Which AI model for which finance task

Model choice for finance tasks depends on complexity (simple text versus multi-step reasoning), context size (one invoice versus annual accounts), cost per token and privacy route (EU hosting versus US default). For controllers and CFOs that means no one-size-fits-all, but a matrix of three or four models for different task types. That yields better output at lower cost.

Model landscape updated 10 September 2026. This is the only page in our resources carrying concrete model names, context windows and rates. The other articles describe roles ("the fast model", "the reasoning model") so they do not go stale. We refresh this page quarterly. Always check rates with the vendor before you budget on them.

"Which AI model is best?" is the wrong question for finance. The right question is which model is best for this finance task. An AP clerk classifying 200 bank transactions a day needs a different model than a controller writing an IFRS disclosure. Leaning on one model by default, usually the heaviest one to be safe, leaves money on the table for routine work and quality on heavy work.

The five axes that matter for finance

1. Reasoning depth. Can the model reason in multiple steps, make assumptions explicit, weigh counterarguments? Crucial for due diligence, contract analysis and heavy IFRS questions. Unnecessary for classification and summarisation.
2. Speed and cost. A fast, cheap model for routine work, the heavy models for what matters. The cost difference per token runs up to a factor of ten.
3. Context window. How much text fits in one request? For annual-report review, long contract portfolios or file analysis this decides whether full-context reading is possible.
4. Capability set. Code execution for number work, file creation for board packs, vision for invoices. A strong model with the wrong capabilities gets you nowhere.
5. Privacy and tier choice. There is no better model that fails the privacy test. For finance data: enterprise tier or API with a data processing agreement, never a consumer account.

The model landscape as of 10 September 2026

Every vendor works with the same three-step ladder: a light model for volume, a mid model as workhorse, and a heavy model for work that matters. Pick the rung first, the brand second.

Anthropic (Claude)

Context window 200K standard, 1M on enterprise. Cache hits are billed at ten percent of the input rate, and the batch API halves both rates for work that does not need an immediate answer. For bulk classification of bank transactions that matters more in practice than the model choice itself.

OpenAI (ChatGPT)

The GPT-5.6 generation comes in three variants: Sol as flagship, Terra as balanced mid tier and Luna as the fast, cheap option. Sol sits at $4 in and $20 out per million tokens, an introductory rate running through late November 2026. There is also a fast mode roughly 2.5 times quicker at double the rate, which finance rarely needs.

Google (Gemini)

Gemini 3.8 Flash is the current generation, at an introductory $0.75 in and $3.75 out per million tokens through 31 December 2026. From 1 January 2027 that doubles to $1.50 and $7.50. If you build a business case on Gemini, budget the 2027 rate straight away. The context window is the largest of the three.

Microsoft Copilot

Copilot is not a model but a layer on top of third-party models, sold per user rather than per token. See the cost article for what that means for your budget.

Perplexity

Not a chatbot but a research assistant. The models underneath vary; the difference sits in retrieval and source handling.

The six players through a finance lens

Claude

Strongest at: reading long documents (close reading of contracts, annual accounts, transfer pricing reports), nuanced writing, careful analysis. Excellent for financial analysis where every clause counts.

Use for: annual-report review, contract analysis, IFRS disclosures, due diligence, final editing of board memos, complex variance analyses.

ChatGPT

Strongest at: versatility, strong code execution for data analysis via Python, agent mode, deep research. The reasoning model is the workhorse for heavy analysis.

Use for: data analysis on large Excel extracts, raw strategic reasoning, deep research into market or benchmark data, multi-round brainstorming.

Microsoft Copilot

Strongest at: sitting where finance work already happens, so Outlook, Word, Excel, Teams and SharePoint. Access to the Microsoft Graph with the right license. Enterprise governance and EU residency are well covered, provided Flex Routing is off, which it is not by default.

Use for: anything involving your own Outlook, Excel models and SharePoint documents. Not for open-ended creative work or heavy reasoning, where it is mediocre.

Gemini

Strongest at: the largest context window, factual grounding via Google, strong multimodal processing.

Use for: entire files (500 pages of transfer pricing documentation, contract portfolios), questions where current search results help, long PDF analysis.

Perplexity

Use for: current VAT rates or tax developments, sector benchmarks such as DSO and EBITDA margins, competitive research for a board memo. Anything where source citation is not optional.

Grok

Little relevance for finance. More important than the brand is the principle: do not pick a platform whose data processing is subject to ongoing regulatory scrutiny, and check that status before you put personal data through a tool.

Decision tree per finance task

• Reading a long contract portfolio or annual report (50 to 300 pages) → a heavy model in the 200K class, or Gemini. Full context, highest quality.
• Huge file (500 pages and up, due diligence) → the largest context window you have, or a RAG approach.
• Multi-step reasoning (M&A analysis, transfer pricing, complex IFRS) → the OpenAI reasoning model or Claude Opus with extended thinking.
• Current finance or tax research with sources → Perplexity for the research, then Claude for the synthesis.
• Memo, board text, variance commentary in proper tone → Claude Sonnet.
• Working inside your Outlook, Excel or SharePoint → Copilot.
• Code, SQL, data analysis on an Excel extract → Claude Opus, the OpenAI reasoning model, or a Q&A agent that works on your Exact data via MCP.
• Bulk classification of bank transactions, invoices or queries → Claude Haiku, the fast OpenAI variant or Gemini Flash. A factor of ten cheaper.
• Extracting receipts or invoices → a model with strong vision, usually Claude or Gemini.

Cost versus quality: when heavy is overkill

The reflex to always take the heaviest model costs money. Three finance rules of thumb:

• The light model costs about a fifth per token of the heavy one. For classifying bank transactions, extracting from invoices and short summaries, light is fine.
• The fast variant is often eighty percent as good as the heavy one at a fraction of the cost, and considerably quicker. For draft emails and internal notes: fast.
• In chains with multiple steps, use fast models in the intermediate steps and deploy the heavy model only on the final synthesis.

Always start with the fast model and upgrade only when the output measurably falls short. For finance there is extra discipline on top: always enable code execution on number work, whichever model you use. Otherwise you are winning the model debate on the wrong dimension.

What this costs per month and how you calculate the saving back sits in what AI costs in a finance team.

Multi-model patterns for finance

• Perplexity, then Claude, then Copilot. Market research for a board pack, Claude builds the analysis, Copilot drops it into the Word template and mails it round.
• Gemini, then Claude. Gemini reads the full 500-page contract portfolio, Claude rewrites the analysis in formal language.
• Light, then heavy. The light model classifies incoming invoices by type, the workhorse produces the final posting proposals for the edge cases. Cheap on volume, strong on the complex cases.

Privacy as the first filter

For finance, model choice starts with privacy rather than quality. For data containing customer or personnel information:

• Consumer accounts (free Claude.ai, ChatGPT Plus): never Tier 3 or Tier 4 finance data.
• API with zero data retention, enterprise tier, Copilot for Business: suitable for Tier 1 through 3, provided there is a data processing agreement.
• A deployment on client-owned infrastructure: Tier 4 as well, because the entire stack runs inside your own environment.

A better model that fails the privacy test is not an option. Privacy first, then compare quality.