Caching in Microsoft Foundry serverless deployments

Question

Hello. I would like to ask if Microsoft Foundry models like DeepSeek V4 Pro and MiniMax 2.7via Fireworks have prompt caching discounts. I want to personally use my balance and use Foundry as an endpoint for prototyping with AI-driven development with an agentic harness (Pi Coding Agent). Are these models available for prompt caching discounts? Or is everything going to be a cache miss?

Answer 1

Hello @Justin Encabo

Thank you for Reaching out to Microsoft Q&A,

Based on the current Azure AI Foundry serverless deployment behavior, models such as:

• DeepSeek V4 Pro

• MiniMax 2.7 via Fireworks

are currently billed on a standard pay-per-use/per-token basis, and there is no broadly documented or guaranteed built-in prompt caching discount mechanism exposed through Azure AI Foundry serverless endpoints today.

In practical terms, you should generally assume:

• Every request is processed independently

• Repeated prompts/shared prompt prefixes are effectively treated as cache misses from a billing perspective

• There is currently no customer-visible cache hit/miss telemetry or cached-token discount model exposed through the Foundry serverless abstraction layer for these provider integrations

While some underlying model providers may internally use KV-cache optimizations for runtime efficiency, Azure AI Foundry serverless deployments do not currently expose deterministic prompt-cache billing discounts similar to some native provider APIs.

Because these models are served through third-party provider integrations (for example via Fireworks), caching semantics can vary depending on:

• The underlying provider implementation

• Whether internal KV-cache reuse exists

• Whether cache reuse is request-scoped or session-scoped

• Whether Azure Foundry passes through any provider-native caching capabilities

• Whether billing systems expose cached-token pricing separately

For your scenario using Foundry endpoints for AI-driven development and agentic workflows such as Pi Coding Agent this is especially important because repeated long-context prompts can significantly increase token consumption and cost.

At this time, Azure AI Foundry serverless endpoints should generally be treated as:

“Full token billing per invocation unless explicit caching support is documented for that model/provider.”

If you would like to reduce or optimize costs, here are a few approaches you can consider:

1. Implement your own application-side cache layer
   • If your agent frequently submits identical or near-identical prompts, you can cache responses in your own database or application layer
   • This is often the most effective workaround today for agentic workflows with repeated context reuse
2. Consider provisioned/reserved capacity options
   • If your workload volume becomes more predictable, provisioned or reserved-capacity deployments may provide better cost efficiency compared to pure serverless pay-per-token usage
3. Explore managed/real-time endpoints
   • Managed compute deployments give you more control over runtime behavior and allow you to implement your own warm-worker or caching strategies within the application/service layer
4. Minimize repeated static context
   • For agentic harnesses, reducing repeated system prompts or shared context can significantly reduce token spend when no caching discounts exist
5. Compare with native provider APIs
   • Some providers may expose prompt caching or cached-token billing more explicitly through their direct APIs than through the Azure Foundry abstraction layer

At the moment, there is no publicly documented indication that DeepSeek V4 Pro in Foundry or MiniMax 2.7 via Fireworks

Currently support customer-visible prompt caching discounts through Azure AI Foundry serverless deployments.

Please refer this

Deploy Models via Serverless API: https://learn.microsoft.com/azure/ai-foundry/how-to/deploy-models-serverless

Microsoft Foundry Models overview (serverless deployments): https://learn.microsoft.com/azure/foundry/concepts/foundry-models-overview#serverless-deployments

Deploy Models via Managed Compute (real-time endpoint): https://learn.microsoft.com/azure/ai-foundry/how-to/deploy-models-managed?tabs=azure-studio#deploy-open-models

Thank you!

Answer 2

Hello Justin Encabo,

Greetings! Thanks for raising this question in Q&A forum.

Great question, and the good news is that the answer is much better than you might expect! Yes, DeepSeek V4 Pro and MiniMax models via Fireworks on Microsoft Foundry do have prompt caching pricing — with a dedicated cached token rate that is significantly lower than the regular input token rate. Let me break this down clearly for you.

DeepSeek V4 Pro — Prompt Caching on Foundry

DeepSeek V4 Pro on Microsoft Foundry via Fireworks has the following serverless pricing: $1.75 per 1M input tokens, $0.15 per 1M cached tokens, and $3.48 per 1M output tokens.

