In September 2026 four vendors offer proprietary, API-served models for enterprise use: OpenAI, Anthropic, Google and xAI. The flagship list prices are identical: GPT-6 Astra and Claude Fable 5.1 both charge $10 per million input tokens and $50 per million output tokens. The mid-range (GPT-5.6 Terra, Claude Sonnet 5, Gemini 3.1 Pro, Grok 4.6) starts at $2 input. The real cost is decided by three other line items: the cache-read price, the long-context price threshold, and whether the model runs in the cloud and region where your data is allowed to be. This article puts the vendors' figures as published on 11 September 2026 side by side, and ends with the reasonable choice for each workload profile.
What is a proprietary model, and how does it differ from open-source?
With a proprietary model the vendor does not release the weights: the model runs on the vendor's or a partner cloud's infrastructure, and your company sends the request and receives the answer through an API. GPT-6 Astra, Claude Fable 5.1, Gemini 3.1 Pro and Grok 4.6 are all in this class. With an open-source model (Llama 4, Qwen 3.5, Mistral 3) the weights can be downloaded, the model can run on your own servers, and the data never leaves your environment, not even for the duration of the call.
That choice is the subject of two separate articles: our open-source vs proprietary comparison works through total cost of ownership, and on-premise vs cloud AI covers where the system runs. This article stays on the proprietary side: once you have decided to call a model over an API, which one should it be.
How much do proprietary models cost on the API?
List prices are per million tokens, quoted separately for input (the prompt and context), cache reads (previously submitted history that is reused) and output (the model's answer). The values below come from the vendors' pricing pages, for standard processing and short context; batch processing at OpenAI and Anthropic is half the list price.
Open full-size chartChart data as a table
| Model | Input | Cache read | Output |
|---|---|---|---|
| GPT-6 Astra | $10 | $1 | $50 |
| GPT-5.6 Sol | $4 | $0.40 | $20 |
| GPT-5.6 Terra | $2 | $0.20 | $12 |
| Claude Fable 5.1 | $10 | $0.25 | $50 |
| Claude Opus 5 | $5 | $0.50 | $25 |
| Claude Sonnet 5 | $2 | $0.20 | $10 |
| Gemini 3.1 Pro | $2 | $0.20 | $12 |
| Gemini 3.8 Flash | $0.75 | $0.075 | $3.75 |
| Grok 4.6 | $2 | $0.50 | $6 |
Per OpenAI's pricing page, GPT-6 Astra is $10 input, $1 cache read and $50 output. GPT-5.6 Sol is $4, $0.40 and $20, on promotional pricing that OpenAI says is available at least through 21 November 2026. GPT-5.6 Terra is $2, $0.20 and $12.
Per Anthropic's pricing page, Claude Fable 5.1 is $10 input, $0.25 cache read and $50 output. Claude Opus 5 is $5, $0.50 and $25. Claude Sonnet 5 is $2, $0.20 and $10; Anthropic announced this as introductory pricing through 31 August 2026 and then made it the standard price instead of the scheduled increase.
Per Google's Gemini API pricing page, Gemini 3.1 Pro (preview) is $2 input, $0.20 cache read and $12 output for prompts up to 200,000 tokens. Gemini 3.8 Flash is $0.75 input, $0.075 cache read and $3.75 output on promotional pricing through 31 December 2026; after that $1.50, $0.15 and $7.50.
Per xAI's models page, Grok 4.6 is $2 input, $0.50 cache read and $6 output for prompts up to 200,000 tokens.
The bare token price misleads on two points. Anthropic states that the tokenizer introduced with Claude 4.7 produces roughly 30% more tokens for the same text than the previous one, so Claude models bill more tokens for the same text than you measured with the old tokenizer. And OpenAI's and Anthropic's per-task cost estimates measure different token consumption, so two models with the same list price cost different amounts per task. The token price is the starting point; the cost measured on your own tasks is the basis for the decision.
Why does the cache-read price decide agentic workloads?
An agent re-reads the system prompt, the tool descriptions and the conversation history at every step. In a longer session most of the submitted tokens are therefore cache reads, and against the list price this is the line item that makes the difference. Fable 5.1's cache read is 2.5% of the input price ($0.25 of $10); on Astra, Sol, Terra, Opus 5, Sonnet 5 and the two Gemini models it is 10%; on Grok 4.6 it is 25% ($0.50 of $2).
