fairlane.systems

LLM PROVIDERS · TOOL COMPARISON

LLM providers compared: OpenAI, Anthropic, Mistral, Google, Cohere, Meta, DeepSeek, xAI, AWS Bedrock, Azure

Ten LLM providers in direct comparison, seven selection axes, one concrete recommendation per use-case. As of May 2026.

Researched & fact-checked by: · As of: 2026-05

What is the provider landscape in May 2026?

Language models no longer come from one vendor but from a market. As of May 2026 about ten serious providers are available – three US hyperscalers (OpenAI, Anthropic, Google), three specialised API vendors (Mistral from Paris, Cohere from Toronto, xAI from Texas), three open-weight vehicles (Meta Llama, DeepSeek, many Llama derivatives), and two aggregators (AWS Bedrock, Azure OpenAI) that host the other models behind a unified API.

For Swiss SMEs the LLM provider choice is the most expensive recurring decision in an AI stack. It drives data residency (data in Switzerland, the EU, the US, or China?), monthly cost (USD 0.27 vs. USD 15 per 1M output tokens), reasoning quality (can the model think through a complex tax argument?), and legal posture against the EU AI Act and the revDSG.

This page is an overview. Each major provider has its own deep-dive page with prices, models, data protection, and concrete setup notes. Click the vendor name in the table below to jump to the detail page.

Why the choice matters

Seven axes decide suitability: hosting region, price, context window, reasoning, code quality, multilinguality, DPA availability. Pick wrong and you pay with compliance risk, triple cost, or poor answer quality.

Hosting region: For fiduciary and law mandates this is the most important axis. OpenAI runs natively in the US; via Azure OpenAI you get Switzerland-North and Sweden-Central. Anthropic runs natively in the US; via AWS Bedrock also in eu-central-1 Frankfurt. Google Gemini has europe-west6 (Zurich!) – the only hyperscaler with a Swiss region. Mistral runs EU-natively in Paris. DeepSeek hosts exclusively in China – for mandate data an absolute no-go.

Price: In May 2026 input token prices range from around USD 0.28/1M (DeepSeek V3.2) to USD 15/1M (OpenAI o3-pro). Output is typically 3-5x more expensive. For typical fiduciary loads (200 queries × 5,000 tokens each) the choice between Claude Haiku and Claude Opus is the difference between CHF 8 and CHF 200 per month.

Context window: the current top Claude model 200k, GPT-4.1 1M (limited), Gemini 2.5 Pro 2M, Mistral Large 2 128k. For RAG setups, 32k is usually enough. For "full dossier in one prompt" 200k to 1M becomes relevant.

Reasoning: o3-pro and Claude Opus with Extended Thinking dominate complex reasoning tasks. Gemini 2.5 Pro with Deep Think is catching up. For pure classification and extraction, reasoning is overkill – Haiku or Mistral Small are enough.

Code: Claude Sonnet and Codestral lead on code generation. DeepSeek V3.2 surprisingly strong. GPT-4o solid mid-pack.

Multilinguality: Anthropic, Google, Mistral excellent on DE/FR/IT. OpenAI very good, not top. xAI Grok weaker on Swiss High German and Schwyzerdütsch.

DPA: Anthropic, OpenAI, Google, Mistral, Cohere offer standard DPAs. AWS Bedrock and Azure OpenAI have enterprise DPAs with annex and SCCs. DeepSeek API: no EU-compliant DPA available. xAI: legal situation May 2026 unclear, no dedicated EU tier.

The ten providers in detail

OpenAI (US, via Azure-EU also Switzerland): GPT-4o (USD 2.50 input / USD 10 output per 1M), GPT-4.1 with 1M context, o3-pro for reasoning (USD 15/60). Market leader in share and reach. Detail page: openai-gpt-evaluation.

Anthropic (US, via AWS Bedrock Frankfurt EU): the current Claude model and 4.7 (Haiku/Sonnet/Opus), Extended Thinking, 200k context. Top in code, reasoning, multilingual quality. Detail page: anthropic-claude-evaluation.

Mistral (EU-Paris): Mistral Large 2, Small 3.1, Codestral. EU hosting native, EUR billing. EU AI Act champion, DPA easy. Detail page: mistral-eu-evaluation.

