Board North

Updated 26 September 2026

The AI market is moving quickly, and for finance leaders the terminology can be confusing.

You may hear names such as Claude Fable 5, GPT-6 Astra, GPT-6 Sol, GPT-6 Luna, GPT-5.6, Jev, Kimi K3, DeepSeek V4, Gemini and Copilot used as though they are all the same type of product. They are not.

Some are AI models. Others are applications or platforms that give you access to AI models.

For CFOs and Finance Directors, the important question is no longer simply, “Which AI tool should we buy?” It is increasingly, “Which model is best suited to the work we are trying to do?” Different models now have materially different strengths around reasoning, speed, cost, vision, automation and infrastructure.

First: Copilot is not a model

One of the most common misunderstandings is that Microsoft Copilot is itself an AI model. It is not.

Copilot is better thought of as an application layer through which AI can be used inside the Microsoft environment.

The model is the intelligence. Copilot is one of the environments through which that intelligence can be accessed.

The same distinction applies elsewhere. ChatGPT is an application through which OpenAI models are accessed. Claude is Anthropic’s AI product and model family. Copilot sits within Microsoft’s ecosystem and can bring AI into applications such as Excel, Outlook, Word, PowerPoint and Teams.

So when a business says, “We use Copilot”, that does not necessarily tell you which underlying model is doing the work. That distinction will matter more as businesses begin using different models for different tasks.

Who are the main AI model providers?

Company	Models / product
Anthropic	Claude Fable, Opus, Sonnet
OpenAI	GPT-6 Astra, Sol and Luna; GPT-5.6 family
Google	Gemini
xAI	Grok
DeepSeek	DeepSeek V4
Moonshot AI	Kimi K3
Microsoft	Copilot is an application layer rather than a model in this comparison
TypeSafe AI	Jev (System One model for structured decisions)

Finance leaders do not need to memorise every model release. What matters is understanding that these models are not interchangeable. They can behave very differently on the same task.

Claude Fable 5: the strongest model overall

Claude Fable 5 is currently the strongest model overall.

Its main strength is the depth of its reasoning. It can work through difficult problems, understand complex objectives and identify things that other models can miss.

That makes it particularly relevant for senior finance work where there is not necessarily one obvious answer. Examples include:

• acquisitions and due diligence
• strategic analysis
• reviewing business cases
• challenging forecasts
• analysing complex board information
• investigating commercial problems
• reviewing large amounts of interconnected information
• working through ambiguous issues
• identifying weaknesses in assumptions

The difference becomes clearer when you compare two types of AI task.

Basic task: “Summarise this board paper.”

Higher-value task: “Review this board paper, identify the assumptions that appear weak, highlight inconsistencies between the operational and financial information, and identify the questions the board should be asking.”

The second requires much more than summarisation. It requires reasoning. That is where Fable 5 is strongest.

Fable is extremely capable, but that does not mean it behaves perfectly. It can sometimes take shortcuts, lose track of what it is doing or act in ways that need managing. Its strength is that it currently shows the deepest capability on complex work.

GPT-6: the new OpenAI generation

GPT-6 is now OpenAI’s current frontier generation. Astra is the highest-capability model in the family, with Sol and Luna bringing the same generation into faster and lower-cost workloads.

For finance leaders, GPT-6 matters because it extends the move away from using one model for everything. The family is designed around different levels of complexity, speed and cost.

GPT-6 Astra

GPT-6 Astra is OpenAI’s most capable model. It is built for complex, multi-step professional work, including research, computer use, document creation and long-running workflows.

For finance, the most relevant use cases include complex due diligence, strategic analysis, large document sets, cross-functional investigations, scenario work and agentic processes that need to use tools as well as reason.

Board North’s current view is that Astra belongs in the same frontier conversation as Claude Fable 5. Fable 5 remains our strongest overall choice in this review, but Astra is the most important new challenger and should be tested directly against it on high-value finance work rather than judged from vendor benchmarks alone.

GPT-6 Sol

GPT-6 Sol brings much of the GPT-6 generation into a lower-cost model. It is relevant for substantial professional work where the task is too demanding for a small model but does not justify using the highest-cost frontier model for every step.

Finance use cases include management information analysis, forecasting, structured research, reporting, large reconciliations and multi-stage analytical workflows.

GPT-6 Luna

GPT-6 Luna is the fast, lower-cost model in the GPT-6 family. Its role is high-volume work: classification, extraction, summarisation and background automation where economics matter as much as raw intelligence.

That makes it relevant to accounts payable triage, transaction classification, document routing, recurring data checks and other workloads that may involve thousands of model calls.

GPT-5.6 remains useful context because it provides a clear comparison of how model capability, token efficiency and cost trade off. However, for new OpenAI deployments, GPT-6 is now the generation finance teams should be evaluating first.

GPT-5.6 Sol: one of the strongest models available

GPT-5.6 Sol is also an exceptionally capable model.

It performs strongly on long-running and difficult tasks and is able to handle substantial pieces of work across multiple stages.

Its token efficiency is also important. A model can appear expensive based on its advertised token price but still prove economical if it reaches a result using significantly fewer tokens than competing models.

