Model behavior in Findit6Agent

Anthropic / Claude

Claude was the first model family I used while developing Agent Mode, and for some time it was effectively the only one. The differences between the Claude models became quite clear during testing. Haiku can operate Findit6 reliably and usually produces solid results, but its research tends to remain comparatively simple and straightforward.

Sonnet is considerably stronger. Depending on the task - a simple knowledge question or an explicitly requested comprehensive report - it can vary its research depth quite substantially. It plans searches well, follows additional research branches when necessary and usually structures the final answer very cleanly. Opus pushes this further, especially in research depth and in the quality and language of the final synthesis. It is the Claude model I would choose when the result matters more than shaving a few minutes or cents off the run.

In my typical test questions, which often contained several related information needs, Haiku usually finished within a few minutes and was the least expensive option. Sonnet tended to take somewhat longer, but API costs were still usually only a few dozen cents. Opus generally took a little more time again and could cost up to roughly twice as much, but in my tests even fairly substantial research runs usually remained below about €0.50.

xAI / Grok - the truffle pig

In our tests Grok behaved a little like a truffle pig. Whether the task was a factual knowledge question or a more statistical investigation across many files, it tended to reason for a while, follow side branches and keep searching until it genuinely seemed convinced that it had found the important material. The resulting summaries were usually excellent and often unusually complete.

That thoroughness has a price in both time and API cost. Our recurring test question - essentially “Who is Axel F., what have I had to do with him, and what does he occasionally have to do with ovens?” - could cost around one euro in a deep Grok run. But the answer then really did contain the professional biography, the different aspects of our cooperation and the very specific oven story in detail.

If you want something simpler, tell Grok explicitly. “Only the basic facts, no detailed research” makes a real difference. Left to itself, Grok tends to assume that finding one correct answer is not the same thing as finishing the research.

Google / Gemini

Gemini Flash has been one of the pleasant surprises. It is fast, relatively inexpensive and has handled the Agent Mode protocol and iterative search logic much better than I would have expected from a model optimized so strongly for speed.

It does not usually have Grok's appetite for digging through every possible side branch, but for many normal research jobs that is an advantage rather than a weakness. It tends to move quickly from search to evidence to answer and can therefore be a very attractive everyday choice when cost and turnaround time matter.

For extremely difficult, ambiguous or forensic research I would still tend to choose one of the heavier models. For routine Agent Mode work, however, Gemini Flash has repeatedly produced results good enough that I am quite happy to keep it in the normal model list.

Perplexity

Perplexity has generally behaved more pragmatically than Grok. It likes concrete evidence, tends to get to useful source documents quickly and has produced strong answers at comparatively low cost.

One early test exposed a useful weakness in our own Agent briefing: after receiving a hitlist that was too broad, Perplexity correctly stated that the list needed further refinement - and then stopped, waiting for the user. We changed the briefing to make the intended behavior explicit: decide on the next research action and execute it without asking for permission. In later runs Perplexity then refined the population autonomously and continued normally.

I like it particularly for straightforward archive research where I want a capable model that does not automatically turn every question into an expedition. It has also been refreshingly sober when asked afterwards to evaluate Findit6Agent itself.

But Perplexity in Guided Mode – Bertie Botts Every Flavour Queries. In my tests, Perplexity occasionally followed very explicit transformation instructions only approximately — for example by altering a required marker sequence or changing a prescribed logical operator. Most runs were correct, some almost perfect...But such small deviations can be enough to break a tightly controlled Guided Mode search pipeline. For that reason, Perplexity currently works very well for me in Agent Mode, but I do not recommend it for Guided Mode.

Mistral

I had actually hoped that Mistral, as a major European AI provider, would play an equal role alongside Claude, Grok, Gemini and the others in Agent Mode. Unfortunately, that was not what I observed in my tests. Mistral was the only large model I tested that repeatedly formulated the deterministic Findit6 search queries noticeably less well than the others and sometimes lost the thread during more complex multi-step research.

This reflects my experience with the models available in August 2026 and may well change with future versions. The API integration itself works perfectly well; the limitation I observed was the model's autonomous research behavior, not technical compatibility. For the time being I therefore cannot recommend Mistral for Agent Mode.

In Guided Mode, however, the picture is completely different. Mistral Medium 3.5 has worked extremely well there. The controlled workflow plays to its strengths: it follows the individual instructions reliably, handles the different reasoning settings correctly and produces rich, high-quality results. In my tests it was clearly more capable than the smaller local models while still being reasonably fast and inexpensive. At the moment, Mistral is therefore one of my preferred external models for Guided Mode

Mistral-hosted / Z.ai GLM

Mistral also hosts z.ai-glm-latest, a model developed in China by Z.ai rather than by Mistral itself. In my initial Agent Mode tests, it performed noticeably better at autonomous archive research than the Mistral-developed models I had previously tried. It planned sensible searches, selected relevant original documents and produced a detailed, well-structured report.

In that first comparison, the results were broadly comparable to those I have seen from Claude Sonnet, GPT Sol and Grok, while the research was relatively fast. API pricing appears attractive, although I have not yet checked the actual cost of this particular run. This is still an early impression, not a thorough model evaluation. Nevertheless, z.ai-glm-latest is already a viable choice for Agent Mode.

OpenAI

In my tests, the OpenAI models generally behaved like competent professional assistants: careful, structured and good at matching the amount of research effort to the task.

Luna was already surprisingly capable on normal Knowledge questions. It worked efficiently and produced a strong answer from a comparatively small number of well-chosen documents, although one test showed a minor protocol-following hiccup. But produced good results for sometimes less than 10 Eurocent.

Sol is a different animal. On a deliberately difficult archive-wide counting question, it spent more than twenty minutes refining the research, separating clear cases from ambiguous variants and eventually identified 56 indisputable productions plus four borderline cases that made a total of 60 defensible. In that kind of task it behaved almost like Grok's “truffle pig” - but with a meticulous accountant checking the final count afterwards.

That depth comes at a price. The Sol run cost about €2.75, roughly the same order as one of our most expensive Grok research runs. For difficult research where completeness matters more than speed or cost, however, the result was excellent.

Groq, STACKIT, Scaleway and other OpenAI-compatible providers

Findit6Agent includes a general OpenAI-compatible provider path. This makes it possible to connect services such as Groq, STACKIT, Scaleway and others without writing a dedicated provider implementation for every API.

The important warning is that these services usually offer many different models. I have not tested every one of them, and I would not expect most smaller models in those catalogues to be suitable for autonomous Agent Mode research. Correct API responses and valid JSON are only the beginning; a useful Agent also needs strong instruction following, planning, source evaluation, context handling and the ability to continue a multi-step investigation without drifting or stopping too early.

So please read “OpenAI-compatible” literally: the interface is compatible. It is not a blanket quality certificate for every model behind that interface. If you experiment with additional models, Guided Mode is often the safer place to start.