What AI models do financial firms actually use in 2026?
Financial firms in 2026 typically run a mixed portfolio: frontier LLMs (OpenAI GPT-5-class, Anthropic Claude 4-class, Google Gemini 2/3-class) for the hardest reasoning; smaller fine-tuned or open-weight models (Llama, Mistral, Qwen) for cost-sensitive extraction; and a firm-owned embedding + retrieval layer over their own memo, filing and transcript corpus. The mix is chosen per workload, run through a firm-owned gateway with evaluation and provenance logging.
Financial firms in 2026 typically run a mixed portfolio: frontier LLMs (OpenAI GPT-5-class, Anthropic Claude 4-class, Google Gemini 2/3-class) for the hardest reasoning; smaller fine-tuned or open-weight models (Llama, Mistral, Qwen) for cost-sensitive extraction; and a firm-owned embedding + retrieval layer over their own memo, filing and transcript corpus. The mix is chosen per workload, run through a firm-owned gateway with evaluation and provenance logging.
Why "which model" is the wrong first question
The most common mistake in AI strategy at financial firms is starting with the model. The right first question is the workload: what specifically does this AI system need to do, at what latency, with what accuracy, at what cost per call, with what governance surface? The model choice falls out of that.
In practice, serious firms end up running a portfolio of models. There is no single "best" LLM in 2026 across all workloads; there is a best model per workload, and the right architecture treats models as swappable infrastructure behind a firm-owned gateway.
The frontier LLMs in production use at financial firms
- OpenAI GPT-5-class — Strong general reasoning, high tool-use reliability, excellent structured output. Widely used for research automation, memo drafting and complex extraction workflows.
- Anthropic Claude 4-class — Best-in-class long-context reasoning, strong safety/refusal behaviour, high adoption for compliance-sensitive drafting. Often preferred where output has to be defensible.
- Google Gemini 2/3-class — Excellent long-context reasoning, strong multimodal (documents, tables, charts), competitive cost per call at scale. Increasingly common in earnings-call and document-heavy pipelines.
- Google Gemini via Lovable AI Gateway — Default model in many of our client builds for cost-sensitive, high-volume extraction workloads; sits in the same portfolio alongside frontier calls.
Open-weight and fine-tuned models in production use
- Meta Llama (3 / 4 series) — Widely used for cost-sensitive extraction and workloads that must run in the firm's own VPC. Fine-tuning on firm-specific tasks is mature and cheap.
- Mistral / Mixtral — European open-weight models with strong benchmark performance; adopted by EU firms with data-residency requirements.
- Qwen (Alibaba) — Increasingly competitive on cost and open-weight; used cautiously by firms with China-exposure concerns.
- Fine-tuned smaller models — Firm-specific fine-tunes on 7B–70B open-weight models for structured extraction, classification and cost-sensitive routing.
Embeddings and the retrieval layer
Every serious financial AI system depends on retrieval — pulling the right chunks from the firm's memo library, filing corpus, transcript archive and internal knowledge base into an LLM's context. Embedding models are the workhorses of this layer, and their choice matters as much as the LLM.
Common production choices in 2026: OpenAI text-embedding-3-large for general use, Voyage AI's finance-specialised embeddings for coverage of financial jargon, Cohere Embed for multilingual coverage, and BGE / E5 open-weight embeddings for firms that need to run the embedding step in their own VPC. Most firms run a firm-owned vector store (pgvector, Pinecone, Qdrant, Weaviate) with embeddings versioned like any other model.
How model choice maps to workload
The pattern in production: research reasoning and complex drafting use frontier models (GPT-5-class or Claude 4-class). High-volume structured extraction from filings and transcripts uses cheaper models — often Google Gemini through the Lovable AI Gateway, or fine-tuned open-weight models running in the firm's own VPC. Compliance-sensitive drafting typically routes to Claude for its safety behaviour. Embedding and retrieval runs against a firm-owned vector store using either OpenAI, Voyage or open-weight embeddings.
Every call runs through a firm-owned gateway that handles model routing, cost accounting, evaluation logging and provenance. That gateway is what makes models genuinely swappable — a critical requirement in a market where the state of the art changes every quarter.
Model selection by workload (illustrative)
| Workload | Typical model choice | Why |
|---|---|---|
| Overnight research briefs across large universe | Google Gemini (cost-effective, long-context) or fine-tuned open-weight | High volume, structured output, cost matters |
| Complex memo drafting / IC papers | GPT-5-class or Claude 4-class | Reasoning quality, tool use, output polish |
| Earnings-call live extraction | Gemini 2/3 (long-context) or GPT-5-class | Speed + long context for full call handling |
| Compliance-sensitive drafting | Claude 4-class | Refusal behaviour, defensibility |
| Structured extraction from filings at scale | Fine-tuned Llama 3 / Mistral | Cheap per call, runs in firm VPC |
| Retrieval / embeddings | OpenAI text-embedding-3-large or Voyage AI finance embeddings or BGE open-weight | Choose by cost, coverage, VPC requirement |
| Classification / routing | Small fine-tuned open-weight (7B–13B) | Cheap, fast, firm-owned |
Model choices change quarterly; the point is a firm-owned gateway that makes them swappable.
Data residency, hosting and governance
Serious financial firms increasingly split their model portfolio across three hosting patterns: frontier LLM API calls (OpenAI, Anthropic, Google) for the best reasoning, with strict data-processing agreements and retention controls; VPC-hosted frontier or near-frontier models (Bedrock, Azure OpenAI, Google Cloud Vertex) for workloads that require enhanced governance; and firm-owned open-weight deployment for the most sensitive workloads.
The choice per workload is driven by data sensitivity, regulator expectations and cost. A well-designed gateway makes the hosting decision independent of the model choice — the same workload can move from a frontier API to a VPC deployment to a firm-hosted open-weight model without rewriting the application.
Related questions
Which single LLM is best for a hedge fund?
There isn't one. The right answer is a portfolio: frontier LLMs for the hardest reasoning, cheaper models for volume extraction, fine-tuned open-weight for firm-specific tasks, all behind a firm-owned gateway. The single-model question is the wrong question.
Do financial firms use open-source models like Llama?
Yes, widely. Fine-tuned Llama / Mistral / Qwen models running in firm VPCs are standard for cost-sensitive extraction and any workload with data-residency requirements. They complement rather than replace frontier LLMs.
Are firms building their own foundation models?
Almost none. Building a foundation model from scratch is a $100m+ investment justified by only a handful of firms globally. Serious firms fine-tune existing open-weight models and build their proprietary advantage in retrieval, tools, evaluation and the reasoning workflow — not in the base model.
What about domain-specific models like BloombergGPT?
BloombergGPT and similar were interesting research artefacts but have largely been overtaken by frontier general-purpose LLMs. Very few firms use domain-specific foundation models in production; most build domain adaptation via prompting, retrieval and fine-tuning against a general model.
How often does the model portfolio need to change?
Expect meaningful model changes at least quarterly. The firm-owned gateway and evaluation harness are what make this a routine ops task rather than a re-architecture — a critical reason to build the gateway before the first production system, not after.
The model landscape in 2026 rewards firms that treat models as swappable infrastructure behind a firm-owned gateway, not as strategic bets. The compounding advantage lives in the retrieval, the evaluation harness, the tool use and the reasoning workflow — not in the choice of which frontier LLM is best this quarter.
The £15,000 Financial AI Blueprint gives you a firm-specific model and gateway plan sized to your workloads.