past examples, or train.

In practice, the best platform is usually the one your company has already approved. The bigger risk is skipping the measurement: no test on unseen rows, no confidence, no review.

StayCharted AI Model Trainer turns your team's past decisions into a model that makes the next ones, tests it before anyone uses it, and sends the answers it's unsure about to a person. It works with any of the five platforms above, in two ways.

1. Your AI Model: your company's own AI, given your team's past work. Connect Amazon Bedrock, Microsoft Foundry, Google Cloud, OpenAI, Anthropic or any OpenAI-compatible API in Settings → AI platforms. Then, for every item:

• Your AI is given your categories, your labeling guide and the most similar items your team already categorized (method 2, done for you).
• Each item is asked three times by default, and the answers are compared. The confidence comes from how consistently your AI answered and whether the similar past items agree.
• Before anyone uses it, it's tested on held-out rows, and on a sample asked again for consistency. The report shows accuracy per category, the calls made and the tokens used, so you can estimate the cost on your platform's bill.

Nothing is trained. Amazon Bedrock connects through a role in your AWS account that only AI Model Trainer can use, so there's no key to paste; the other platforms use a key or a service account, stored encrypted.

2. A Dedicated AI Model trained in your own Amazon Bedrock (Business). Tick "Also train models here" on your Bedrock connection, and Bedrock fine-tunes Amazon Nova on your examples in your AWS account (method 3). The model stays in your AWS account and answers there, and your own systems can call it too.

Then compare, and review. On the same rows, you can also train StayCharted's own models: a Classifier Model, an AI Classifier Model, or a Dedicated AI Model on our servers. Every kind is tested on the same held-out rows, so the reports compare like for like, and you publish the one that makes the fewest mistakes that matter. Whichever you publish, answers below your cutoff from files, Google Sheets, the API, Zapier, Power Automate or Claude and ChatGPT come into one Review Queue on Essentials and Business, and what your reviewers decide goes into the next version.

Usage on your AI platform is billed by your provider, on your cloud account. Setting up and testing Your AI Model are included on Essentials and Business; publishing it needs the Your AI Models add-on.

Compare StayCharted models with your own AI → · See the plans · Start free →

Platform documentation checked October 17, 2026. Model availability and prices can change.

• Amazon Nova fine-tuning
• Amazon Bedrock pricing
• Microsoft Foundry fine-tuning
• Microsoft Foundry cost management
• Google platform pricing
• OpenAI supervised fine-tuning
• OpenAI deprecations
• Claude in Microsoft Foundry
• Claude batch processing
• StayCharted benchmark results and method