Should Institutions Still Bet on One Model?

Opening perspective

The Case for One Frontier Model Is Strong

A rational institution can make a strong case for standardizing on one leading closed frontier model. The best systems offer exceptional reasoning, mature tooling, managed infrastructure, rapidly improving agentic capability, and a development pace that few enterprises could reproduce internally. Stanford’s 2026 AI Index still showed the leading closed model ahead of the leading open model on the Arena Leaderboard, with six of the top ten models closed.¹

But June 2026 exposed a different kind of risk. After a U.S. government directive restricting foreign-national access to Anthropic’s most advanced models, Anthropic disabled Fable 5 and Mythos 5 for all customers to ensure compliance. The restrictions were later lifted and the models were redeployed.⁶ The models had not suddenly become less capable. Access had become a policy decision.

That distinction matters because institutions are simultaneously being offered more credible alternatives. In September, Vercel reported that open-weight models processed 56% of all tokens routed through its AI Gateway in August, up from 7% in December 2025.² Days later, the Financial Times reported growing corporate adoption of open-weight models, including activity at companies such as AT&T, Siemens, and PNC Financial Services.³

The question is therefore no longer simply which company has the best model. It is whether a critical institutional workflow should depend on one model, one provider, and one access path.

Model access risk

What If Access Is No Longer Entirely Your Decision?

The Anthropic episode is useful precisely because it was temporary. It demonstrates that model availability can change even when the underlying technology has not failed. Reuters reported that Anthropic disabled Fable 5 and Mythos 5 globally after a U.S. export-control directive, then later restored the models after the restrictions were lifted.⁶

The policy issue did not remain confined to the provider. At the June G7 summit, leaders discussed a trusted-partner framework for access to advanced U.S. models.⁷ In July, a U.K. government AI adviser described British banks’ lack of access to Anthropic’s Mythos model as a wake-up call for domestic AI capability.⁸ By October, Anthropic itself warned in its IPO prospectus that government actions could cause business disruption and material revenue losses.⁹

Provider policies already make geography explicit. OpenAI states that API access from unsupported countries or territories may result in blocking or suspension.¹⁰ Anthropic reserves the right not to provide commercial access to entities whose majority ownership is attributable to unsupported nations.¹¹

This is not an argument that governments will broadly prohibit frontier-model use. It is a more limited point: access is no longer governed only by technical merit and commercial demand. National-security policy, provider policy, jurisdiction, and regulation can become part of the model lifecycle.

Open-weight momentum

Production Usage Is Moving Faster Than the Narrative

The open-weight market has changed rapidly enough that enterprise architecture assumptions formed even a year ago deserve re-examination. Vercel’s September 2026 AI Gateway Production Index reported that open-weight models rose from 7% of token volume in December 2025 to 56% in August 2026, the first month in which they processed a majority of gateway tokens.²

Those figures should not be mistaken for global market share. Vercel’s data reflects production traffic routed through its own gateway, and the company notes that its open-weight classification is broader than in earlier reports. But the dataset is large, production-oriented, and directionally important: customers are moving meaningful workloads onto models whose weights can be downloaded and deployed through multiple infrastructure paths.²

The Financial Times reached a similar conclusion from a different angle, reporting that companies were increasingly exploring open-weight systems to reduce costs, improve customization, and keep sensitive data under tighter control. Its reporting cited activity at AT&T, Siemens, and PNC Financial Services.³

Open-weight is not the same as open source

Available weights do not necessarily mean that training data, training code, or the full development recipe is open. Licensing also varies. The relevant institutional question is not whether a model carries an “open” label, but what rights, operational control, inspection, customization, and deployment choices the institution actually receives.

The capability gap

Open Weight No Longer Means Second-Rate

Any favorable argument for open-weight AI should begin by acknowledging the strongest counterpoint: the best closed systems still lead several public measures. Stanford reported a 3.3% gap between the top closed and top open model on Arena as of March 2026, and six of the top ten models were closed.¹

But Stanford also found that leading performance had become increasingly concentrated across several developers, shifting competitive pressure toward cost, reliability, and domain-specific performance.¹ That matters because institutions do not consume leaderboard rank. They consume performance on specific workloads.