Google Gemini (GCP, incl. europe-west6 Zurich): Gemini 2.5 Pro with 2M context, Flash for high throughput. Only hyperscaler with a Swiss region. Detail page: google-gemini-evaluation.

Cohere (Toronto, via AWS Bedrock EU): Command R+, R-08-2024, Rerank-3. Enterprise-RAG-focused, very good multilinguality. Detail page: cohere-evaluation.

Meta Llama (open-weight, via Groq/Together/Replicate/self-host): Llama 4, Llama 3.3 70B Instruct. Free for self-host, very cheap via inference providers. Detail page: meta-llama-evaluation.

DeepSeek (PRC, Hangzhou hosting): DeepSeek V3.2 and R1 (reasoning). Around USD 0.28/USD 0.42 (V3.2) – extremely cheap. IMPORTANT: NOT for mandate data via API – China hosting, no EU DPA. Self-host of the open weights is OK. Detail page: deepseek-evaluation.

xAI Grok (US): Grok 3, Grok 4. USD 5/15 per 1M. Twitter/X integration. As of May 2026 no EU tier, DPA status unclear. Not recommended for Swiss SMEs.

AWS Bedrock (AWS regions incl. eu-central-1 Frankfurt): aggregator hub for Claude, Mistral Large, Llama, Cohere, AI21, Titan embeddings. Enterprise default for mandates already on the AWS stack.

Azure OpenAI (Azure regions incl. Switzerland-North and Sweden-Central): GPT models with data in CH or EU. Microsoft-stack-native enterprise setup, DPA plus SCCs standard.

Selection workflow in 7 steps

  1. 01Clarify residency: must data stay in Switzerland (Azure CH, Vertex Zurich) or is EU enough (Bedrock, Mistral)? Third countries (US, CN) are forced out or in by this.
  2. 02Classify use cases: triage/classification (Haiku, Mistral Small), standard RAG (Sonnet, Mistral Large, GPT-4o), reasoning-heavy (o3-pro, Opus, Gemini 2.5 Pro Deep Think).
  3. 03Compute cost budget: expected tokens per month × model price. Typical SME setup CHF 10-200/month – model choice changes 10-20x.
  4. 04Check multilingual profile: DE/FR/IT dominant → Claude, Mistral, Gemini. EN-only → all of them.
  5. 05Get the DPA: pull the template from each candidate provider, have your DPO review it, sign before production.
  6. 06Multi-provider setup as default: LiteLLM or a custom gateway between application and provider. Always allows fallback and routing.
  7. 07PoC with 30 real tasks: benchmark the top three candidates on the same 30 tasks (quality + latency + cost). Then go production.

Recommendation by use-case

Swiss fiduciary/law with Swiss residency required: Azure OpenAI Switzerland-North (for GPT stack) or Google Vertex AI europe-west6 Zurich (for Gemini). Both deliver a real Swiss region plus enterprise DPA.

Swiss SME, EU hosting acceptable, top reasoning quality: Anthropic Claude Sonnet via AWS Bedrock eu-central-1. Top on DE/FR/IT, DPA easy via AWS.

Swiss SME with EU vendor preference: Mistral Large 2 on La Plateforme Paris. EU-native, EUR billing, EU AI Act champion.

RAG-specialised, enterprise setup: Cohere Command R+ via AWS Bedrock. Enterprise-trimmed for retrieval, very good multilinguality.

Open-weight, on-prem required, own GPU: Llama 4 or Mistral Large-2 open weights via Ollama or vLLM on Hetzner. No token pricing, but hardware investment.

High-volume classification tasks, cost focus: Haiku (Claude Haiku, USD 0.80/4 per 1M) or Mistral Small 3.1 (EUR 0.20/0.60). Overkill to spend USD 15 on o3-pro for triage and extraction.

Multi-provider setup with routing: LiteLLM or Multi-LLM gateway. Standard queries to Haiku/Mistral Small, reasoning queries to the current top Claude model/o3-pro. Saves 60-80% cost without quality loss.

Research/hobby setups without mandate data: DeepSeek V3.2 is extremely cheap and excellent at code. BUT: not for mandate data via API – only self-host of the open weights.

Providers that drop out for Swiss SMEs

DeepSeek API: hosted in mainland China. No EU DPA possible, no way to ship data in an revDSG-compliant manner. Mandate data does not go there. The open weights can be run on-prem – that is a different story.

xAI Grok: legal status as of May 2026 unclear. No dedicated EU tier, no standardised DPA. Not recommended for mandate-relevant use.