Open full-size chartChart data as a table
| Model | Blended price (7:2:1) |
|---|---|
| GPT-6 Astra | $7.70 |
| Claude Fable 5.1 | $7.17 |
| Claude Opus 5 | $3.85 |
| GPT-5.6 Sol | $3.08 |
| GPT-5.6 Terra | $1.74 |
| Gemini 3.1 Pro | $1.74 |
| Claude Sonnet 5 | $1.54 |
| Grok 4.6 | $1.35 |
| Gemini 3.8 Flash | $0.58 |
The chart follows the Artificial Analysis weighting: 70% cache reads, 20% input, 10% output. On that mix GPT-6 Astra comes to $7.70, Claude Fable 5.1 to $7.17, Claude Opus 5 to $3.85, GPT-5.6 Sol to $3.08, GPT-5.6 Terra and Gemini 3.1 Pro to $1.74, Claude Sonnet 5 to $1.54, Grok 4.6 to $1.35 and Gemini 3.8 Flash to $0.58 per million tokens. Between the two flagships the gap is 7% at an identical list price; between the flagships and the mid-range it is four to five times.
Cache writes cost money too: at Anthropic a 5-minute cache write is 1.25 times the input price and a 1-hour write is twice. Your own mix will differ from the chart's 7:2:1 weighting; during the pilot, measure what share of submitted tokens are cache hits and calculate with that.
How large a context window do you get, and where does it get expensive?
The context window is the combined upper limit of input and output in one request. The vendors offer between 500,000 and 1,050,000 tokens, but not all of them sell the large window at one price.
Open full-size chartChart data as a table
| Model | Context window | Price threshold | Rate above the threshold |
|---|---|---|---|
| GPT-6 Astra | 1,050,000 tokens | 272,000 tokens | 2x input, 1.5x output |
| GPT-5.6 Sol | 1,050,000 tokens | 272,000 tokens | 2x input, 1.5x output |
| GPT-5.6 Terra | 1,050,000 tokens | 272,000 tokens | 2x input, 1.5x output |
| Claude Fable 5.1 | 1,000,000 tokens | none | standard rate across the window |
| Claude Opus 5 | 1,000,000 tokens | none | standard rate across the window |
| Claude Sonnet 5 | 1,000,000 tokens | none | standard rate across the window |
| Gemini 3.1 Pro | 1,000,000 tokens | 200,000 tokens | 2x input, 1.5x output |
| Gemini 3.8 Flash | 1,000,000 tokens | none | standard rate across the window |
| Grok 4.6 | 500,000 tokens | 200,000 tokens | 2x input, 2x output |
The model pages for GPT-6 Astra, GPT-5.6 Sol and GPT-5.6 Terra give a 1,050,000-token window and a 128,000-token maximum output; above 272,000 input tokens the whole request is billed at 2x input and cache rates and 1.5x output. Claude Fable 5.1, Opus 5 and Sonnet 5 have a 1,000,000-token window and 128,000-token output per Anthropic's models overview, billed at the standard rate across the full window; Claude Haiku 4.5's window is 200,000 tokens. Gemini 3.1 Pro and Gemini 3.8 Flash have a 1,000,000-token window and 64,000-token output per Google DeepMind's model pages; above 200,000 input tokens the Pro bills $4 input and $18 output, while the pricing page lists one price for the Flash. Grok 4.6's window is 500,000 tokens on the xAI API, with $4 input and $12 output above 200,000; in Microsoft Foundry the same model runs with a 200,000-token window per Microsoft's documentation.
In practice: if some of your requests exceed 272,000 tokens (a full contract file, a large codebase, a long log), on the OpenAI models you pay the higher rate for those requests in full; on the Claude models the same requests run at the standard rate. During the pilot, calculate what share of requests crosses the threshold.
Model selection is a joint decision on workload profile, data path and cloud contract; we shortlist the candidates and run the comparison on your own eval set within a consulting engagement. AI Strategy Consulting
Where does the model run, and where does the data go?
A proprietary model runs where the vendor or the partner cloud runs it. That settles two questions at once: which cloud contract it can be billed under, and which region processes the data.
Open full-size chartChart data as a table
| Vendor | Own API | Microsoft Foundry | Amazon Bedrock | Google Cloud |
|---|---|---|---|---|
| OpenAI (GPT-6 Astra, GPT-5.6) | yes | yes | yes | not listed |
| Anthropic (Claude 5) | yes | yes | yes | yes |
| Google (Gemini 3) | yes | not listed | not listed | yes |
| xAI (Grok 4.6) | yes | yes | not listed | not listed |
OpenAI's GPT-6 Astra announcement lists the model on its own API, on Microsoft Azure and on AWS Bedrock. Anthropic makes Fable 5.1, Opus 5 and Sonnet 5 available on Amazon Bedrock, Google Cloud and Microsoft Foundry alongside the Claude API. Google offers Gemini on the Gemini API and on Google Cloud (Gemini Enterprise Agent Platform, formerly Vertex AI); Google Cloud's model list carries the three Claude 5 models, Llama 4 and the Mistral models next to Gemini 3.8 Flash and 3.1 Pro. xAI offers Grok 4.6 on its own API and, per xAI's announcement, in Microsoft Foundry from 26 August 2026, in preview there.
