AI strategy / Cost model

Enterprise AI Risk Adjusted ROI

Price the chance that you must change models, not only the token price.

Download worksheet

Editable CSV · Open in your spreadsheet app

A token price tells you what a model costs when nothing changes. Enterprise AI programs also pay when something does change. A law, a customer contract, a license, or a vendor decision can force you to replace a model. This guide adds that cost to the comparison. It is for the people who choose models and approve AI budgets.

The interactive calculator runs the same model in your browser. The original Google Sheet has the source values. This guide corrects some dates and legal statuses in the sheet after a check of the primary sources.

Limits of this model. Prices and benchmark scores come from public pages, checked on November 18, 2026, and they change often. The reasoning score is the text-only Humanity's Last Exam, with no tools. The chance of a forced change is Devekko's judgment, not a measurement. Change each input to fit your own exposure. This guide is not legal advice.

See the risk-adjusted ranking

The example workload uses 2,000 million input tokens and 400 million output tokens each month, for 12 months. A forced change of model costs $250,000 in the example. The table ranks the models by cost for each reasoning point.

Model Origin Token cost per year Chance of forced change (percent) Risk cost Risk-adjusted annual cost Reasoning score (HLE) Cost per reasoning point Coding score (SciCode) Cost per coding point
Mistral Large 4 Mistral, France $26,880 5 $12,500 $39,380 35.0 $1,125 54.2 $727
Nemotron 3 Ultra NVIDIA, USA $25,920 5 $12,500 $38,420 28.4 $1,353 40.3 $953
Muse Spark Open Meta, USA $50,400 10 $25,000 $75,400 43.9 $1,718 58.4 $1,291
DeepSeek V4.1 Flash DeepSeek, China $12,960 40 $100,000 $112,960 43.3 $2,609 55.9 $2,021
GLM-5.3 Z.ai, China $54,720 25 $62,500 $117,220 42.3 $2,771 59.0 $1,987
Kimi K3.1 Pro Moonshot, China $96,000 25 $62,500 $158,500 52.5 $3,019 60.8 $2,607
Claude Opus 5.5 Anthropic, USA $192,000 10 $25,000 $217,000 61.4 $3,534 66.9 $3,244
Gemini 4 Argon Google, USA $192,000 10 $25,000 $217,000 57.1 $3,800 61.8 $3,511
GPT-6 Astra OpenAI, USA $480,000 10 $25,000 $505,000 53.1 $9,510 55.4 $9,116

The worksheet has the same rows, so you can enter your own values.

Calculate the cost in four steps

  1. Calculate the token cost: monthly input millions times the input price, plus monthly output millions times the output price. Multiply the sum by the months in view.
  2. Calculate the risk cost: the chance of a forced change times the cost of a migration.
  3. Add the two costs to get the risk-adjusted annual cost.
  4. Divide the risk-adjusted cost by a benchmark score to get the cost for each point.

The model uses list prices for each million tokens:

Model Input price Output price Price note
Mistral Large 4 $0.70 $2.10 EU routing costs $0.77 and $2.31
Nemotron 3 Ultra $0.60 $2.40 One host's price. Other hosts are cheaper
Muse Spark Open $1.25 $4.25
DeepSeek V4.1 Flash $0.30 $1.20 Peak price. Off-peak, which is most hours, is $0.15 and $0.60
GLM-5.3 $1.40 $4.40 Other hosts can be cheaper
Kimi K3.1 Pro $2.00 $10.00 Other hosts can be cheaper
Claude Opus 5.5 $4.00 $20.00
Gemini 4 Argon $4.00 $20.00 Rate from December 1, 2026. Until November 30 the rate is $2 and $10
GPT-6 Astra $10.00 $50.00 A request above 272,000 input tokens costs more

The model does not include prompt caching, batch discounts, or a fee from a model router. Each of these can change the token cost a lot, so add them if you use them.

The risk input changes the ranking

Without the risk cost, DeepSeek V4.1 Flash is the cheapest model per reasoning point, at $299. Mistral Large 4 is second at $768. With a 40 percent chance of a forced change, DeepSeek moves to fourth place at $2,609.

That result depends on a judgment, so test it. DeepSeek ties Mistral per reasoning point at a chance of about 14 percent. Below that value, DeepSeek leads.

