🔍 Read the full analysis: Running AI In-House Could Be A Volume Play, IBM’s CEO Says on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
IBM CEO Arvind Krishna said on the company’s Oct. 21 earnings call that open-weight AI models can be about 90% as effective as leading proprietary models at 10% to 25% of their token cost. An Artificial Analysis comparison available Oct. 26 put the best downloadable model at 80% of the top proprietary score; reaching 96% depends on a model whose weights had not yet been released.
IBM CEO Arvind Krishna said on the company’s Oct. 21 earnings call that open-weight AI models can cost 10% to 25% as much per token as frontier models while being “90% as effective,” making them a potential fit for some enterprise workloads. An Artificial Analysis comparison read on Oct. 26 put the leading downloadable model at 80% of the top proprietary model’s score, underscoring that the performance and savings depend on which models and deployment costs a company compares.
Krishna made the remarks during the call’s question-and-answer session. He said performance differences between leading and open models had narrowed, describing a broad benchmark gap that he said moved from roughly 70 versus 20 or 30 in late 2024 to 70 versus 50 in late 2025. He framed the business case around routine tasks such as summarizing email, examining contracts and answering customer calls, which may not require the most capable available model.
On the Artificial Analysis Intelligence Index, Xiaomi’s MiMo-V2.6-Pro scored 46.3, compared with 57.6 for proprietary Claude Opus 5.5. That makes the downloadable model’s score about 80% of the proprietary leader’s, not 90%. The comparison uses the index’s current version and its set of evaluations; it is one measure of model performance, not a guarantee that models will perform identically on a particular company’s work.
A potential exception was Kimi K3.1 Pro, which scored 55.1, or about 96% of Claude Opus 5.5’s score. But the source material said Moonshot had promised to release its weights by Nov. 4, 2026. Until those weights are available, the model does not qualify as downloadable open weights for this comparison. Its reported score does not establish that enterprises can already run it themselves.
IBM's 90% claim, checked
Open weights reach 80% of the leader
Krishna says open-weight models are "90% as effective". On an independent index the best model you can download today reaches 80% of the best closed one, up from about half two years ago.
Best open model as a share of the best closed
Artificial Analysis Intelligence Index, current version, best downloadable vs. best proprietary model at each date.
What Krishna said
Illustrations from the earnings call, not a named benchmark.
- Late 2024"70" vs. "20 or 30"Krishna, Q3 call, 21 Oct · no named benchmark
- Late 2025"70 to 50"Krishna, Q3 call, 21 Oct · no named benchmark
- Now"90% as effective"Krishna, Q3 call, 21 Oct · no named benchmark
Top downloadable models today
Intelligence Index score and licence type. Read the licence before the leaderboard.
| Model | Index | Licence |
|---|---|---|
| MiMo-V2.6-Pro Xiaomi | 46.3 | MIT |
| DeepSeek V4.1 Flash 1017 DeepSeek | 45.1 | MIT |
| GLM-5.3 Z AI | 44.8 | own commercial licence |
| Step 5 Preview StepFun | 43.7 | own commercial licence |
| Muse Spark Open Meta | 38.9 | licence with conditions |
In points the gap did not shrink: 7.5 points at the end of 2024, 11.3 today. Closed models moved too.
Sources: Artificial Analysis, Intelligence Index v4.3.2 (read 26 Oct 2026), IBM Q3 2026 earnings call transcript (21 Oct 2026), Artificial Analysis on Kimi K3.1 Pro (23 Oct 2026).
Powered by Thorsten Meyer AI
The Economics of Running Models
Krishna’s argument is about more than model prices. Running a model on a company’s own infrastructure can give it more control over where data and logs are kept, a consideration he cited for banks and trading firms. Open weights can also give organizations the option to deploy on-premises or in private clouds, subject to the model’s licence. That control may matter to companies with strict data-handling needs, but it comes with responsibility for hardware, staffing, networking and operations.
IBM sells on-premises systems and has a commercial interest in this approach; Krishna said he expects demand to support sales of the company’s Spyre cards. He also acknowledged a constraint: enterprises need people who can manage model kernels and the underlying network. The source account said this expertise gap remains for many organizations. The proposed savings, then, are not simply a matter of comparing a hosted token price with a model running on purchased equipment.