EU data residency differs by vendor. Per OpenAI's data controls documentation, the Europe region (the eu.api.openai.com endpoint) offers regional storage and regional processing; Modified Abuse Monitoring and Zero Data Retention (ZDR) need OpenAI's approval, and for models released on or after 5 March 2026 the regional endpoint carries a 10% uplift, so it applies to Astra, released on 3 September. Astra's Fast mode is not available with EU data residency. By default OpenAI retains abuse-monitoring logs for up to 30 days, and since 1 March 2023 it trains on API data only with explicit opt-in.
Per Anthropic's data residency documentation, the inference_geo parameter accepts two values, us and global; there is no first-party EU region, and the workspace geo can only be us. An EU region for the Claude models comes from the regional endpoints of Amazon Bedrock and Google Cloud, which bill a 10% premium over the global endpoint. Per the data retention policy, Anthropic does not retain prompts and responses by default, with the exception of Fable 5.1, Mythos 5.1, Fable 5 and Mythos 5: those require 30-day retention, and ZDR is available only with Anthropic's express authorization; for the other models ZDR is requested per organization through sales. It never uses the data for training without express permission.
Per the Gemini API terms, Google does not use prompts and responses from the paid service to improve its products; on the free tier it does. Check the EU-region availability of Gemini models model by model in Google Cloud's per-region model table, because the Flash generations receive regional endpoints at different times.
For xAI, Microsoft's documentation lists Grok 4.6 in Global Standard deployment only; Grok 4.3 and the earlier Grok models are also available in Data Zone Standard (US) deployment, and the table lists no EU data zone for any Grok model.
What ties you to the vendor, and how do you keep the switch open?
A proprietary model ties you to the vendor's roadmap. Anthropic publishes retirement commitments in its models overview: Fable 5.1 no sooner than 1 September 2027; Opus 5 no sooner than 24 July 2027; Sonnet 5 no sooner than 30 June 2027; Haiku 4.5 no sooner than 15 October 2026. OpenAI guarantees Sol's promotional price through 21 November 2026 and Google the 3.8 Flash promotional price through 31 December 2026; after that a different price applies.
The second tie is the platform. The same model runs with a different window in a different cloud (Grok 4.6 with 500,000 tokens on the xAI API, 200,000 in Foundry), and the fast tier is first-party only at several vendors: Anthropic's Fast mode (Opus 5 and Opus 4.8, $10 and $50) is limited to the Claude API, and OpenAI's Astra Fast mode cannot be called with EU data residency.
The third tie is the API. Astra does not support fine-tuning, the Assistants endpoint or the Realtime endpoint; Fable 5.1 brought three API changes relative to Fable 5, which we described in our GPT-6 Astra or Claude Fable 5.1 article.
This is how you keep the switch open:
- One model layer in your own code. The model call happens in one place, and the model ID and parameters come from configuration. The enterprise AI platforms we build are made this way.
- Your own eval set. Fifty to a hundred real tasks from your own data, with expected answers. Every candidate model runs on it with the same prompt and the same cost budget.
- Cost per task. Next to the token price, measure token consumption per task; between the two flagships this is where the real difference lies.
- Cache hit rate and the share of long requests. These two numbers decide how much the cache-read price and the price threshold matter for you.
- Data path in the contract. Write down which region the model runs in, who the data processor is (the vendor or the cloud provider), and how long retention lasts.
Which proprietary model should you choose?
Based on the vendors' published figures, by workload profile:
- Long, complex agentic work where quality decides: GPT-6 Astra or Claude Fable 5.1. At an identical list price the cache read decides: $7.70 versus $7.17 blended. We compared the two models' benchmarks in a separate article.
- Mixed enterprise workloads at a good price-performance ratio: Claude Opus 5 ($3.85 blended) or GPT-5.6 Sol ($3.08, promotional through 21 November).
- High volume, low price: Claude Sonnet 5, GPT-5.6 Terra and Gemini 3.1 Pro between $1.54 and $1.74; Gemini 3.8 Flash at $0.58 on promotional pricing.
- EU data residency from the first party: OpenAI on the eu.api.openai.com endpoint, with a 10% uplift and approval. For Claude, an EU region comes from Amazon Bedrock or Google Cloud. Grok 4.6 in Foundry currently runs in global deployment only.
- Under an existing cloud contract: on Azure OpenAI, Claude and Grok; on AWS OpenAI and Claude; on Google Cloud Gemini and Claude.
No vendor leads on every criterion. First narrow down to two or three candidates by your workload profile and your data path, then measure them on your own eval set with identical prompts; that is where the decision is made.