How to choose an AI model: frontier, open-weight or local

What is the difference between frontier, open-weight and local models?

A frontier model is the most capable model a lab offers, available only through its API or a cloud partner; an open-weight model is one whose trained parameters are published so anyone can download and run it; a local model is any model, usually open-weight, running on hardware you own or rent exclusively.

The 3 categories trade capability for control. Frontier models such as Claude Fable 5.1, GPT-6 Astra or Gemini 3.8 Flash are the strongest and the simplest to use, but your prompts go to the vendor's servers. Open-weight models such as Mistral Small 4, DeepSeek V4, Qwen 3.8 or Kimi K3 can be served by a cloud provider of your choice or by you, and the licence decides what you may do with them. Local means the data never leaves the building.

Which frontier models exist in September 2026 and what do they cost?

The frontier tier in September 2026 is Anthropic's Claude Fable 5.1 and Opus 5, OpenAI's GPT-6 Astra and GPT-5.6, and Google's Gemini 3.8 Flash and 3.1 Pro, with list prices between $0.75 and $10 per million input tokens and between $3.75 and $50 per million output tokens.

A token is roughly 0.75 words, so 1 million tokens is about 700,000 words. According to Anthropic's pricing page, Claude Fable 5.1 costs $10 per million input tokens and $50 per million output tokens, Claude Opus 5 costs $5 and $25, Claude Sonnet 5 costs $2 and $10 and Claude Haiku 4.5 costs $1 and $5, with a 1 million token context window on the Claude 5 models.

OpenAI's pricing page lists GPT-6 Astra, its flagship released in September 2026, at $10 and $50; GPT-5.6 Sol at $4 and $20 under promotional pricing through at least 21 November 2026; GPT-5.6 Terra at $2 and $12; and GPT-5.6 Luna at $0.20 and $1.20. Google's Gemini API pricing lists Gemini 3.8 Flash at $0.75 and $3.75 through 31 December 2026, rising to $1.50 and $7.50 afterwards, and Gemini 3.1 Pro Preview at $2 and $12 up to 200K tokens. Batch processing is 50% off at all 3 vendors, and cached input costs 10% of the base price or less at Anthropic and OpenAI.

Which open-weight models are worth considering?

The open-weight models worth considering in 2026 are Mistral Small 4 and Large 3 (Apache 2.0), DeepSeek V4 Pro and V4 Flash (MIT), Qwen 3.8 (Apache 2.0 for the 27B model) and Kimi K3 (custom licence), because each has a permissive licence and a context window of 256K tokens or more.

Mistral is the European option. Mistral Small 4, released in March 2026, has 119 billion total parameters with 6 billion active per token, a 256K context window, an Apache 2.0 licence and an API price of $0.15 per million input tokens and $0.60 output, according to Mistral's announcement. DeepSeek V4 Pro is a 1.6 trillion parameter mixture-of-experts model with 49 billion active parameters, a 1 million token context and an MIT licence, according to its Hugging Face model card. Check the licences of the rest: the largest Qwen 3.8 model and Kimi K3 carry attribution clauses above 100 million monthly users or $20 million in monthly revenue, and Meta's newest Muse Spark models are API-only. Read the licence before you build, not after.

• Mistral Small 4: 119B total / 6B active, 256K context, Apache 2.0, $0.15 / $0.60 on Mistral's API.
• DeepSeek V4 Pro: 1.6T total / 49B active, 1M context, MIT; V4 Flash: 284B / 13B, MIT.
• Qwen 3.8: 27B model under Apache 2.0, the 2.4T flagship under a custom licence.
• Kimi K3: 2.8T total / 104B active, 1M context, custom licence with attribution clauses.

When does a local model make sense?

A local model makes sense when the data must not leave your control, when you need predictable latency without network round trips, or when your volume is high enough that owning hardware beats paying per token. If none of the 3 applies, use an API.

The hardware has caught up faster than most CEOs realise: a workstation with 512 GB of unified memory can now load the largest open-weight models entirely on device, and a single 32 GB consumer GPU runs a 27B model comfortably. In my companies local models handle classification, extraction and first drafts on confidential documents. Anything that needs judgement across a long context still goes to a frontier model, because a wrong answer costs more than the tokens.

• Use local when: regulated or confidential data, on-premise requirements, offline use, very high steady volume.
• Do not use local when: fewer than 10 million tokens a month, no engineer to own it, tasks that need frontier judgement.

How do you keep AI data in the EU?

You keep AI data in the EU by choosing an EU inference endpoint, which every major vendor now offers through its own API or a cloud partner: Claude through the EU regions of Amazon Bedrock or Google Cloud, OpenAI through its European API region, Mistral hosted in the EU by default, and any open-weight model on servers you choose.

According to Anthropic's documentation, Claude on Amazon Bedrock supports EU inference profiles in Frankfurt, Zurich, Stockholm, Milan, Spain, Ireland, London and Paris with a 10% premium for regional endpoints, and on Google Cloud the Claude 5 models use the "eu" multi-region endpoint. Anthropic's first-party API offers a US-only option but no EU-only option yet, so European residency for Claude goes through the cloud partners. OpenAI's data residency guide states that its Europe region performs inference in region, and Mistral's help centre states that customer data is hosted in the European Union by default.

Then there is the law. The EU AI Act's obligations for general-purpose AI models have applied since 2 August 2025, its transparency rules came into effect in August 2026, and the AI Omnibus, in force since 27 July 2026, moved most high-risk obligations to 2 December 2027, according to the European Commission. For a company using models in support, back office or sales, the practical duties today are transparency and data protection, not high-risk conformity.

What is a simple framework for choosing an AI model?

Split your AI work into 3 buckets and assign a tier to each: judgement tasks go to a frontier model, volume tasks go to a mid-tier model, and confidential or regulated tasks go to an open-weight model on EU or local infrastructure. Then route every request to the cheapest tier that passes your own evaluation.

Your own evaluation is the key phrase. Leaderboards such as Arena and Artificial Analysis are useful to shortlist, and in September 2026 the Artificial Analysis Intelligence Index shows Claude Fable 5.1, GPT-6 Astra and Claude Opus 5 at the top with small gaps. But a model that wins a coding benchmark may lose on your Spanish invoices. Build 50 to 100 real examples with correct answers, run every candidate and measure cost per completed task, including retries and human corrections.

This is how my companies run it: roughly 10% of requests go to Fable 5.1 or Opus 5, around 80% to Sonnet 5, Gemini 3.8 Flash or GPT-5.6 Terra, and the rest to an open-weight model on infrastructure we control. Re-evaluate every quarter: Sonnet 5 already had a planned price increase cancelled and Gemini 3.8 Flash doubles in price in January 2027.