Which LLM is best for UK businesses in 2026?

There is no single best LLM; the right choice depends on use case, residency requirements, existing stack, and cost tolerance. For most UK SMEs in April 2026, Claude Sonnet is the common workhorse on quality-to-cost. For Microsoft-heavy stacks, Azure OpenAI is the default. For very long context or multimodal work, Gemini Pro on Agent Platform is competitive. The 12-criterion framework and a two-week bake-off produce a defensible choice faster than vendor benchmarks do.

How do we keep data in the UK when using an enterprise LLM?

Choose a route that explicitly supports the required UK storage or processing boundary; an EU region does not meet a UK-only requirement. Anthropic’s own API offers US or global inference only. UK GDPR regulates restricted transfers rather than requiring UK storage. As at 14 October 2026, AWS lists London (eu-west-2) in-region availability for Claude Sonnet 4.6 on bedrock-runtime, but the same model card gives its in-region endpoint as N/A. Confirm the endpoint and account access with AWS before relying on UK inference. AWS separately lists Claude Haiku 4.5 on bedrock-mantle for in-region inference in Ireland (eu-west-1) and Stockholm (eu-north-1). Storage, retention and transfers need separate assessment. OpenAI offers UK storage at rest on ChatGPT Enterprise and the API, but its inference residency covers only the United States, Europe (EEA and Switzerland) and the United Arab Emirates. On Azure, the current GPT-6 models are available from UK South only as Global Standard deployments, which Microsoft says may process in any Azure region, and the UK is outside Microsoft's EU Data Zone. On Google Cloud, assess the named model and endpoint; an EU processing commitment is not a UK commitment. Mistral hosts data in the EU by default and commits to EU inference only on its EU regional endpoint, at a 10% upcharge. In each case a UK region is not the same thing as UK processing. A signed data processing agreement with a no-training clause from every vendor in the pipeline is required, along with a documented international transfer mechanism if any component sits outside adequate jurisdictions. Verify current residency on the provider's trust documentation at the time of contracting.

How much does an enterprise LLM cost in 2026?

List prices are usually quoted per million input and output tokens and vary widely by model tier. In April 2026, a typical UK business running 20 staff against a Sonnet-class model spends £400 to £1,500 per month on inference; a heavier workload with longer contexts and tool use can reach £3,000 to £8,000 per month. Frontier models such as Opus or the most capable GPT variants are two to five times more expensive per token, which matters for high-volume workloads. A hybrid routing strategy usually reduces total inference cost by 40% to 70%.

How do we actually run a bake-off?

Prepare a held-out task set of 50 to 150 prompts drawn from real work, with agreed correct outputs or acceptability criteria. Score blinded where possible. Measure latency from your infrastructure in your region under realistic load. Calculate cost-per-task at list price. Include operational factors such as onboarding ease, billing fit, and vendor responsiveness. A defensible bake-off finishes in two to three weeks with a written report to procurement stakeholders.

When should we fine-tune versus use the base model?

Fine-tune when you need a specific output format the base model does not produce reliably, a tone or style you cannot prompt into, or lower latency and cost that a smaller fine-tuned model can deliver on your task. Do not fine-tune to improve factual knowledge; retrieval-augmented generation is cheaper, updateable, and more auditable. Plan for the fine-tuning dataset to fall under the same governance as production data, and for periodic re-tuning as the base model family evolves.

Back to Blog

strategy

Enterprise LLM selection for UK businesses in 2026

By The AI ConsultancyPublished 24 April 2026Last reviewed 10 October 2026

A 12-cell scoring grid overlaid on silhouettes of five candidate language models, with a bake-off workflow diagram below

What "enterprise LLM" means in 2026

An enterprise LLM is a large language model offered under commercial terms designed for business use: a signed data processing agreement, a no-training guarantee by default, role-based administrative controls, audit logging, regional data processing options, and either single sign-on integration or an API surface that supports it. The term is useful because it excludes consumer tiers (free ChatGPT, free Claude.ai, free Gemini) which share a model family with the enterprise offers but lack the contractual and administrative surface UK businesses need. A UK business evaluating an LLM in 2026 is choosing between enterprise offers, not between consumer products.

