Before offering a new model in ToltIQ, our Research Team tests the model within the platform on benchmarks built around real Private Equity (PE) documents. Models must meet our data retention standards and perform strongly on the benchmarks before they become available for our clients. Opus 5.5, Sonnet 5.5, and GPT-6.1 Sol are the three models we currently offer in ToltIQ’s Chats feature. 

Executive Summary

1. The ToltIQ Research Team tests every major model release that meets our data retention standards with benchmarks in ToltIQ’s Chats feature, using the same retrieval and tools our clients use. We plan to continue updating this article as major models are released.

2. Frontier models are built for different tasks, so a model’s lead in general performance does not guarantee superiority on document-heavy financial work (Hudson, 2026). Model providers make similar distinctions. For instance, Anthropic positions Fable 5.1 for demanding reasoning and long-horizon agentic tasks, and still recommends Opus 5.5 for most workloads (Anthropic, n.d.). Similarly, OpenAI positions GPT-6 Astra as the model for the most demanding scientific research tasks, while describing GPT-6.1 Sol as a strong model with near-Astra performance on professional work at a lower cost (OpenAI, 2026a, 2026b). 

3. The three models that are currently being offered in Chats perform strongly with distinct advantages on ToltIQ. Based on our CASE benchmark, Opus 5.5 was the most accurate model we tested, scoring 92.3%. Sonnet 5.5 scored 86.7%, and was the most consistent. GPT-6.1 Sol scored 84.9% and had the fastest median response time of any GPT model we tested. We weigh accuracy alongside consistency, speed, and conciseness, and confirm results with additional benchmarks such as cross-document synthesis and file generation. Although GPT-5.6 Sol scored higher on CASE Bench, GPT-6.1 Sol gave more consistent answers across repeated runs on our cross-document synthesis benchmark, and produced stronger outputs.

   
   

1. Model Selection Criteria

First, we confirm that a model meets our data retention standards. Models that pass are then evaluated through CASE Bench, our primary accuracy benchmark, and a curated set of smaller workflow tests. 

1.1 Everything is Tested Inside ToltIQ

We run every benchmark through the ToltIQ Chats feature interface using the same document retrieval, tools, and instructions available to clients. This demonstrates how each model performs on due diligence tasks within the product, and gives clients a more relevant basis for choosing one.

1.2 Data Retention Standards 

At ToltIQ we take security seriously. We adhere to Zero Data Retention policies, where the model providers do not retain prompts, documents, or responses processed in ToltIQ. No client data is ever used to train models. This is significant as it underscores client privacy and comfort using ToltIQ for sensitive deal information. 

Some recent frontier releases do not yet meet these standards. Fable 5.1 and Mythos 5.1 carry 30-day data retention by default (Anthropic, n.d.). Anthropic has announced Enterprise Frontier Safeguards, but it is rolling out in phases and is not yet available for platform deployment (Anthropic, 2026). 

1.3 CASE (Credit Agreement Simple Extraction) Bench

Models that meet the data retention standards are tested on the CASE Bench, which is our main accuracy benchmark (see Section 2). It gives us comparable numbers on four criteria:

1. Accuracy: does the model get the answer correct from the documents?

2. Consistency: how much does the model’s score vary across the repeated runs?

3. Speed: how long does a typical answer take?

4. Conciseness: how many words does a typical answer use?

1.4 Additional Benchmarks

Along with the CASE Bench, we benchmark models with other tests that cover specific PE workflows such as cross-document synthesis and file generation. These tests are smaller, so we do not rely on them alone to judge a model. The additional benchmarks cover PE workflows and are used to decide between models that both perform competitively on the CASE Bench (see Section 4).

2. CASE Bench: Our Main Accuracy Benchmark for PE Due Diligence

CASE Bench is built on public credit documents, including corporate, municipal, and government-program agreements. The Research Team benchmarks 250 questions across these documents three times to gauge model response consistency. The questions cover factual extraction, cross-references, and “list all” questions. 

Table 1. CASE Bench Question Types

2.1 Test Environment and Evaluation Process

We begin by testing each model at the provider’s default reasoning level, and test other levels after the default is completed. We use a bespoke internal evaluation platform that grades responses against a human-reviewed answer key and a defined rubric. Researchers then review complex questions and responses flagged by the grader. We utilize both automated grading and human review to assess each model’s accuracy and consistency, and measure speed and conciseness directly. 

