How good are the best large language models in 2026?

How well do the newest large language models (LLMs) perform on real world tasks, and how fast are they advancing?

In the beginning of 2026, a series of new releases of the Claude AI model by the company Anthropic which included plugins for legal and finance collectively wiped billions off the value of software companies like Salesforce and Adobe, as investors became concerned that agentic AI could replace the functionality of traditional software-as-a-service offerings.

Meanwhile, Anthropic has filed two lawsuits against the Trump administration after the US Defense Department designated Anthropic a “supply chain risk”, essentially blacklisting Pentagon suppliers from using Claude. Anthropic’s usage policy prohibits using Claude for domestic surveillance or in fully autonomous weapons.

So, could the new AI models pass a law exam and has this been a significant improvement over 2025’s models?

The top large language models can now score consistently close to 80% on a law exam when connected to a retrieval augmented generation (RAG) system. However, improvements are becoming more incremental.

Above you can see a graph of large language models by release date, showing their score on a mini “law exam” prepared for the Insolvency Bot paper. We now have over two years of data points, dating from the early days of LLMs (GPT 3.5).

How I evaluated the LLMs

I have evaluated the new LLMs that have been released in 2026, such as the latest GPT versions, Claude from Anthropic, and Qwen from the Chinese company Alibaba Cloud (Aliyun, 阿里云). I tested them on our insolvency law exam which we used for our Insolvency Bot publication[1]. I did this same evaluation a year ago and since then a large number of ground-breaking new models have been released.

As reported in the media, the Chinese large language models such as DeepSeek and Qwen have generally performed as well as the American models, although they have been trained for a fraction of the cost, and sometimes without the access to the hardware that the American companies have.

I have also compared each large language model in its unmodified form, together with a RAG (Retrieval Augmented Generation) setup where I am sending prompts containing relevant parts of statute, case law and other information.

What’s interesting to see is that the latest large language models seem to have inbuilt knowledge of some English law. Since I am using the models via their API, they are not retrieving data from the internet (unless this is specifically turned on). However, the models’ performance improves dramatically with the RAG setup, which forces the responses to be given in relation to English and Welsh law (as opposed to American law, or any other content which may be on the internet).

It was also informative to see the latest Qwen perform worse than the other LLMs without RAG, but just as well as the other LLMs once RAG was included. This could reflect Alibaba Cloud not having much English legal text in their training dataset.

Example question and answer pair

Here is an example question which we used to evaluate the chatbots:

I set up a limited liability company in 2015. My business suffered greatly during the Covid-19 pandemic, and it still hasn’t recovered yet. I fell into arrears, and one of my suppliers submitted a winding-up petition against my business. A friend of mine is willing to invest money in my company. Do I have good chances of avoiding a liquidation order if my friend commits to pumping money into my business?

Below is Claude 4.6 Opus’s answer. For each question, multiple answers were generated: two for every LLM that I evaluated, one with RAG (the black data points in the graph above) and one without RAG (the green data points). You can find the full evaluation data set on our Github: https://github.com/fastdatascience/evaluate_insolvency

# Resisting a Winding-Up Petition Through New Investment

This is a significant situation that requires careful consideration. Let me walk you through the key issues, though I must stress you should **seek urgent legal advice from an insolvency practitioner or solicitor** given the time-sensitive nature of winding-up proceedings.

## The Court's Key Consideration

When deciding whether to make a winding-up order, the court primarily asks: **is the company unable to pay its debts?**

Under the Insolvency Act 1986, this typically means:
- The company cannot pay debts as they fall due (**cash-flow insolvency**), or
- The company's liabilities exceed its assets (**balance-sheet insolvency**)

## Can Your Friend's Investment Help You Resist the Petition?

Potentially, but it depends on several critical factors:

### Factors That Would Strengthen Your Position

- **Concrete, unconditional commitment** — not just a vague promise but a binding, documented agreement to inject funds
- The investment being **sufficient to pay the petitioning creditor's debt in full**, plus costs
- The investment being enough to demonstrate the company can also **meet its other obligations** going forward
- Evidence of a **viable business going forward** — the court will want to see this isn't just delaying the inevitable
- A **credible business plan** showing how the company will return to solvency

### Factors That Could Undermine Your Position

- If the commitment is **conditional or uncertain**
- If the investment only covers the petitioning creditor but leaves other debts unpaid — the court may still find the company is insolvent
- If there are **multiple creditors** supporting the petition
- If the court views the investment as merely **"throwing good money after bad"**
- If there's a history of broken promises or failed rescue attempts

## Important Practical Points

**The court has discretion.** Even if you can pay the petitioning creditor, the court can still make a winding-up order if it believes the company is fundamentally insolvent. The court will look at the **overall picture**, not just one debt.

**Timing matters enormously.** You need to:
- Act quickly — there are strict procedural deadlines
- Ideally pay or secure the petitioner's debt **before the hearing**
- If you pay the petitioning creditor in full before the hearing, they may consent to the petition being dismissed — though other creditors can substitute themselves as petitioner

**Other creditors can support the petition.** Even if your friend's money satisfies the original petitioner, other creditors may appear at the hearing and support the winding-up order.

## Alternative Options to Consider

- **Company Voluntary Arrangement (CVA)** — a formal agreement with creditors to repay debts over time
- **Administration** — which imposes a moratorium on creditor actions
- **Negotiated settlement** — using your friend's investment to negotiate directly with creditors
- **Informal arrangement** — paying the petitioning creditor to withdraw and restructuring other debts

## What You Should Do Now

1. **Instruct an insolvency solicitor immediately** — this is not something to navigate alone
2. **Get your friend's commitment in writing** with clear terms
3. **Prepare a realistic business plan** showing viability post-investment
4. **Identify all your creditors and total liabilities** — not just the petitioner
5. **Consider whether the investment is genuinely sufficient** to restore solvency, not just pay one debt

## Honest Assessment

Having a willing investor is certainly better than not having one, but it is **not a guarantee** of avoiding liquidation. The court will scrutinise whether the investment genuinely restores the company to solvency or merely delays an inevitable collapse. The stronger and more concrete the evidence of a viable future, the better your chances. **This is genuinely urgent — please seek professional legal advice as soon as possible.** The consequences of a winding-up order are severe and largely irreversible.

Here is the mark scheme for the above question:

You can read our original paper (which predates the release of DeepSeek) here:

• Marton Ribary, Paul Krause, Miklos Orban, Eugenio Vaccari, Thomas Wood, Prompt Engineering and Provision of Context in Domain Specific Use of GPT, Frontiers in Artificial Intelligence and Applications 379: Legal Knowledge and Information Systems, 2023. https://doi.org/10.3233/FAIA230979

And you can try the Insolvency Bot here: https://fastdatascience.com/insolvency

Are people prepared to seek advice online or use chatbots?

In our survey of 109 small business owners which we conducted for the Insolvency Bot study, the most commonly cited port of call for advice was Citizens Advice. People would also use the free thirty minute consultation offered by many solicitors, but many small business owners would consider the Insolvency Bot, the internet, or other chatbots. You can see their responses in the graph below:

References

1. Ribary, Marton, et al. Prompt Engineering and Provision of Context in Domain Specific Use of GPT. Legal Knowledge and Information Systems. IOS Press, 2023. 305-310.