OpenAI’s gpt-oss-120b is illustrative. OpenAI describes it as a 117B-parameter mixture-of-experts model with 5.1B active parameters, near-parity with o4-mini on core reasoning benchmarks, and efficient operation on a single 80 GB GPU. It is released under Apache 2.0 and is fine-tunable.⁴

NVIDIA’s Nemotron 3 Super offers another signal: 120B total parameters, 12B active parameters, up to one million tokens of context, and an architecture optimized for reasoning and efficient inference. NVIDIA reports higher or comparable accuracy to several peer open models across a range of benchmarks.⁵

The relevant question is therefore not whether an open-weight model is universally superior. It is whether a model clears the institution’s performance threshold for the work while delivering forms of control, portability, or economics that matter to the institution.

From selection to portfolio construction

Asset Managers Already Know the Problem With Concentration

Investment professionals would not normally allocate an entire portfolio to one asset simply because it had the highest recent return. Expected performance matters, but so do concentration, correlation, liquidity, drawdown, and the ability to rebalance. AI architecture is developing an analogous problem.

The analogy should not be pushed too far. Models are not securities, and adding models can increase latency, cost, operational burden, and governance complexity. But the underlying risk principle is familiar: concentration can create fragility even when the concentrated asset is excellent.

Diversification does not mean using every model

A multi-model strategy should not become a model zoo. Institutions need a small, governed set of models that have demonstrated value for specific workflows. A closed frontier model may belong in that set. An open-weight model may belong in it. The objective is not ideological purity; it is model optionality supported by evidence.

Once more than one model can credibly perform the work, the architectural question changes from “Which model wins?” to “Which models should be qualified, how should they be used, and how easily can they be replaced?”

The Model Council

The Multi-Model Pattern Is Already Established

The idea of combining multiple language models is not new. Public research and commercial implementations have explored ranking, peer evaluation, aggregation, and synthesis for years. The architectures differ materially, but the underlying premise is established: multiple models can contribute to the same question and their outputs can be evaluated collectively.

This history matters for a practical reason. The Model Council should not be treated as a magic product feature. It is a design pattern. Differentiation begins with what an institution does around the pattern: which models are admitted, which evidence they receive, how their outputs are evaluated, how disagreements are interpreted, how permissions are enforced, and how the system is re-tested over time.

That is especially important as open-weight models multiply. More choice increases the opportunity for model independence, but it also increases the burden of qualification.

Council governance

A Council Does Not Eliminate Model Risk

Adding models does not automatically make a system safer or more accurate. A Council can reproduce the same error across several models, amplify a bad source, increase cost and latency, or create false confidence because several systems agree. Correlated failure is still failure.

The evaluator is a model too

The Language Model Council research explicitly starts from a weakness in single-model judging: a single LLM judge can introduce its own bias, particularly on subjective tasks.¹³ That concern does not disappear when an institution adds a separate evaluator. The evaluator itself has preferences, failure modes, configuration sensitivity, and potential drift.

This is why a multi-model architecture without assurance can be worse than a well-understood single-model workflow. The Council needs evidence not only that its members are capable, but that the combined system contributes something worth the additional complexity.

From model selection to Model Assurance

Who Has Earned a Seat?

As model choice expands, selection has to become a qualification process. A public leaderboard can tell an institution which models deserve investigation. It cannot determine which models should influence an investment-research workflow, interpret the institution’s evidence, or operate inside its controls.

A serious assurance process should preserve enough evidence to answer a narrower set of questions: what exactly was tested, under which configuration, against which institutional questions and evidence, how the model failed, whether the failure was consequential, and what changed since the last evaluation.

No model gets tenure

A model should remain in a Council because current evidence supports its contribution, not because it won a benchmark six months ago. Models improve. Models regress. Providers change them. New open-weight alternatives appear. Council members become redundant. Evaluators drift.

