SAP is replacing some expensive artificial-intelligence models powering its products with cheaper alternatives, a move it expects will generate “more than €2 billion in efficiency gains over the next two years”, Reuters reported on 28 October 2026. Chief Financial Officer Dominik Asam told Reuters the company had removed some of the most expensive AI models from products including its internal consulting assistant.
SAP uses more than 70 AI models from providers including OpenAI, Anthropic and Google, according to Reuters. Asam said the company increasingly chooses the cheapest model capable of performing a task to the required standard, while still expecting overall spending on AI to rise. “The goal is healthy margins as scale grows, not an absolute reduction in AI expenditure,” SAP said in answers to Reuters questions.
Short explainer · SAP AI models brief
A vertical news Short narrated from the published headlines and opening lines of this piece and four related FourWeekMBA pieces on SAP and AI model prices.
Business Pill · Model routing
An evergreen Business Engineer explainer on model routing: sending each request to the cheapest model that can handle it. It explains the general idea only and does not discuss SAP.
The key insight: As we read it, SAP is treating its more than 70 AI models as interchangeable inputs: it picks the cheapest one that meets the standard for each task, while it bills its own customers per completed task.
SAP uses more than 70 AI models from providers including OpenAI, Anthropic and Google and expects more than โฌ2 billion in efficiency gains over the next two years from swapping some expensive models for cheaper ones. Source: Reuters, 28 Oct 2026 (CFO Dominik Asam).
The swap. Asam told Reuters SAP had replaced some of its most expensive models with “lower-cost alternatives where performance remained strong”, in products including its internal consulting assistant.
The target. Reuters reported that SAP expects the move to generate more than €2 billion in efficiency gains over the next two years.
The rule. Asam said SAP increasingly chooses the cheapest model capable of performing a task to the required standard.
The spending. SAP still expects overall spending on AI to rise as adoption grows, Reuters reported.
List prices per million tokens on two published price lists: Mistral Large 4 vs DeepSeek V4.1 Flash on Mistral’s API, and GPT-5.6 Terra vs GPT-6.1 Terra on OpenAI’s. GPT-5.6 Terra prices and the percentages are our arithmetic. SAP did not say which models it swapped. Sources: Mistral AI pricing page (28 Oct 2026); OpenAI (26 Oct 2026).
Why Per-Task Pricing Matters
Reuters reported that, unlike many AI providers, SAP charges customers for completed tasks rather than for the underlying computing resources used to generate responses, meaning lower operating costs can help improve profitability.
SAP’s third-quarter figures, published on 21 October, showed its cloud gross margin at 74.3%, down from 75.1%, as we reported on 22 October. Reuters did not connect the model swap to that quarter’s margin, and SAP did not say how much of the €2 billion would show up in its cloudmargin.
Where Model Origin Comes In
Customers are also paying closer attention to where AI models come from, Reuters reported. Asam said government agencies and companies in regulated industries increasingly ask specific requirements and exclude models based on origin, while European customers increasingly demand local solutions.
How Wide the Price Gaps Are
SAP did not say which models it removed or which replaced them. Two published price lists show how far apart list prices on the same platform can be, per million tokens:
On Mistral’s API, DeepSeek V4.1 Flash is listed at $0.30 for input and $1.20 for output, against $1.36 and $4.18 for Mistral Large 4, as we reported on 28 October.
OpenAI priced GPT-6.1 Terra at $1 for input and $5 for output and said those prices are 50% lower than GPT-5.6 Terra’s, as we covered on 26 October.
By our arithmetic, the cheaper option in these two pairs costs between 22% and 50% of the pricier one per token. These are list prices; SAP’s own contracts and model mix are not public.
Measured per task instead of per token, Artificial Analysis puts the cost of running its index at $10.04 per task for Claude Opus 5.5 (max) and $1.27 for GPT-6.1 Sol (max), as we reported on 28 October.
Connecting the Coverage
Axios reported on 28 October that an AI token price index from Ramp, the corporate expense management firm, put token prices at roughly 59 cents per million tokens, down roughly 35% since the end of June, which we added to our GPT-6.1 Terra piece. What’s new here: a large software buyer is putting a euro figure on what choosing cheaper models is worth to it.
The Structural Read
As we read it, per-task billing is what turns cheaper models into margin for SAP: Reuters reported that SAP charges for completed tasks rather than for computing resources, so lower operating costs can help improve profitability.
SAP still expects its overall AI spending to rise, Reuters reported, so by our reading the €2 billion is about cost per task, not a smaller AI budget.
Reuters reported that customers are also paying closer attention to where AI models come from, and Asam said European customers increasingly demand local solutions, which narrows the pool SAP can route to for some customers.
SAP, in answers to Reuters questions, 28 October 2026
“The goal is healthy margins as scale grows, not an absolute reduction in AI expenditure.”
Three Implications
FOR MODEL VENDORS SAP uses models from providers including OpenAI, Anthropic and Google, per Reuters, and says it increasingly chooses the cheapest one capable of the task.
FOR SAP’S MARGIN SAP charges customers for completed tasks, Reuters reported, so a cheaper model on the same task can help profitability; its Q3 cloud gross margin was 74.3%, down from 75.1%.
FOR BUYERS OF AI SOFTWARE Asam said government agencies and regulated industries increasingly exclude models based on origin, so which models are allowed varies by customer.
The framework treats the token as the unit AI is billed in and says: “The new job of finance is to manage the ratio between the two โ work produced per token burned.”
As we read it, SAP’s rule of picking the cheapest model that meets the standard applies that ratio to its own products. The framework was published in June 2026 and does not discuss SAP.
What Is Not Established
Reuters’ report does not name the models SAP removed or the ones that replaced them, does not break down how the more than €2 billion in efficiency gains is measured, and does not say whether the figure is a cost saving, a margin effect or both. We did not find an SAP press release on the model swap in its newsroom feed on 28 October. We did not contact SAP.
This Week In The AI Supercycle
Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls
This week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.
SAP told Reuters it is replacing some expensive AI models with cheaper ones where performance holds, across a portfolio of more than 70 models from providers including OpenAI, Anthropic and Google, and expects more than €2 billion in efficiency gains over two years, while still expecting its total AI spending to rise.
A note on sourcing. SAP’s statements come from Reuters’ report of 28 October 2026, read in full through a syndicated copy. Price comparisons restate list prices from Mistral’s and OpenAI’s published pages and Artificial Analysis’s leaderboard, as reported in our earlier pieces. Percentages are our arithmetic. We did not contact SAP.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