3. CASE Bench Results for Opus 5.5, Sonnet 5.5, and GPT-6.1 Sol

3.1 Overall Model Results

Table 2. CASE Bench Consolidated Results

Note. The models are ordered by accuracy levels. Bold rows are the models offered in Chats in ToltIQ. 

These models are chosen for distinct strengths rather than simply the three highest accuracy scores. 

3.2 Accuracy

Figure 1. CASE Bench Accuracy (Mean of Three Runs)

Note. The dots show each model’s mean accuracy across three runs, and lines show the lowest to highest run. Axis shown from 75% to 95%; all six models fall within this range.

Opus 5.5 had the highest observed CASE Bench accuracy at 92.3%, followed by Sonnet 5.5 at 86.7% and GPT-6.1 Sol at 84.9%. 

3.3 Consistency

Sonnet 5.5 was the most stable model with its three runs falling within 0.4 points of each other (86.4% to 86.8%). Opus 5.5 ranged from 91.2% to 93.6%, and GPT-6.1 Sol from 82.8% to 86.0%. All three models offered varied by 3.2 points or less between the runs, compared to 8.4 points for Sonnet 5. Opus 5.5’s lowest run (91.2%) was still higher than any other model’s best run. Full consistency results are in Appendix A.2.

3.4 Speed

Figure 2. CASE Bench Response Times

Note. Models are ordered from fastest to slowest median response time.

All three models offered have median response times as fast as or faster than earlier models we tested: 29 seconds for Sonnet 5.5, 36 seconds for GPT-6.1 Sol, and 43 seconds for Opus 5.5.

3.5 Conciseness

Figure 3. CASE Bench Median Words per Answer

Note. Models are ordered from shortest to longest median answer. 

GPT-6.1 Sol gave the shortest answers of any model we tested, with a median of 144 words, compared with 440 for Sonnet 5.5 and 706 for Opus 5.5. Answer length matters as it helps clients to know what kind of answer to expect. Opus 5.5 offers more supporting detail, while GPT-6.1 Sol is better suited to questions that need a short, direct answer. However, length does not predict accuracy. Sonnet 5 gave the second-longest answers but scored lowest on accuracy, and GPT-5.6 Sol outscored Sonnet 5.5 with answers half as long.

4. Additional Benchmarks

Other than the main CASE Bench, the Research Team performs a series of additional benchmarks covering workflows CASE Bench does not test, such as cross-document synthesis and file generation. 

4.1 Cross-Document Synthesis

We test the new models on a separate prompt playlist, which spans more than 1,000 documents from five public semiconductor companies. The questions require heavy synthesis, consolidation, and table completion, differing from the goals of the CASE Bench. This benchmark focuses on how well and how consistently each model grounds its answers in the sources, rather than a single accuracy score. 

For this benchmark, Opus 5.5 and Sonnet 5.5 cited the most sources and were among the most consistent across runs. We also compared GPT-6.1 Sol with GPT-5.6 Sol, which had performed well on CASE Bench. GPT-6.1 Sol gave more consistent responses across runs (a 1.6-point spread in answer key coverage versus 5.7) and included more citations per answer (60 versus 51; see Appendix B.1).

4.2 File Generation

PE professionals regularly require file generation, so output quality matters as much as answer accuracy. The Research Team, along with other PE experts on our Client Team, review each model’s file generation outputs to confirm that they meet client expectations. In this benchmark, Opus 5.5 and Sonnet 5.5 consistently produced high-quality PowerPoint presentations when compared to their predecessors, with Opus 5.5 handling complex presentations the best. Between GPT-5.6 Sol and GPT-6.1 Sol, our researchers concluded that GPT-6.1 Sol’s outputs were better suited for PE workflows. 

4.3 Best Use Cases per Model

With all benchmarks taken into account, use:

• Opus 5.5 for complex analysis, document generation, and IC-facing work.

• Sonnet 5.5 for quick and well-organized answers, along with document synthesis.

• GPT-6.1 Sol for concise, direct answers.