Scope of this review, 14 October 2026: provider controls and model availability have been checked. References to April 2026 procurement experience and the consultancy’s illustrative cost and routing estimates below remain historical judgements, not current vendor quotations or measured savings.

Four providers have substantive enterprise LLM offerings relevant to most UK businesses: Anthropic (Claude Opus and Sonnet, via direct API, AWS Bedrock, Google Vertex), OpenAI (GPT-series via API, Azure OpenAI for Microsoft customers), Google (Gemini Pro via Agent Platform), and Mistral (Mistral Large via their API and Azure). Other credible options exist, particularly for open-weights self-hosting, but these four cover the majority of enterprise procurement conversations we see in the UK in April 2026.

A 12-criterion scoring framework

The framework below is designed to be scored out of 10 per criterion, with weightings set by your own context. It is deliberately criterion-heavy; a three-criterion "price, quality, speed" framework will not surface the questions that actually matter for an enterprise deployment. For a companion view on choosing between the two most common providers specifically, see our Claude vs ChatGPT for UK SMEs comparison.

Data residency. Where does the model run, where is data processed, and what regions are guaranteed in the contract? UK or EU residency matters for UK GDPR posture; a provider that can only process in the US adds an international transfer burden.
Single sign-on and SCIM. Does the provider support enterprise SSO (SAML, OIDC) and automated provisioning (SCIM)? Without these, user lifecycle management is manual and account sprawl becomes an audit finding.
DPA terms. Is the data processing agreement balanced, with clear data controller/processor roles, no-training by default, and clean UK or EU international transfer mechanisms? Are sub-processors disclosed and updated?
Audit logs. What is logged, for how long, and how is it accessible? Per-user prompt and response logs, API call logs, admin changes, and policy violations are the minimum. Logs should be exportable to your SIEM.
Safety evaluations. What published evaluations has the provider run on the model, in which languages and domains, and how have the results moved across model generations?
Regional pricing and billing. Can the provider bill in GBP, issue VAT-compliant invoices, and support UK entity purchasing? Surprisingly, this still trips up enterprise procurement in 2026.
Latency and throughput. What is the typical time-to-first-token and tokens-per-second in the region you will operate in? Cross-region latency to US endpoints can add 300 to 800 ms per request, which matters for agents and voice use cases.
Context window. What context length is supported, how is it priced, and how does the model behave as context fills? Claim and reality diverge here; test under load.
Function calling and tool use. Does the model support structured function calling, parallel tool calls, and JSON mode reliably? For agentic uses, this is the single biggest differentiator between models on paper versus in production.
Fine-tuning options. Does the provider offer fine-tuning, and if so, what is the data handling posture? Some vendors fine-tune in-region with strict isolation; others do not offer it on enterprise plans at all.
Retention controls. Can retention be set to zero for prompts and completions? What is the default? Zero-data-retention eligibility is model-, endpoint- and contract-specific. It is often an approved API arrangement, not a standard enterprise-chat setting. Obtain written confirmation of coverage and exceptions.
Independent benchmark scores. How does the model perform on independent benchmarks relevant to your use case: MMLU, GPQA, HumanEval, SWE-bench, and domain-specific suites? Treat vendor-published benchmarks as one input among several, not as the whole picture.

Comparison table across leading enterprise LLMs

The table below summarises the five model families we most often see shortlisted by UK businesses. The residency column was re-verified against each provider's own documentation on 10 October 2026; the strengths and watch items are our working assessment and are there to be tested in your own bake-off. Refresh this table quarterly: vendor prices, residency options, and capabilities change on a short cadence, and the details below are a snapshot intended as a starting point, not a final selection.

