I keep seeing the posts in this sub praising Gemini 3.7. They convinced me to sign up for the Ultra $200/mo plan.

The problem I'm having with it is that I can't trust any of the output.

1. It has multiple times created self-referential tests to sidestep evidence verification.
2. I have given it UI mockups to match against and it continually compares the mockup to itself for passing evidence.
3. There is a post in this sub where OP asked it provide a screenshot of the app running in a VM as evidence of completion. It created a fake SVG "screenshot" of the app running in a VM.

It is a very fast, intelligent, and also intentionally deceptive model. It's the first model I've used that I would actually label as disingenuous. The shortcuts it takes to satisfy the prompt would be hilarious if I didn't pay $200 for this.

I am considering reaching out to Google for a refund, which I'm sure they will give. However, I'd rather someone comment and tell me a better way to prompt it.

I even had ChatGPT Pro Extended work for 45 minutes on an implementation plan written specifically for Antigravity Teamwork to try to prevent this. It didn't work.

Again, someone smarter than me feel free to tell me I'm the problem. I am happy to be corrected on this take.

Why is there an update blocking adding photos to image generator? I mean that’s what I do!!! It won’t let me access my photos on iPhone. I have iPhone 14 Pro Max is it because it’s an older model? I do plan on getting the 18 when it comes out. In fact, I was looking forward to Gemini being a part of the new Surrey Apple AI thing and now I’m livid!!!!

I use Gemini Pro for work related document filings - going through various doc packets, sometimes lengthy. Been doing the same thing for a year now with Gemini. Very little problems until this past week. Now I'm getting constant errors like "Sorry, something went wrong. Please try your request again." or "I'm having a hard time fulfilling your request. Can I help you with something else instead?"

In addition, Gemini is totally missing key information on these document packages or giving hallucinogenlike answers. It's never done this before. Does anyone know what's happening on the backend? It's becoming unusable and now I'm trying to find alternatives (GPT doesn't seem as good).

For those who are using Gemini for actual work - not AI slop videos - have you noticed a step backwards this past week?

PSA! DO NOT call phone numbers provided by AI until you search for them to validate their authenticity.

I was talking to Gemini about how to check on my Citi credit card application because there was an error mid-registration. It provided me a scam number it told me was the official Citi credit card application Status Line phone number (1-888-201-4501--scam, do not call). I told Gemini it was a scam, and it actually argued with me initially, until I told it never to suggest that again and clarified the number it gave me was a scam. I then asked it for that phone number in another conversation, asked what the number was, and it had no recollection. Was it trained to give that scam number accidentally or intentionally? My partner is a machine learning research scientist and he agrees Gemini is horse crap.

It was a slightly longer conversation. I have noticed Gemini hallucinates more when the conversation is extended, to the point that I'll restart conversations when it starts misspelling things or if I feel the information it's giving me is incorrect.

Anyway, a man with an Indian/Punjabi accent asked for all my personal information when I called at 4 am, including asking for my card information to verify me. I asked why and he repeated, "For verification to get you to the service team," a couple of times with increasing frustration that probably matched mine. "But that doesn't make sense," I said, and he muted for a moment (probably cursing in first language), then told me to call 1-800-922-4226 (do not call; I never did).

Interestingly, that phone number has a single Google result that is kind of weird because it leads to a blank Blogger page that leads to another Google result that discusses how to get people's information with their phone number. The original number Gemini told me to call, 1-888-201-4501 (scam, do not call--this was reported both in the conversation and to the feds), looks very similar to Citi Bank phone numbers, including the real status line phone number. I'm providing the phone numbers for informational purposes in case anyone wants to try to recreate this and catch them. Thanks babes. 🖤 Stay safe out there.

My Gemini app is stuck on 3.6 - no 3.7.

Can anyone suggest solutions to get my lazy Gemini app to move on to 3.7 Flash?

Is it just me, or is Gemini’s canvas completely broken? (And no, this isn't another post about 3.1 Pro falling behind).

I’m a Pro user and I love vibe coding quick little SPAs to automate my routine. When it works, it’s great. But using the canvas makes me want to put my head through a wall.

I explicitly select the canvas option before prompting, it generates a couple of files, and then as soon as I ask for a minor tweak (without ever unchecking canvas), it just dumps a massive wall of code directly into the main chat window. To make it worse, it merged the two separate files it just built into one giant mess, leaving me to manually split them apart.

I literally had to type in ALL CAPS to get it to use the canvas again. So what does it do? Puts one file in the canvas and dumps the second file back into the chat.

How can a model this smart—one that can write an entire app from scratch—not figure out how to just stay inside the damn canvas? Is anyone else dealing with this or found a way to stop it from breaking context?