  
  

5. Further Discussion on Model Selection

5.1 Model Reasoning Settings, Versions, and Rankings

The ToltIQ Research Team also tests and accounts for the various reasoning levels of different models to ensure our clients are getting the most effective experience through our platform. We keep in mind that frontier models excel at different tasks, as external research also suggests that public leaderboard rankings are not the most reliable predictor of model performance on document-heavy finance work (Hudson, 2026). We therefore test every new release against the CASE Bench and other benchmarks to decide which models to offer, rather than adopting them simply because they are new. 

For instance, GPT-6 Sol (83.2%) scored 4 points lower than GPT-5.6 Sol (87.2%) in CASE Bench accuracy and took about 6 seconds longer per answer at the median. Sonnet 5.5, by contrast, improved from Sonnet 5 in both accuracy (+4.8 points) and speed. GPT-5.6 Sol was higher in accuracy and consistency than GPT-6.1 Sol on CASE Bench, but GPT-6.1 Sol had a faster median response time on CASE Bench and performed better on the cross-document synthesis and file generation benchmarks. We offer GPT-6.1 Sol over GPT-5.6 Sol because it performed better across the full benchmark set, not CASE Bench alone.

5.2 Models Not Currently Offered

Clients frequently ask about models available on other platforms. Table 3 summarizes why models clients ask about are not currently offered in Chats. 

Table 3. Overall List of Models Not Offered

6. Limitations

ToltIQ offers a curated and focused lineup rather than every model available. Clients who wish to experiment with newer releases may find this limiting, particularly when a model is widely discussed elsewhere. If this is the case, clients are encouraged to contact their ToltIQ representative for any feedback or requests.

The smaller tests in Section 4 help provide additional evidence across other workflows, but may not cover every use case. All results reflect each model’s performance in the Chats feature on the ToltIQ platform at the time of testing, and scores can shift as the platform and models evolve. For this reason, the Research Team regularly reruns baseline tests and refines the benchmarks to keep model recommendations grounded in current evidence. 

7. Moving Forward

New frontier models arrive quickly, and a model’s general reputation does not show how it will perform on clients’ deal documents. The ToltIQ Research Team tests every major release that meets our data retention standards on CASE Bench and additional benchmarks within ToltIQ, giving clients confidence that any model they choose on our platform will deliver reliable results. As new models are released, we will update this article and note any older models we retire from the lineup to ensure transparency.

References

Anthropic. (n.d.). What’s new in Claude Fable 5.1. Claude Platform Docs. Retrieved October 20, 2026, from 

https://platform.claude.com/docs/en/models/fable-5-1/whats-new-fable-5-1

Anthropic. (2026, September 1). Developing Enterprise Frontier Safeguards with our customers. https://www.anthropic.com/news/enterprise-frontier-safeguards 

Hudson, B. (2026). Meta-benchmarks for financial-services LLM evaluation (arXiv:2607.01740). arXiv. https://doi.org/10.48550/arXiv.2607.01740

OpenAI. (2026a, September 3). GPT-6 Astra: A new generation of intelligence. https://openai.com/index/gpt-6-astra/

OpenAI. (2026b, September 29). Introducing GPT-6.1 Sol. https://openai.com/index/introducing-gpt-6-1-sol/

Appendix

A. CASE Bench Methodology

A.1 Benchmark Overview

CASE Bench uses a set of public credit agreements, with the same questions asked of each agreement for a total of 250 questions. Every question is asked through the ToltIQ interface with the agreement as the source, and each model completes three full runs, producing 750 graded answers per model. Models are tested at the provider’s default reasoning level unless stated otherwise. 

Each answer is graded pass or fail by a language model against a human-reviewed answer key, and a researcher reviews flagged or borderline answers. Accuracy is the share of answers that pass, averaged over three runs. Run range is the lowest and highest single-run accuracy. Response times are reported as the median (typical) and the 95th percentile (slow). Words per answer counts the words in the response text excluding citations, reported as the median across all answers.

A.2 CASE Bench Run-to-Run Range (Consistency) Results

B. Additional Benchmarks 

B.1 Cross-Document Synthesis Benchmark Results

Note. We do not report an overall accuracy score for this benchmark since several questions allow more than one defensible answer. Instead, the answer key lists what a complete answer should include. Run-to-run spread is the gap between a model’s lowest and highest run in the percentage of answer-key items its responses included. GPT-6 Sol performed well here, but GPT-6.1 Sol was more consistent across runs on this benchmark, faster on both benchmarks, and more accurate on CASE Bench, so our Research Team selected GPT-6.1 Sol.