Model	UK / EU residency	Strengths	Watch items
Claude Opus 5.5	AWS bedrock-runtime London availability is disputed within its Opus 5.5 documentation, checked 14 October 2026; obtain confirmation before purchase	Strong reasoning, reliable tool use, careful safety posture, long context	Higher per-token cost; do not treat EU processing as UK processing
Claude Sonnet 5.5	Checked 14 October 2026: Claude Sonnet 5.5 has no European in-region option on AWS bedrock-runtime or bedrock-mantle. Google Cloud Agent Platform lists an EU multi-region endpoint; Anthropic-direct API offers US or global inference. Confirm route, access and contractual coverage	Strong quality-to-cost ratio; most common workhorse choice for UK SME deployments	Lower ceiling than Opus on hardest reasoning tasks
GPT-series (latest, via Azure OpenAI in Microsoft Foundry)	Deployable from Azure UK South, but the current GPT-6 models are offered there as Global Standard only, which Microsoft says may process in any Azure region. UK-geography processing is limited to older models: GPT-4o and GPT-4.1 mini on regional Standard, and GPT-5.6 Sol and Luna on Regional Provisioned. The UK is not in Microsoft's EU Data Zone. Data at rest stays in the resource's Azure geography	Deep ecosystem, mature tooling, strong function calling	Check the deployment type, not only the region; rapid release cadence means behaviour changes frequently, so evaluate on the specific snapshot
Gemini (via Google Cloud's Agent Platform, formerly Vertex AI)	Google publishes processing commitments by model and endpoint. Check the selected model’s row and regional endpoint before procurement; the global endpoint is not a regional processing guarantee	Very long context, native multimodal, competitive pricing	Check the exact model version and endpoint in the current provider tables; test tool-use reliability on the intended workload
Mistral Large (via Mistral API and Azure)	Data is hosted in the EU by default on Mistral's API. Inference location is committed only on the EU regional endpoint (EU and EFTA data centres), at a 10% upcharge and with feature limits; the default global endpoint carries no location commitment. No UK region. On Azure, Mistral Large 3 is in the EU Data Zone, and in UK South and UK West as Global Standard only	European provenance, competitive cost, strong on European languages	Mistral Large 4 is in public preview (6 October 2026); smaller ecosystem of enterprise integrations; less deep governance documentation than the US providers

The residency column was verified on 10 October 2026: the Claude rows against Anthropic's platform documentation, as set out in our Claude data residency guide; the GPT row against Microsoft's Foundry deployment-type and region-availability pages; the Gemini row against Google Cloud's data residency and endpoint pages; and the Mistral row against Mistral's regional inference documentation and Help Center. The pattern across all four providers is the same: a region name in a console does not tell you where inference runs. A UK region can be a place to deploy from without being the place your prompts are processed. Ask each vendor which deployment type or endpoint carries a processing commitment for the specific model you intend to use, and confirm it on the current documentation before committing, because this is a fast-moving area.

How to run a bake-off

A bake-off is a structured comparison between two or three shortlisted models on your actual tasks. It is the step most UK businesses skip, and the decision quality suffers when they do. A defensible bake-off has five components.

A held-out task set. 50 to 150 prompts drawn from real work, with agreed correct outputs or acceptability criteria. The set stays fixed across all providers to allow a like-for-like comparison.
A scoring rubric. Quality dimensions defined in advance: factual correctness, format adherence, tone, handling of edge cases, willingness to say "I don't know." Score blinded where possible: remove provider identifiers from responses before the reviewer scores them.
Latency measurement under realistic conditions. Time to first token and end-to-end completion time, measured from your infrastructure in your region, at a load resembling your expected production traffic.
Cost-per-task calculation. Token counts and list price give a per-task cost; sum across the task set for the bake-off total. Include retrieved context in the input token count if your production flow uses retrieval.
Operational factors. Onboarding time, sandbox availability, billing fit, and the provider's responsiveness during the evaluation. These often outweigh model quality differences for shortlisted candidates that score within a point or two of each other.