So cached tokens are billed at roughly 91% less than regular input tokens — a very significant discount that makes agentic use cases with repeated context much more cost-efficient.

MiniMax / Kimi K2.6 — Prompt Caching on Foundry

Kimi K2.6 (MoonshotAI) has serverless pricing of $0.95 per 1M input tokens, $0.16 per 1M cached tokens, and $4.00 per 1M output tokens. Both models are available per-token serverless and via PTU through the Foundry model catalog with a single Azure endpoint and the same enterprise controls.

How does the caching work on Fireworks models?

For Fireworks models, cached input tokens are by default priced at 50% for all text and vision language models unless otherwise specified. However, as you can see from the Foundry-specific pricing above, DeepSeek V4 Pro and Kimi K2.6 on Foundry have even deeper cached token discounts (around 91%) compared to the standard 50% discount.

Practical tips for your Pi Coding Agent agentic harness

To maximize cache hits and minimize costs in your agentic workflow, here are the key things to keep in mind:

Step 1: Structure your prompts for cache hits Keep the beginning of your prompts — especially the system prompt, static context, tools definitions, and code context — consistent and identical across calls. The caching mechanism rewards prompts where the prefix is unchanged between requests.

Step 2: Use serverless pay-per-token for prototyping Serverless pay-per-token inference is ideal for experimenting securely and quickly with Data Zone Standard — this is exactly the right option for prototyping with your balance without committing to PTUs.

Step 3: Switch to PTUs when your usage patterns stabilize Once you've validated your agentic harness and have predictable throughput, provisioned throughput units (PTUs) offer predictable, steady-state performance for base or custom models which is better for production agentic workflows.

Step 4: Monitor your cached vs non-cached token usage In the Foundry portal, go to your project > Monitoring to track token usage. Look at the ratio of cached to total input tokens — a good agentic setup with stable system prompts should achieve high cache hit rates over time, keeping your costs very close to the cached token rate.

Step 5: Check for the latest pricing before committing Pricing for these models can change. Always verify the current rates at the Microsoft Foundry model catalog page for DeepSeek V4 Pro and MiniMax before planning your budget. The pricing shown above is current as of May 2026.

If this answer helps you kindly accept the answer which will help others who have similar questions.

Best Regards,

Jerald Felix.

Answer 3

Prompt caching support and discounts are determined by whether the underlying model and deployment type support Azure OpenAI-style prompt caching, not just by being a Microsoft Foundry model.

From the available information:

• Prompt caching is described and priced specifically for Azure OpenAI models (for example GPT‑4o, GPT‑4.1, GPT‑5.x series) under Foundry/OpenAI endpoints. Cached tokens are billed at a discount for Standard deployments and up to 100% discount for Provisioned deployments, with behavior controlled via parameters like prompt_cache_retention and (optionally) prompt_cache_key.
• The prompt caching documentation lists only GPT‑4.x and GPT‑5.x families as supporting extended prompt cache retention, and states that all Azure OpenAI models GPT‑4o or newer support in‑memory prompt cache retention.
• Fireworks-based models (such as DeepSeek and MiniMax variants) are listed separately in the Foundry catalog as partner models, with per‑token or PTU offers, but there is no mention in the provided material that they participate in Azure OpenAI prompt caching or receive prompt caching discounts.

Based on the context provided, prompt caching discounts are documented only for Azure OpenAI models (GPT‑4o and newer, GPT‑4.1, GPT‑5.x families) and not for Fireworks partner models like DeepSeek or MiniMax. For those partner models, behavior should be assumed to be regular per‑token billing with no documented prompt caching discount, and any reuse of computation via caching is not described.

To get prompt caching discounts for an agentic development harness, use one of the Azure OpenAI models that explicitly support prompt caching (for example GPT‑4o+ or GPT‑4.1 / GPT‑5.x families) on Standard or Provisioned deployments and structure prompts so that the first 1,024 tokens are identical across calls.

---

References:

• Prompt caching
• Fireworks models on Microsoft Foundry (preview)
• Microsoft Foundry Models overview
• Explore Microsoft Foundry Models in Azure Machine Learning
• Foundry Models frequently asked questions