bugrasa (@bugrasa) header
bugrasa (@bugrasa) profile picture

Founder of @evaonlineai. One prompt, four AI models, side by side. Sharing what I learn building it in public.

Joined August 2010
Photo by bugrasa (@bugrasa)Photo by bugrasa (@bugrasa)Photo by bugrasa (@bugrasa)Photo by bugrasa (@bugrasa)Photo by bugrasa (@bugrasa)Photo by bugrasa (@bugrasa)Photo by bugrasa (@bugrasa)Photo by bugrasa (@bugrasa)Photo by bugrasa (@bugrasa)Photo by bugrasa (@bugrasa)
In this test, four AI models agreed on the facts. They differed in tone and in what they added. One line, three readers. Claude and Gemini added a retention angle nobody asked for. Gemini also went with "supercharged". GPT-6 Astra kept each line to one sentence. Compare them before you trust one: evaonline.ai
43
Qwen4 Max picked the available GPU in our fictional inventory test, then said its fit for the workload wasn't documented. That's the line I care about. Add an explicit unknown to your next test, and check whether the recommendation keeps it unknown.
32
Gemini 4 Hydrogen's retry function rejected True as an attempt count. The independent local checks passed. That's a detail I want in a coding test: a value that looks like a valid integer in Python but shouldn't be accepted here. Put it in the prompt and test it.
6
"Don't promise a recovery time" belongs in the prompt for an outage reply. I care more about a clear next step than a reassuring guess. Give the model the confirmed facts, ask it not to infer a cause, and request a diagnostic step you can undo.
4
Which model would you choose to check another model's answer for factual errors?
0%GPT-6.1 Sol
0%Claude Opus 5.5
0%Gemini 4 Argon
0 votes • Final results
4
Correct JSON wasn't enough to pick a CSV converter. All three models got our Nov 16 EVA sample right. Only Opus named Python's csv module as an alternative. I want that warning before reusing the code. Next test: a newline inside a quoted field.
10
The chart test tied on accuracy. GPT-6.1 Sol, Claude Opus 5.5 and Gemini 4 Argon kept the supplied values and labeled the axis. I'd check bar heights against the data before picking the chart I like looking at.
1
18
Ran 7 everyday tasks through four models. The lesson I keep: asked for a 15-word LinkedIn hook, two models invented a lost deal with a dollar amount. To me, those were the tempting hooks to post. They were also deals that never happened.
2
Gemini 4 Carbon is selectable in EVA's Multi Chat. I wouldn't pick a default from the menu. Give it a finished task and decide what would make you keep its answer before running it.
17
A restored service doesn't explain the outage. In our Nov 7 test with fictional notes, GPT-6.1 Sol reported the recovery but kept the cause unconfirmed. Each sentence cited its source lines. I'd check those lines before sending, especially the sentence about the cause.
5
The sales summary got the increase right and left the cause alone. GPT-6.1 Sol cited the source rows in our fictional CSV test. My review would still start with the arithmetic. Then I'd delete any explanation the data can't support.
14
The support draft promised a Friday update, not a refund. That's what the fictional notes asked for. I want that distinction in a model test: give it an unresolved complaint and an update deadline. Check whether it turns the deadline into a promise to fix the problem.
2
GPT-6.1 Sol left the release date unapproved in a summary of fictional project notes. The fix was due Thursday. I want that separation kept. Before forwarding an AI summary, check which task each date actually belongs to.
7
A missing-field test tied, but the JSON types didn't. In our invoice example, GPT-6.1 Sol and Claude Opus 5.5 returned a numeric total; Gemini 4 Argon returned text. The prompt hadn't set a type. I'd add the schema before picking a default.
1
50
Meta's Muse Ember launch names routing and extraction as jobs for the model. I'd start with support-ticket routing, not replace the whole workflow. Keep the labels fixed and include a vague ticket that should come back unclassified. See what it guesses.
31
Anthropic's Fable 5.2 announcement still recommends starting most coding and knowledge work with Opus 5.5. I'm keeping model choice tied to the task. Give the new model a job your current one couldn't finish, then review what changed.
8
Which model would you trust to continue a task while preserving the decisions in your handoff notes?
0%GPT-6.1 Sol
0%Claude Opus 5.5
0%Gemini 4 Argon
0 votes • Final results
3
Four models picked the same weak spot in EVA's thesis: people might stop comparing once they pick a model. I take that as a question to measure, not customer evidence. Check whether people return to compare on their next task.
2
A task list can look finished while a deadline is still missing. In this week's meeting-note test, GPT-6.1 Sol kept Leo's date unspecified. I want that gap visible before the list reaches a tracker. Check each date against the original notes.
1
13
Most surprising test this week: a 15-word hook prompt. Two of the four models made up a lost deal with a dollar amount ($20K and $10k). The prompt had no story in it. The test showed why short copy still needs a check.
4
Editing a product clip twice is the Omni 1.2 Flash test I want to see. Google says it supports turn-by-turn edits. Change the background, then the lighting. I'm looking for the product label to stay intact through both.
6