A useful bake-off finishes in two to three weeks with a written report shared to the procurement stakeholders. A bake-off that stretches beyond a month typically means the scope was not tight enough or the task set was not defined clearly up front.

When to fine-tune (and when not to)

Fine-tuning an enterprise LLM on proprietary data is right in three situations: when you need a specific output format the base model does not produce reliably, when you need a tone or style the base model cannot prompt into, or when you need lower latency or cost and have demonstrated that a smaller fine-tuned model matches the base model's quality on your task. Fine-tuning is usually wrong for improving factual knowledge; retrieval-augmented generation is cheaper, updateable, and more auditable for that purpose. Our RAG for UK SMEs guide covers the retrieval path in more detail.

Two practical cautions. Fine-tuning data becomes part of the risk surface: the fine-tuning dataset must be handled with the same care as production data, including UK GDPR compliance, access controls, and retention rules. And fine-tuned models often lag the base model's capabilities across the next one or two releases; a model fine-tuned on one Sonnet release does not inherit the next release's improvements. Plan for periodic re-tuning or for a path back to base models if the capability gap grows.

Governance documentation buyers will ask for

Enterprise procurement processes in the UK increasingly ask a consistent set of governance questions before contracting. Assembling the documentation up front shortens the sales cycle materially and avoids the pattern where a chosen model is later blocked by legal or compliance review. Four documents are usually required.

A signed data processing agreement (DPA) with the provider, configured to reflect the controller-processor roles for your intended use. Read it against the provider's standard sub-processor list and the stated data retention defaults; confirm that training on your data is disabled by default and cannot be re-enabled without your consent.

A transfer impact assessment if any component of the pipeline sits outside the UK or adequate jurisdictions. This document names the transfer mechanism (UK International Data Transfer Agreement, or the Addendum to the EU Standard Contractual Clauses), records the transfer risk analysis, and notes any supplementary measures applied.

A model-use policy setting out which staff may use which models for which tasks, what data classifications are permitted for each model, and what audit evidence is kept. The policy should be referenced by staff induction and refreshed when a new model is added to the approved list.

A bake-off record: the held-out task set, the scoring rubric, the results, the operational observations, and the rationale for the final selection. This document is disproportionately valuable at year two when the decision is revisited or when a new model release forces a re-evaluation. Without it, the business typically re-runs the bake-off badly and without the original context.

Hybrid model strategies

Most mature UK enterprise AI deployments in 2026 are not single-model. A typical pattern is a "router" that selects between a higher-capability model for hard tasks and a smaller, cheaper model for routine tasks, with a third specialist model for particular use cases. Hybrid strategies can reduce inference cost by 40% to 70% against a single-model baseline without materially changing user-facing quality, when the routing logic is tuned against your actual task mix.

The trade-off is complexity: two or three models means two or three sets of contract terms, two or three audit surfaces, and more prompt-engineering work to keep behaviour consistent. A hybrid strategy is usually the right choice only after a single-model deployment has proven its value and the cost pressure has become material.

Where to start

A sensible first step is to populate the 12-criterion framework with your organisation's weightings, shortlist two or three candidates, and run a two-week bake-off on a held-out task set. The bake-off report becomes the foundation for procurement approval and for the governance documentation the legal and compliance teams will ask for later. For structured support with shortlisting, bake-off design, and enterprise rollout, see our AI strategy service, our Enterprise AI service, and our AI implementation service. For sector-specific context, our financial services industry page covers the additional procurement and audit expectations in regulated environments. For wider background, the Knowledge Hub strategy section collects related guides on AI strategy and governance.

