Skip to content
llm-spend
LLM API pricing · prompt-cache economics · cost verification

What LLM APIs actually cost.

Real LLM API prices, in USD and CHF. Every number is sourced, and when a rate comes from billing data instead of a pricing page, we say so. Mostly measured on Microsoft Foundry (formerly Azure AI Foundry), plus direct APIs to compare against.

Open the cost calculator →Read the cache case study
97%
Cost cut from caching alone, measured
$12.42→​$1.74
Same task, cache off vs on (CHF 10 → 1.40)
4.5×
Widest cloud resale markup vs direct
7
Providers, all in USD + CHF
Reference

Providers

One page each: the full table in USD and CHF, plus the quirks a pricing page hides, like cache meters, Responses-API-only models, and deployment-tier premiums.

KimiMoonshot AI

Kimi K3.1 Pro and K3 have first-party API lanes; K2.6/K2.7 Code also publish BatchJob rates, with separate Global and Data Zone Foundry rows.

from $0.60 /M in→
DeepSeek

V4.1 Flash now lowers the direct API price, while V4 Pro remains available after DeepSeek reversed its planned September 14 reroute. Microsoft Foundry still exposes the older V4 Flash meters with resale markups from ~10% to a reported 4.5x.

from $0.15 /M in→
GLMZhipu / Z.ai

GLM-5.3 now has Fireworks-hosted Foundry Global and Data Zone lanes, while Z.ai's Flash and faster FlashX variants anchor the direct multimodal lanes.

from $0.15 /M in→
OpenAI / Azure OpenAI

GPT-chat-latest adds a current Direct and Foundry lane at $5/$0.50/$30 per M.

from $0.10 /M in→
OpenAI Decisions APIOpenAI

GPT-6 Luna's general-availability Decisions API discounts cached input.

from $0.10 /M in→
Microsoft Decision ModelsMicrosoft

Microsoft-Decision-1 scores bounded choices at $0.042 per million input tokens.

from $0.042 /M in→
CohereCohere

Rerank 5 direct API rates are $0.05/M tokens for Pro and $0.02/M for Fast.

from $0.02 /M in→
ClaudeAnthropic

Claude Opus 5.5 Fast mode is available on Azure-hosted Foundry, alongside the standard and Direct price lanes.

from $0.10 /M in→
GeminiGoogle

Gemini 4 Argon is Google's new flagship reasoning model, with introductory API rates through November 30.

from $0.75 /M in→
GrokxAI

Grok 4.7 is GA on Microsoft Foundry Global and US Data Zone; Grok 4.6 remains a Global Standard preview.

from $0.20 /M in→
QwenAlibaba

Qwen3.8 Max is the flagship, while Qwen3.8 Flash adds a new $0.15/$0.47 multimodal direct lane — but Foundry still has no per-token Qwen meter.

from $0.10 /M in→
MistralMistral AI

Mistral Large 4 launch pricing and Mistral Medium 3.5 on Foundry.

from $1.36 /M in→
MiniMaxMiniMax AI

MiniMax M2.5 and MiniMax 3 are resold on Microsoft Foundry as Data Zone-only serverless listings; the direct API lineup now includes M3.1 Flash.

from $0.15 /M in→
Perplexity

Perplexity's Agent API hosts NVIDIA Nemotron 3.5 Super and DeepSeek V4 Pro; its Decisions API now has v1.2 and Flash v1.2 input-only lanes.

from $0.01 /M in→
Microsoft AI

MAI-Thinking-1 and MAI-Code-1.1-Flash bring Microsoft's own reasoning and coding models to Microsoft Foundry.

from $0.20 /M in→
Embeddings

The retrieval layer. Input-only pricing, and the cheapest model is rarely the right one for code RAG.

from $0.02 /M in→
The framework

Don't trust the sticker price

Seven things that decide what you actually pay. None of them is the headline number.

01

Blended cost depends on your input:output ratio

Output usually costs 3-4x input. Two models with the same input rate can differ 2x on a real workload, so your token split matters more than the headline price.

02

Prompt caching is the single biggest lever

Often bigger than model choice. A task that costs $12.42 / CHF 10 without cache came in at $1.74 / CHF 1.40 at a 97% hit rate.

03

Microsoft Foundry can publish incomplete pricing

Foundry's public pages often skip the cached-input column for third-party models even though the cache meter exists and gets billed. A billing export exposes it.

04

Resale markups vary wildly

From near 0% (resold at the provider's own rate), to ~10% (Data Zone vs Global), to as much as 4.5x (a 'Global' tier vs the direct API, per public support forums).

05

Context window is a real autonomy lever

Bigger windows (1M tokens on DeepSeek V4 and GLM-5.2) mean less forced context compaction, so less babysitting on long runs.

06

Benchmark leaders don't always win

GPT-5.3-Codex tops coding benchmarks but can lag in real agentic work, with poor context retention, versus DeepSeek V4 Pro and GLM-5.2.

07

Don't trust in-app 'estimated cost' widgets

Portal estimates can show $0 or wildly wrong numbers while real charges pile up. Trust billing exports grouped by meter instead.

Reusable method

How to verify real LLM costs

Six steps for finding what a workload really costs, including how to back out an undocumented cache rate from a billing export.

  1. 1

    Get the provider's published per-million rate for input, output, and cached input if it's listed.

  2. 2

    Run a real workload and read the actual billed cost from Azure Cost Management, not a chat client's estimate.

  3. 3

    If real cost is far below the flat-rate math, suspect an undocumented cache discount before assuming a bug.

  4. 4

    Export a cost report grouped by meter, not by service name. Match known meters to confirmed rates to back out token counts, then solve for the unknown: cached_rate = cached_meter_cost / cached_token_count.

  5. 5

    Cross-check those token counts against the usage dashboards to validate your assumptions.

  6. 6

    Treat anything not on an official page as an estimate, not fact. Re-check once real numbers land or a billing cycle closes.

Confidenceofficial published pagederived reconciled from billing †estimate pattern-inferred ‡