OpenAI without the Azure bridge: the direct OpenAI API hosts in the US. For Swiss professional-secrecy data, plain OpenAI is a third-country transfer with unresolved Schrems-II issues. Fix: use Azure OpenAI Switzerland-North or Sweden-Central, not api.openai.com directly.

Anthropic direct: same issue as OpenAI. The direct Anthropic API is US. Fix: use AWS Bedrock eu-central-1 Frankfurt.

Self-host without operations capacity: Llama 4 70B needs an A100 or H100, care, updates, monitoring. Without DevOps, self-host buys you only headache. API is cheaper than owner hours.

One provider for everything: vendor lock-in. An SME should keep at least two providers as fallback (e.g. Anthropic primary + Mistral secondary). A multi-LLM gateway enables this in 1-2 days of setup.

Trade-offs

STRENGTHS

  • Swiss region available: Azure OpenAI Switzerland-North + Vertex AI europe-west6 Zurich
  • EU hosting for almost any provider via Bedrock (eu-central-1) or Azure
  • Open-weight models (Llama 4, Mistral, DeepSeek) enable on-prem without vendor lock-in
  • Multi-LLM routing cuts cost by 60-80% at comparable quality

WEAKNESSES

  • OpenAI/Anthropic direct: US hosting, Schrems-II issue for professional secrecy
  • DeepSeek API: China hosting, no EU DPA, not for mandate data
  • xAI Grok: no EU tier, unclear DPA status, not recommended for SMEs
  • Self-host (Llama 4 70B, Mistral Large): GPU hardware from CHF 5,000 + operations effort

FAQ

Which provider is the best all-round pick for Swiss SMEs in May 2026?

Anthropic Claude Sonnet via AWS Bedrock eu-central-1 Frankfurt. Top reasoning, very good multilinguality, EU hosting via AWS, easy DPA. For Swiss residency: Azure OpenAI Switzerland-North with GPT-4o, or Vertex AI Zurich with Gemini 2.5 Pro.

Is DeepSeek really off-limits?

The API yes, for mandate data. Hosting in China, no EU DPA. The open weights (DeepSeek-V3, R1) are permissively licensed and can run on your own hardware or via EU inference providers like Together AI – then data stays under control. The API is the issue, not the model.

Is multi-LLM routing really worth it?

Yes, almost always. A typical fiduciary stack: 70% simple classification (Haiku or Mistral Small, ~USD 0.80/4 per 1M), 25% standard RAG (Sonnet or Mistral Large), 5% reasoning (Opus or o3-pro). Without routing all queries hit Opus – 5-10x more expensive without quality gain. LiteLLM or a custom gateway pays back in 1-2 months.

What does the EU AI Act change for provider choice?

From August 2026 (high-risk provisions take effect) providers must supply technical documentation, training-data summaries, and risk management. Mistral, Anthropic, and Google have largely documented this in May 2026. DeepSeek and xAI have not as of May 2026. For SMEs deploying in high-risk areas (HR, school grades, credit scoring) this becomes critical.

Related topics

ROUTING · AI CONCEPTMulti-LLM routing: which model when, for how muchMULTI-LLM GATEWAY · SERVICEMulti-LLM Gateway: eight providers, one entry point, compliance routingSELF-HOSTED VS. CLOUD · AI CONCEPTSelf-hosted vs. cloud LLM: a decision framework for SMEs and fiduciariesTIA · COMPLIANCEThird-country transfer and Transfer Impact Assessment (TIA): Swiss data in US and PRC cloud LLMsEU AI ACT · COMPLIANCEEU AI Act 2026: high-risk duties from 2 August 2026 – what Swiss providers must do now

Sources

  1. Anthropic – Claude Pricing and Models · 2026-05
  2. OpenAI Pricing – GPT-4o, GPT-4.1, o3-pro · 2026-05
  3. Mistral La Plateforme – EU Pricing · 2026-04
  4. Google Vertex AI – Regions (incl. europe-west6 Zurich) · 2026-05
  5. AWS Bedrock – Available Models and Regions · 2026-05
  6. Azure OpenAI – Available regions (Switzerland-North) · 2026-05

FITS YOUR STACK?

What this looks like in your business – a 30-minute intro call.

Book a call