Frequently asked questions

Which LLM is best for UK businesses in 2026?: There is no single best LLM; the right choice depends on use case, residency requirements, existing stack, and cost tolerance. For most UK SMEs in April 2026, Claude Sonnet is the common workhorse on quality-to-cost. For Microsoft-heavy stacks, Azure OpenAI is the default. For very long context or multimodal work, Gemini Pro on Agent Platform is competitive. The 12-criterion framework and a two-week bake-off produce a defensible choice faster than vendor benchmarks do.
How do we keep data in the UK when using an enterprise LLM?: Choose a route that explicitly supports the required UK storage or processing boundary; an EU region does not meet a UK-only requirement. Anthropic’s own API offers US or global inference only. UK GDPR regulates restricted transfers rather than requiring UK storage. As at 14 October 2026, AWS lists London (eu-west-2) in-region availability for Claude Sonnet 4.6 on bedrock-runtime, but the same model card gives its in-region endpoint as N/A. Confirm the endpoint and account access with AWS before relying on UK inference. AWS separately lists Claude Haiku 4.5 on bedrock-mantle for in-region inference in Ireland (eu-west-1) and Stockholm (eu-north-1). Storage, retention and transfers need separate assessment. OpenAI offers UK storage at rest on ChatGPT Enterprise and the API, but its inference residency covers only the United States, Europe (EEA and Switzerland) and the United Arab Emirates. On Azure, the current GPT-6 models are available from UK South only as Global Standard deployments, which Microsoft says may process in any Azure region, and the UK is outside Microsoft's EU Data Zone. On Google Cloud, assess the named model and endpoint; an EU processing commitment is not a UK commitment. Mistral hosts data in the EU by default and commits to EU inference only on its EU regional endpoint, at a 10% upcharge. In each case a UK region is not the same thing as UK processing. A signed data processing agreement with a no-training clause from every vendor in the pipeline is required, along with a documented international transfer mechanism if any component sits outside adequate jurisdictions. Verify current residency on the provider's trust documentation at the time of contracting.
How much does an enterprise LLM cost in 2026?: List prices are usually quoted per million input and output tokens and vary widely by model tier. In April 2026, a typical UK business running 20 staff against a Sonnet-class model spends £400 to £1,500 per month on inference; a heavier workload with longer contexts and tool use can reach £3,000 to £8,000 per month. Frontier models such as Opus or the most capable GPT variants are two to five times more expensive per token, which matters for high-volume workloads. A hybrid routing strategy usually reduces total inference cost by 40% to 70%.
How do we actually run a bake-off?: Prepare a held-out task set of 50 to 150 prompts drawn from real work, with agreed correct outputs or acceptability criteria. Score blinded where possible. Measure latency from your infrastructure in your region under realistic load. Calculate cost-per-task at list price. Include operational factors such as onboarding ease, billing fit, and vendor responsiveness. A defensible bake-off finishes in two to three weeks with a written report to procurement stakeholders.
When should we fine-tune versus use the base model?: Fine-tune when you need a specific output format the base model does not produce reliably, a tone or style you cannot prompt into, or lower latency and cost that a smaller fine-tuned model can deliver on your task. Do not fine-tune to improve factual knowledge; retrieval-augmented generation is cheaper, updateable, and more auditable. Plan for the fine-tuning dataset to fall under the same governance as production data, and for periodic re-tuning as the base model family evolves.
Should we standardise on one LLM or use several?: Most mature UK deployments in 2026 are hybrid. A higher-capability model handles hard tasks, a cheaper model handles routine tasks, and specialist models cover particular use cases. Hybrid deployments often cut total inference cost by 40% to 70% against a single-model baseline. The trade-off is complexity in contracts, audit surface, and prompt engineering. Standardise on one model first, prove the value, and introduce a second model only when cost pressure or a specific capability gap justifies it.

Get new briefings by email

The AI Consultancy newsletter delivers briefings and analysis for UK businesses. We use your address only to send it, and you can unsubscribe at any time.