The risk cost also stays the same when volume increases, but the token cost does not. At ten times the example volume, DeepSeek leads per reasoning point even with the 40 percent chance. Its risk-adjusted cost is $229,600, and Mistral's is $281,300. At low volume, the risk cost decides the ranking. At high volume, the token price decides it.

The migration cost is the third lever. At $50,000 instead of $250,000, DeepSeek leads again, at $761 per reasoning point against $839 for Mistral. The break-even migration cost is about $60,000.

Know what can force a change of model

A forced change means that you must stop using a model before you planned to. The model adds a probability for each model, but the causes are different:

  • Government rules on origin. Section 1532 of the FY2026 NDAA bars DeepSeek in the performance of US defense contracts. A DHS order, signed July 24, 2026, removes it from federal civilian agencies. Senate bill S. 4784 would add ten more Chinese AI companies, including Zhipu AI and Moonshot AI. It also covers derivative models, so a copy from another host would count. A Senate committee reported the bill, but it is not law. GLM and Kimi are not restricted yet, so the model gives them a lower chance.
  • Your customers' contracts. A federal contractor or a regulated customer can require that you exclude a model. For that business, the chance is not a probability. It is a requirement.
  • Vendor and government action on US labs. In February 2026, the federal government directed agencies to stop using Anthropic. A court enjoined that directive on March 26, 2026. In September, an appeals court upheld a separate defense supply-chain exclusion, and a rehearing petition is pending. A closed model also carries vendor lock-in. The vendor can change the price or the terms, or retire the model.
  • Licenses. Open weights do not always mean an open license. Under the Muse license, a company with more than $20 million revenue that offers the model as a service must agree separate terms with Meta. Mistral Large 4 needs a commercial license for a model-as-a-service business. Read each license before you depend on it.

Open weights reduce some risks and not others. You can host the model yourself, so a vendor cannot remove it. A rule that bars a model by its origin still applies to a copy that you host.

Check quality before cost

Cost per point is a ratio, and it can hide a model that is not good enough. A model that fails your task is not cheap at any price. Set a minimum score for the task first, and compare cost only among the models that pass it.

The two scores measure different work, and leaderboards report slightly different values. One other leaderboard gives GPT-6 Astra about 1.5 points more on each test. HLE (Humanity's Last Exam) measures hard reasoning questions, and SciCode measures scientific code tasks. Mistral Large 4 has the lowest cost per coding point, at $727. Claude Opus 5.5 has the highest scores in both columns. Public benchmarks do not measure your own tasks, so test each candidate on a sample of your real requests.

Reduce the cost of a forced change

The risk cost is the chance times the migration cost. You cannot control the chance, but you can control the migration cost. A lower migration cost reduces the risk cost of every model in the table.

  1. Send each model call through one internal interface, so that a change of model is a configuration change.
  2. Keep an evaluation set of your real tasks, so that you can qualify a replacement in days.
  3. Keep prompts and tool definitions in version control, and test them on a second model.
  4. Qualify one second model for each important workflow before you need it.
  5. Record which customers, contracts, and regions restrict which models.

Use the worksheet

  1. Enter your monthly input and output tokens and the months in view.
  2. Enter the current list prices, and record the date of each price.
  3. Estimate your migration cost from your last change of model or vendor.
  4. Set the chance of a forced change for each model. Use your customer contracts and your markets.
  5. Remove each model that fails your minimum quality score.
  6. Compare the risk-adjusted costs, then repeat the comparison at a higher volume.

The platform-neutral cost guide applies the same method to platform decisions. The AI application stack shows where a model gateway fits. Devekko's Digital strategy service covers model and vendor decisions in a roadmap.

Your next step

Discuss your AI model strategy

Share the task, your current systems, and the decision you need to make. We’ll discuss a useful first scope.

Discuss your AI model strategy

Devekko / chat

AI project guide

Good ideas start with
a conversation.

Tell me what you’re building. I can help you find a starting point and the right Devekko service.

Automatic security check.

Enter to send · Shift + Enter for a new line

AI can make mistakes. Messages are processed by our AI service provider. Leave out private information. Privacy

Prefer a person?hello@devekko.dev