AI model pricing is set in each provider's rate card: a table with one row per model and one column per token type (input, cached input, cache write, output), in US dollars per million tokens, repeated for every processing tier, with per-unit prices for images, audio and tools beside it. The input price that comparisons quote is one cell of that table. What an application costs depends on how its tokens spread across the cells, which tier and region it runs in, and whether the model is bought from its provider or through a cloud.
How it works
Rows and columns. Each row is a model, each column a token type; which types exist and how the providers' cache columns differ is explained in tokens, requests and context. Output costs a multiple of input: on the lists read, five times for gpt-6.1-sol and Claude Sonnet 5.5 and six times for Gemini 3.1 Pro Preview. Tools carry per-unit prices, such as web search at USD 10 per 1,000 calls at OpenAI and Anthropic plus the tokens of the results.
Uplifts and discounts. Three modifiers change the rate of the same cell:
- Context length. OpenAI prices input above 272,000 tokens in separate long-context columns, for gpt-6.1-sol at twice the input and one and a half times the output rate; Google applies the same factors to Gemini 3.1 Pro Preview prompts above 200,000 tokens. Anthropic bills the full one-million-token context window of Claude 4.6 and later models at the standard rate, except Claude Haiku 5.5, which costs five times as much above 100,000 tokens.
- Processing tier. Batch costs half on all three lists. Flex, at batch rates on OpenAI and Google, trades price for slower responses. Preferential processing costs more: OpenAI renamed Priority to Fast mode on 2026-07-30 and charges twice the standard rate for gpt-6.1-sol in Fast mode and six times in Ultrafast mode, Google charges 1.8 times for Priority on Gemini 3.1 Pro Preview, and Anthropic twice for Fast mode on Claude Opus 5.5.
- Region. OpenAI adds 10 percent on data residency endpoints for models released on or after 2026-03-05; Anthropic applies a 1.1 multiplier to US-only processing through
inference_geoon Claude 4.6 and later models. The Gemini API price page states no data residency terms. What residency covers is explained in data residency in the cloud.
At Anthropic the modifiers compound, for example cache multipliers on top of batch prices.
The same model through a cloud. Bought through Amazon Bedrock, Microsoft Foundry or Google's Gemini Enterprise Agent Platform, a model is priced on the cloud's list, with the cloud's regions and currency. On Bedrock, Priority costs 75 percent more and Flex 50 percent less than Standard where the list states the ratio, and Claude models carry separate prices for global and geographic inference; Anthropic gives regional endpoints on Bedrock and Google Cloud a 10 percent premium. Azure OpenAI prices the deployment types Global, Data Zone (EU or US) and Regional separately; how Microsoft converts its dollar prices is covered in EUR and GBP cloud cost reporting. Google charges more on non-global endpoints for generally available models from Gemini 3 on. Which invoice, in which currency, each route produces is covered in AI provider billing, endpoint and training charges in AI and machine learning platform services.
Model IDs and lifecycle. The model ID decides which row applies. At Anthropic every ID from the 4.6 generation on is a fixed snapshot; older short names such as claude-sonnet-4-5 point to the latest dated snapshot. OpenAI pins versions as snapshots, dated for older models such as gpt-4o-2024-08-06, and Google swaps a latest alias with every new release. A new model brings its own rates: Claude Opus 5.5 costs USD 4 input and USD 20 output, Claude Opus 5 USD 5 and USD 25. Anthropic promises at least 60 days' notice before a retirement; for previews, OpenAI and Google name two weeks. For generally available models OpenAI states at least six months, yet on 2026-10-14 it announced the end of GPT-4o and GPT-4.1 for 2027-01-14, three months later.
What it means for your cost
Comparing models needs the token mix of one real workload, not the input price. The usage and cost exports of the AI providers give uncached input, cache reads, cache writes and output, and those quantities turn any rate card into a bill. Which cells a workload hits:
| Workload trait | Cell or modifier | What to check |
|---|---|---|
| long, stable system prompt | cached input, cache write | share of cache hits |
| long answers or heavy reasoning | output | output tokens per request |
| prompts above 100,000, 200,000 or 272,000 tokens | long-context rates | share of requests above the model's threshold |
| no answer needed now | batch, flex | time window until the result |
| tight latency targets | priority, Fast mode | premium per token |
| processing required in the EU | data residency, Data Zone, regional endpoint | premium of the chosen endpoint |
| tools such as web search | per-call price | calls per request |
List prices are the starting point: Anthropic negotiates volume discounts case by case, and how clouds agree private pricing is covered in enterprise agreements and private pricing. Applying cache, batch and routing is covered in caching, batching and routing, what training and dedicated endpoints add in training, fine-tuning and inference.
Where it goes wrong
- Input prices compared alone. The model with the lower input price turns out dearer, because cache rates, tiers and thresholds sit differently at each provider.
- The output multiple overlooked. For answer-heavy features, such as long summaries or agents that reason a lot, output decides the bill at five to six times the input rate.
- A cloud price read as the provider's. The estimate misses the cloud invoice, because the cloud's deployment type, region and currency apply; an Azure Data Zone costs more than Global.
- An alias moves to a new release. Cost changes without a code change, because a
latestalias or short name now points to another model, possibly with another rate or tokenizer. - Cached rates assumed without cache hits. The plan uses the read rate while the usage data shows few cache reads; where writes carry a premium, each request costs more than with no cache.
Example
List prices per million tokens in USD, standard tier, read on 2026-10-16: gpt-6.1-sol on the OpenAI API, Claude Sonnet 5.5 on the Claude API with global processing, Gemini 3.1 Pro Preview on the Gemini API paid tier.
| Rate | gpt-6.1-sol | Claude Sonnet 5.5 | Gemini 3.1 Pro Preview |
|---|---|---|---|
| Input | 2.00 | 2.00 | 2.00 |
| Cached input (read) | 0.10 | 0.10 | 0.20 |
| Cache write | 2.50 | 2.50 (five minutes) | storage 4.50 per hour |
| Output | 10.00 | 10.00 | 12.00 |
For a workload with I million uncached input tokens, C million cache hits and O million output tokens below the thresholds, gpt-6.1-sol and Claude Sonnet 5.5 cost the same, 2.00 × I + 0.10 × C + 10.00 × O USD plus cache writes. Gemini 3.1 Pro Preview costs 2.00 × O + 0.10 × C USD more, with cache storage instead of writes.
Through Azure, the Azure Retail Prices API lists gpt-6.1-sol in Germany West Central (effective 2026-10-01) at the same USD 2.00 input and USD 10.00 output in Global Standard, and at USD 2.40 and USD 12.00 in Data Zone Standard, 20 percent more; the API's euro prices are EUR 1.7599 and EUR 8.7997 for Global.
Sources
Read on 2026-10-16. OpenAI pricing, Flex processing, GPT-6.1 Sol, GPT-4o and its snapshots and deprecations; Anthropic pricing, model IDs and versioning and model deprecations; Google Gemini API pricing, Gemini models, Gemini deprecations and Gemini Enterprise Agent Platform pricing; AWS Amazon Bedrock pricing; Microsoft Azure OpenAI pricing and the Azure Retail Prices API, queried for Germany West Central in USD and in EUR.