We ran seven AI models from Anthropic, OpenAI and Google on 18 real analyst tasks, once in each lab's own harness and once in LinqAlpha's. Every model scored 8 to 21 points higher in ours, and two cheaper models working together beat the best single model at 25% lower cost.
| Model | Own lab → LinqAlpha | Cost |
|---|---|---|
| Claude Fable 5.1 | 70.1 → 84.4 | $4.38 |
| Gemini 4 Argon | 72.2 → 83.1 | $1.86 |
| Opus 5.5 | 68.8 → 81.4 | $1.82 |
| GPT-6 Astra | 60.0 → 80.6 | $3.85 |
| GPT-6.1 Sol | 57.0 → 74.1 | $0.72 |
| Gemini 3.8 Flash | 50.6 → 60.5 | $0.72 |
| Gemini 3.1 Pro | 33.6 → 41.7 | $1.08 |
| LinqAlpha System | 89.4 | $3.28 |
I. The wrong first question
The question clients ask us most is which AI model they should use. Often it comes in another form: why not just give our analysts Claude or ChatGPT directly?
It is fair to ask. New models come out every month, and the rankings change almost as often. A choice made this month can be out of date by the next. So we tested what actually drives the quality of an answer: the model, or what surrounds it.
By "system" we mean everything around the model: the data foundation, the domain expertise, the checks, and how models are combined.
II. Two harnesses, one model
A harness is everything around an AI model that lets it do real work: the data it can reach, the tools it can use, the instructions it follows and the checks on what it produces.
Every AI lab now ships its own harness: Claude Code from Anthropic, Codex from OpenAI and Antigravity from Google. They are strong, general-purpose tools, built by the people who made the models. Think of a brilliant new analyst with a laptop and a web browser.
A laptop and a browser
- DataWhatever it can find on the open web
- Know-howGeneral instructions, written for any kind of task
- ChecksNone specific to financial research
A research desk and a senior colleague
- Data foundationA security master that ties every company, security and identifier together, with licensed filings, transcripts, fundamentals, estimates and prices
- Domain expertiseWhich data to use for each kind of task, and how to read it
- ChecksOn every answer before it reaches the user
Only the harness changes between the two runs.
The research desk is the data foundation. A security master ties every company, security and identifier together, so a question about TSMC finds the Taiwan-listed company's filings and calls, not just its US-listed shares.
The senior colleague is domain expertise. Which fiscal year does "FY25" mean for this company? Is a segment's "operating income" really that, or a contribution before corporate costs? Should a bank be judged on interest margin and capital, and a software company on customer retention? The harness carries these judgments for each kind of task.
III. The test: real work, not quiz questions
Most AI benchmarks are lists of questions written for the test. Ours come from what clients actually ask. In part 1 we sorted 33,453 anonymized analyst questions into the kinds of work that keep coming back, and picked 18 of them for this test.
Each task comes with a checklist of what a good answer must contain. Three AI models act as judges, one each from Anthropic, OpenAI and Google, and each grades every answer against the checklist. The score is their average, out of 100.
IV. Every model does better in LinqAlpha
All seven models scored higher in LinqAlpha's harness, by 8 to 21 points. In their own labs' harnesses, none scored above 72.2. In LinqAlpha's, five of the seven did.
| Model | Own lab | LinqAlpha | Gain |
|---|---|---|---|
| GPT-6 Astra | 60.0 | 80.6 | +20.6 |
| GPT-6.1 Sol | 57.0 | 74.1 | +17.1 |
| Claude Fable 5.1 | 70.1 | 84.4 | +14.3 |
| Opus 5.5 | 68.8 | 81.4 | +12.6 |
| Gemini 4 Argon | 72.2 | 83.1 | +10.9 |
| Gemini 3.8 Flash | 50.6 | 60.5 | +9.9 |
| Gemini 3.1 Pro | 33.6 | 41.7 | +8.1 |
OpenAI's GPT models gained the most: GPT-6 Astra rose from 60.0 to 80.6. Google's Gemini 3.1 Pro gained the least and finished last either way.
V. A cheap model beats an expensive one
GPT-6.1 Sol, one of the cheapest models we tested, scored 74.1 in LinqAlpha's harness at $0.72 a task. Claude Fable 5.1, the most expensive, scored 70.1 in Claude Code at $5.50 a task.
The cheaper model, set up properly, gave better answers for about one eighth of the price.
LinqAlpha's harness does not always cost more, either. Fable was cheaper inside it ($4.38 against $5.50), and Opus 5.5 cost the same.
VI. Two models beat the best single model
The best single model in LinqAlpha's harness was Claude Fable 5.1, at 84.4. We wanted to know whether we could beat it without paying for a more expensive model.
So we gave each task to two leading models at once, each cheaper than Fable. A final step merges their drafts into one answer, keeping the evidence and numbers from both. We call this the LinqAlpha System.
Across two full runs of the test, it scored 89.4 at $3.28 a task. Fable alone scored 84.4 at $4.38. That is 5 points higher and 25% cheaper. It is a little slower: 331 seconds per task against 295.
Is it just writing longer answers?
A merged answer might score higher just by being longer. To rule that out, we capped it at Fable's length. It still scored 88.1, 3.7 points ahead of Fable, at $3.18 a task.
Do the judges favor their own lab?
Our judges come from the same labs as the models we tested, so we looked at each judge separately. All three preferred the LinqAlpha System. The OpenAI judge gave the biggest lead, 7.0 points. The Anthropic judge, whose lab makes Fable, still put the LinqAlpha System 6.4 points ahead. The Google judge showed the smallest lead, 1.8 points, because it tends to score every answer highly.
| Judge | Fable alone | System | Lead |
|---|---|---|---|
| GPT-6 Astra judge | 77.6 | 84.6 | +7.0 |
| Fable 5.1 judge | 83.1 | 89.5 | +6.4 |
| Gemini 3.1 Pro judge | 92.4 | 94.2 | +1.8 |
VII. Why the system makes the difference
We have not yet measured how much each part of the harness adds. But the results point to three things.
Better data, read the right way
In its own lab's harness, a model works from the open web and has to decide on its own which company a number belongs to and which numbers to trust. In LinqAlpha, it also has licensed data, tied to the right company, and knows which figures matter for the task.
More of the work gets done
In Codex, OpenAI's own harness, the GPT models spent about half as much per task and lost the most points, a sign that they did less research before answering.
Two drafts cover more ground
The LinqAlpha System gained about 7 points on reports, briefings and explanations, and under 1 point on screens and lookups, where there is one right answer (Fig. 6). Merging keeps the evidence from both drafts, so fewer checklist items are missed.
VIII. What this means if you use AI for research
Choosing a model matters less than how you set it up. A leading model in its own lab's harness is a capable generalist. Give it the data foundation and the domain expertise an analyst relies on, and the same model does the work better, often for less.
It also means you do not have to bet on one lab. When Google opened Gemini 4 Argon to all developers on October 13, it was running inside LinqAlpha within days. Clients can use the strongest new model without changing how they work.