Sol performs particularly well where a model needs to stay engaged with a substantial task over an extended period rather than simply answer a single question.

For finance teams, that makes it relevant to work such as:

• analysing management information
• researching complex issues
• reviewing large datasets
• producing structured reports
• analysing multiple documents together
• carrying out multi-stage financial analysis
• supporting longer automated workflows

Fable and Sol sit materially ahead of many other models when the work becomes genuinely complex. The gap is not simply about producing slightly better wording. It is about being able to complete categories of work that weaker models still require a person to manage closely.

GPT-5.6 Luna: high-volume, lower-cost processing

Luna serves a different purpose. It is fast, inexpensive and capable enough for a large number of routine processing tasks.

Its strengths include:

• summarisation
• classification
• creating structured data from unstructured information
• background processing
• simple tool-based tasks

That makes it particularly interesting for finance automation.

Consider an accounts payable inbox receiving thousands of messages. The AI might need to:

1. identify the supplier
2. identify the invoice number
3. understand the nature of the query
4. classify it
5. route it to the appropriate person
6. structure the information for another system

There is little commercial logic in using the most powerful model available for every one of those steps. A smaller model can often perform repetitive processing much more economically.

The same principle could apply to transaction classification, email triage, extracting data from documents, routine report summarisation, categorising information and first-pass processing before exceptions are escalated.

The strongest model is not automatically the right model for every job.

Jev: a different type of AI for automated decisions

Jev is important because it is not trying to be another chatbot or general-purpose writing model. It is TypeSafe AI’s first System One model, designed to make fast, structured decisions inside software.

Instead of generating paragraphs of text, Jev takes unstructured information and returns typed decisions, probabilities and confidence scores. The possible outputs are defined in advance, which makes it particularly suited to workflows where software needs to decide what happens next.

For finance, that opens up a different set of use cases:

• classifying supplier or customer queries
• routing invoices and exceptions
• scoring transactions or cases against defined criteria
• deciding whether an item can proceed automatically or needs human review
• prioritising collections, approvals or exceptions
• acting as a fast decision layer inside larger AI workflows

Jev is designed for very low latency and very low input cost, with no output-token charge because it does not generate free-form text. It also returns a confidence measure with every decision, allowing a business to set thresholds for when the system can act automatically and when it should escalate to a person.

That does not mean Jev cannot be wrong. It can make an incorrect decision. The difference is that it is constrained to defined outputs and reports uncertainty, which makes it easier to build controls around than open-ended text generation.

For finance automation, Jev is therefore less a replacement for Fable, Astra or Sol and more a complementary decision engine. A workflow could use Jev to classify and route thousands of items, then escalate the uncertain or complex cases to a frontier reasoning model.

Kimi K3: strong capability with vision

Kimi K3 is a highly capable open-weight model. It performs particularly well on longer, more complex tasks and has strong visual capability.

That visual capability matters because business information increasingly comes in more than one format. AI may need to understand text, screenshots, charts, images, documents, diagrams and data presented visually.

Kimi K3 can handle that type of multimodal work well.

However, there is an important commercial point: open weight does not automatically make a model cheap. Large models require substantial infrastructure to run.

For businesses evaluating AI economics, the calculation needs to include model cost, infrastructure, compute, integration, implementation and maintenance.

DeepSeek V4 Flash: strong value and deployment flexibility

DeepSeek V4 Flash is another strong model, particularly where cost and deployment flexibility matter.

It is open weight, which gives organisations more choice over how and where it is deployed. For businesses with specific security, infrastructure or data requirements, that can be important.

Its major limitation is that it does not have vision capability. That means it is less suitable where the workflow requires the AI to interpret visual information.

For finance teams, that could matter where processes involve scanned documents, screenshots, charts, images embedded in reports or visual information in PDFs.

The question is not simply, “How intelligent is this model?” It is, “Can this model work with the type of information our process actually uses?”

Grok 4.6 and Muse Spark: useful for specific tasks

Not every model needs to be the strongest overall to be commercially useful.

Grok 4.6 can handle complex and multi-stage tasks, although its efficiency is weaker than some alternatives.

Muse Spark is particularly notable for speed. It can process and analyse large amounts of information quickly, which could make it useful where rapid processing matters more than maximum reasoning capability.

Sometimes the right model is simply the fastest model that is sufficiently accurate for the task.

Gemini: the importance of token efficiency

Gemini currently has a significant problem around efficiency. The issue is not simply the advertised token price. It is the number of tokens the model may require to complete the work.

This is important because finance leaders can easily be misled by headline pricing.

Model A costs £5 per million tokens.
Model B costs £10 per million tokens.

At first glance, Model A looks cheaper. But if Model A requires five times as many tokens to complete the same task, Model B is actually considerably cheaper.

The better commercial measure is cost per successfully completed task rather than cost per million tokens. This becomes particularly important when AI is being used thousands or millions of times within business processes.

Models from the same company can perform very differently

Another mistake is assuming that choosing a provider means all of its models are broadly equivalent. They are not.