The purchase-versus-rental calculation in the source illustrates how utilization changes the equation. It lists an eight-GPU H200 server at $172,643 and AWS rental of the same GPU class at $63.30 an hour. On those listed prices, the server’s purchase cost would equal about 114 days of rental at full utilization, or about 15 months at 25% utilization. Those estimates exclude staff, cooling and networking, and do not establish a universal break-even point.
Products Worth Considering
Bose QuietComfort Wireless Noise-Cancelling Over-Ear Headphones, Black
Amazon Kindle Paperwhite (12th Generation, 2024), 16 GB, 7-inch Display, Jade
Logitech MX Keys S Wireless Keyboard, Graphite
As an affiliate, we earn on qualifying purchases.
A Narrowing Gap, Not Parity
The Artificial Analysis figures in the source show the best open-weight model’s share of the leading proprietary score rising from 51% at the end of 2024 to 74% at the end of 2025, then 68% at the end of June 2026 and 80% on Oct. 26. These percentages compare the top-ranked model in each category at each point in time; they do not show that every open model is close to every proprietary model. The source also notes that the absolute score gap was 7.5 points at the end of 2024 and 11.3 points in October 2026, because proprietary models improved too.
Cost comparisons also depend on whether a company hosts the model itself or uses a provider. The source says DeepSeek’s hosted V4.1 Flash 1017 was priced at 3% to 6% of Claude Opus 5.5’s list price, while OpenAI’s closed GPT-6 Luna had a listed price of $0.10 per million input tokens and $0.50 per million output tokens. These are specific listed prices and models, not evidence that every open-weight model is cheaper in every deployment.
Licence terms are another part of the decision. The source identifies MIT licences for MiMo-V2.6-Pro and DeepSeek V4.1 Flash 1017, while other models use custom commercial or community licences. Access to weights alone does not establish that every commercial use is permitted. Companies need to check each model’s terms as well as its benchmark position and operating costs.
Costs and Weights Still Matter
The “90% as effective” figure is Krishna’s characterization, not a result tied in the source material to a named benchmark. Artificial Analysis offers a separate comparison, but its scores do not settle how models perform on particular enterprise tasks or whether the performance difference is acceptable to a customer. The source does not provide a like-for-like accounting of total operating costs across self-hosted and hosted deployments.
It also remains unclear whether Moonshot will release Kimi K3.1 Pro’s weights on the stated schedule and under what licence. The source says no licence had been published and describes the model’s reported 2.8-trillion-parameter size as a substantial infrastructure requirement. The listed hardware calculation likewise omits operating expenses and assumes particular utilization levels. Those factors can materially change the economics for an individual organization.
Watch for Releases and Real-World Costs
The next point to watch is whether Moonshot makes Kimi K3.1 Pro’s weights available by the promised Nov. 4 date and publishes licence terms. If the weights are released, the model could be included in open-weight comparisons, though companies would still need to establish whether its requirements and terms suit their use.
For enterprises considering on-premises AI, the practical test will be workload-specific: model quality on the tasks they need, token volume, hardware utilization, staffing and data requirements. IBM’s earnings-call comments set out the company’s case for that approach; the source material does not provide an independently verified total-cost comparison or a universal break-even point.
Key Questions
What did IBM’s CEO say about open-weight AI models?
Arvind Krishna said open-weight models can cost 10% to 25% as much per token as frontier models while being “90% as effective.” That is his characterization on IBM’s Oct. 21, 2026, earnings call, not a result tied in the source to a named benchmark.
How close are the best downloadable models to the proprietary leaders?
In an Artificial Analysis comparison read on Oct. 26, 2026, MiMo-V2.6-Pro scored 46.3 against Claude Opus 5.5’s 57.6, or about 80% of the leader’s score. Scores can differ by benchmark and do not predict performance on every company’s tasks.
Is Kimi K3.1 Pro available to run on a company’s own servers?
The source material said Moonshot had promised the weights by Nov. 4, 2026, but they had not yet been released as of the account. A model score by itself does not mean its weights are available; its licence and infrastructure requirements also matter.
Does running a model in-house always cost less?
No. The economics depend on utilization and on costs such as staff, cooling, networking and hardware. The source’s H200 purchase-versus-rental calculation excludes several of those expenses, so it is an illustration rather than a general break-even estimate.
Source: ThorstenMeyerAI.com