That is the practical connection between open-weight progress and the Model Council. Open-weight models expand the candidate pool. A Council creates a structure for model diversity. Model Assurance determines which models have earned the right to participate and whether they continue to justify their place.

Closing perspective

The Best Model May Still Be One You Use. It Should Not Be One You Cannot Replace.

The case for frontier models remains strong. For some workflows, the strongest closed system may deliver a capability advantage that easily justifies its cost and dependency. A sensible multi-model architecture should be able to use that model rather than exclude it.

But capability leadership and architectural concentration are different questions. Recent events have shown that access can be affected by government policy, jurisdiction, provider rules, and model lifecycle decisions. At the same time, open-weight models have become capable and deployable enough to provide credible alternatives for a growing class of workloads.

That combination changes the enterprise decision. Institutions no longer have to ask only which model is best today. They can ask which models are qualified, how much dependency they are willing to accept, and whether the architecture can absorb the loss or replacement of a model without losing the institution’s accumulated intelligence.

For asset managers and asset owners, that may be the more durable way to think about AI architecture: not as a permanent wager on one provider, but as a governed portfolio of capabilities whose membership can change as evidence changes.

Evidence used in this paper

1. Stanford Institute for Human-Centered Artificial Intelligence. “Technical Performance,” 2026 AI Index Report, 2026. hai.stanford.edu/ai-index/2026-ai-index-report/techn…
2. Vercel. “AI Gateway Production Index — September 2026,” Sept. 17, 2026. vercel.com/blog/ai-gateway-production-index-septembe…
3. Financial Times. “Corporate America embraces cheaper ‘open’ AI models,” Sept. 2026. www.ft.com/content/d9de4776-1fc9-4f2b-aaaf-9961c35d8…
4. OpenAI. “Introducing gpt-oss,” Aug. 5, 2025; gpt-oss-120b model documentation. openai.com/index/introducing-gpt-oss/…
5. NVIDIA Research. “NVIDIA Nemotron 3 Super,” March 10, 2026. research.nvidia.com/labs/nemotron/Nemotron-3-Super/…
6. Reuters. “Anthropic disables top-tier AI models after US order limiting foreign access,” June 13, 2026. www.reuters.com/technology/us-blocks-foreign-access-…
7. Reuters. “G7 leaders discuss ‘trusted partners’ access to cutting-edge US AI models,” June 16, 2026. www.reuters.com/legal/government/g7-leaders-discuss-…
8. Reuters. “UK banks’ lack of Mythos access a wake-up call, says government AI adviser,” July 14, 2026. www.reuters.com/business/finance/uk-banks-lack-mytho…
9. Reuters. “Anthropic warns government attitudes may hurt customer ties, IPO prospectus shows,” Oct. 2, 2026. www.reuters.com/business/media-telecom/anthropic-war…
10. OpenAI. “API — Supported Countries and Territories,” accessed Oct. 2026. help.openai.com/en/articles/5347006-which-countries-…
11. Anthropic. “Supported Regions Policy,” accessed Oct. 2026. www.anthropic.com/supported-countries…
12. Jiang, Ren & Lin / ACL. “LLM-Blender: Ensembling Large Language Models with Pairwise Ranking and Generative Fusion,” ACL 2023. aclanthology.org/2023.acl-long.792/…
13. Zhao et al. “Language Model Council: Democratically Benchmarking Foundation Models on Highly Subjective Tasks,” June 2024. arxiv.org/abs/2406.08598…
14. Wang et al. “Mixture-of-Agents Enhances Large Language Model Capabilities,” June 2024. arxiv.org/abs/2406.04692…
15. Andrej Karpathy. “LLM Council,” public GitHub repository, Nov. 2025. github.com/karpathy/llm-council…
16. Perplexity. “Introducing Model Council,” Feb. 5, 2026. www.perplexity.ai/changelog/what-we-shipped---februa…