Within Anthropic, Fable, Opus and Sonnet have materially different capabilities. Within OpenAI, GPT-6 Astra, Sol and Luna – as well as the GPT-5.6 family – have different capability, speed and cost profiles. Within DeepSeek, V4 Flash and V4 Pro have different economics and characteristics. Jev is different again: it is built for structured decisions rather than open-ended generation.

So saying “We use Claude” or “We use GPT” will increasingly be as incomplete as saying “We use Microsoft.” The specific model matters.

Which models are most relevant to finance?

Claude Fable 5
Strongest overall model. Best suited to difficult reasoning, complex knowledge work, strategic analysis and problems where the AI needs to understand what is really happening rather than simply process instructions.

GPT-6 Astra
OpenAI’s highest-capability model. Relevant to complex, multi-step professional work and a direct model to evaluate alongside Fable 5 for high-value finance tasks.

GPT-6 Sol
Strong option for substantial professional workloads where frontier capability needs to be balanced against cost.

GPT-6 Luna
Fast, lower-cost GPT-6 model for high-volume automation, classification, extraction and summarisation.

Jev
A specialist decision model for structured, probabilistic choices inside software. Particularly relevant to routing, scoring, triage and exception handling at scale.

GPT-5.6 Sol
One of the strongest models for substantial and complex work. Particularly relevant where the AI needs to sustain performance across a long or multi-stage task.

GPT-5.6 Luna
Strong option for fast, inexpensive processing. Particularly useful for high-volume automation, classification, extraction and summarisation.

Kimi K3
Strong multimodal and open-weight model. Relevant where visual capability and deployment flexibility matter.

DeepSeek V4 Flash
Strong value and open-weight option. Relevant for text-based workloads where deployment flexibility is important, although lack of vision limits some use cases.

Grok and Muse
Useful specialist options. Potentially valuable where their particular strengths around orchestration or speed match the process.

Gemini
Currently less compelling on efficiency. Its token usage means businesses need to look carefully at cost per completed task rather than headline token pricing.

The bigger opportunity: use different models for different tasks

The most important development is that businesses do not need to choose one model. A finance workflow can use several.

Imagine an automated month-end review.

Stage 1: Process the volume

A lower-cost model – or a decision model such as Jev – analyses thousands of transactions, classifies them and identifies anomalies.

Stage 2: Investigate exceptions

A more capable model examines the unusual items and determines which require further investigation.

Stage 3: Apply deeper reasoning

The most complex issues are then passed to a model such as Fable 5 for deeper analysis.

Stage 4: Present the result

The output is converted into a management summary or board-level report.

The user may never need to know which model completed each step. The system can select the most appropriate model depending on the task. This is known as model routing.

For businesses, it could provide an important balance between capability, speed and cost.

What should CFOs and Finance Directors actually assess?

Finance leaders do not need to become AI engineers. They do, however, need enough understanding to ask the right questions.

1. Capability
Is the model intelligent enough for the complexity of the task?

2. Reliability
Does it consistently produce acceptable results on real business examples?

3. Cost
What does each completed task or process actually cost?

4. Speed
Does the workload require a response in seconds, minutes or hours?

5. Data
Can the model work with the types of information involved?

6. Integration
Can it access the applications and systems needed to complete the process?

7. Governance
Where does the data go, who can access it and how are outputs controlled?

Those questions are considerably more important than simply choosing whichever AI brand is most familiar.

The Board North view

The AI market is becoming increasingly specialised.

• Claude Fable 5 remains Board North’s strongest overall choice for difficult reasoning and complex knowledge work.
• GPT-6 Astra is now OpenAI’s highest-capability model and should be evaluated directly alongside Fable 5 for complex, high-value finance work.
• GPT-6 Sol and GPT-6 Luna extend the OpenAI range into substantial professional workloads and high-volume, lower-cost automation respectively.

Jev adds a different capability: fast, structured, confidence-scored decisions for software workflows, making it particularly relevant to finance automation at scale.

• Kimi K3 provides strong complex and visual capability with open-weight deployment options.
• DeepSeek V4 Flash combines strong value with open-weight flexibility, although its lack of vision limits some applications.
• Grok and Muse offer useful characteristics for particular workloads.
• Gemini currently has weaker economics where high token usage drives up the real cost of completing tasks.

Microsoft Copilot is not another model to add to that list. It is part of the application layer through which AI can be delivered into the working environment.

From choosing an AI tool to designing an AI operating model

Most businesses are still asking, “Should we use Copilot, ChatGPT or Claude?” That question is becoming too simplistic.

The more important question is: “Which parts of our finance function should use AI, which model is best for each part, and how should those models interact safely with our systems and data?”

That moves AI beyond individual productivity.

Different models can begin to sit inside:

• month-end
• management reporting
• forecasting
• accounts payable
• financial analysis
• data processing
• research
• controls
• workflow automation

The future is unlikely to be one business selecting one AI. It is much more likely to be different models performing different jobs within the finance operating model – with the strongest models reserved for the work where their additional intelligence genuinely adds value. Model availability and pricing change quickly. Board North reviews the model landscape as part of its AI and automation work with finance teams and updates recommendations as new

#AI #Models #Finance #CFO #FD #FP&A #FC