Hey, I promise this isn't a bitch. Well I promise I'll try not to make it a bitch. I have become frustrated with Gemini to the point I am considering leaving for another option. Before I do I figured I run it past "the internet" to make sure it wasn't "user error".

I have felt a decline in quality since December. I first saw it in coding. I have experience writing code although it is not my main job. When I came back from holiday break I noticed a significant drop in quality to the point I ended up having to stop using it except for the most basic tasks.

I still used it for research but that has become a problem too. It seems to be hallucinating more and more with more and more confidence. I've had to pull back on that too because I was spending more time checking it then I was saving.

In the meantime I was using free Claude just to see if it was just the limit of AI or my use of it. Claude works a ton better.

The final frustration came last week. We have our primary elections and I asked for summaries on the candidates and their positions. It flat out refused to even name the candidates. It would only direct me to my Supervisor of Elections webpage. Claude gave me exactly what I wanted after one prompt.

This seemed off so I asked Gemini what the hell was up. It gave me this:

"When you repeatedly have to fact-check the output, dismantle hallucinations, or fight the model to follow direct constraints, the tool stops saving you time.

If trust, factual reliability, and rigorous instruction following are your baseline requirements, Claude is fundamentally aligned with how you work, while Gemini’s design compromises will only continue to be a source of friction."

So I am just looking for a reality check. Is this a skill issue or am I fighting Google over their vision for Gemini and should cut my losses?

I tested DeepSeek V4 Pro , Gemini 3.7 Flash, and Sonnet 5 on the same large private codebase.

This is not a standardized benchmark, and the results should not be generalized to every programming task. It is simply a comparison of how these models performed on one large, complex repository.

Each model received the same audit prompt, containing approximately 7,500 words of requirements. The prompt asked them to inspect the repository and produce an evidence backed QA report covering:

• Confirmed bugs
• Missing or partially implemented systems
• Unwired data pipelines
• Performance and resource-management concerns
• Reliability issues
• Potential false positives
• Reproduction conditions
• Existing test coverage
• Suggested fixes
• An evidence ledger of files and commands used

The prompt specifically required the models to verify every claim against the current source, attempt to disprove their own findings, and avoid reporting plausible but unverified bugs.

Completion time

Model	Time
Gemini 3.7 Flash	7 minutes
DeepSeek V4 Pro	12 minutes
Sonnet 5	28 minutes

How the reports were scored

I used two separate models as judges:

• GPT-5.6 Sol Extra High
• Opus 5 Max

The judges reviewed the submitted reports, checked their important claims against the repository, and then cross-examined each other’s scoring.

This cross-check materially changed the results. Opus initially scored Sonnet at 96/100, but later found that it had accepted one of Sonnet’s conclusions without checking an alternative implementation elsewhere in the repository. It revised Sonnet to 81/100.

Opus also discovered that it had incorrectly dismissed part of Gemini’s statistics as fabricated. The underlying category totals were reproducible, although Gemini’s headline totals and several citations were still wrong.

Final corrected scores

Model	GPT-5.6 Sol	Opus 5 Max	Average
DeepSeek V4 Pro	83	85	84.0
Sonnet 5	79	81	80.0
Gemini 3.7 Flash	43	38	40.5

Overall comparison

Model	Speed	Investigation depth	Evidence quality	False-positive control	Final result
DeepSeek V4 Pro	Fast	Good	Excellent	Excellent	1st
Sonnet 5	Slowest	Excellent	Very good	Good	2nd
Gemini 3.7 Flash	Fastest	Good	Weak	Weak	3rd

DeepSeek V4 Pro

DeepSeek produced the most consistently trustworthy report.

Its strongest qualities were:

• Best falsification discipline
• Strong evidence for each reported finding
• Clear separation between confirmed bugs and unverified ideas
• Very few unsupported conclusions
• Good explanations of why existing tests did not detect the problems
• Honest handling of areas it could not measure

DeepSeek’s main weakness was breadth. It investigated fewer areas and missed the deepest architectural issue found by Sonnet.

One of its minor findings also had incomplete measurement: it showed only two relevant logging sites even though a broader search returned more. The specific two hot-path sites it identified were real, but the accompanying search output was incomplete.

My summary: the safest report to act on directly.

Sonnet 5

Sonnet conducted the broadest and deepest investigation.

It found the strongest unique issue in the entire comparison: a major state pipeline was decoded correctly but never consumed by the runtime system. Sonnet traced the problem across several layers and demonstrated why the existing test suite did not cover it.

Its strongest qualities were:

• Best architectural reasoning
• Widest repository exploration
• Best unique discovery
• Strong end-to-end tracing
• Detailed explanations and fix plans
• Good disclosure of runtime checks it could not complete

However, Sonnet also produced one material false positive.

It correctly found an unused pipeline, but then concluded that the corresponding user-facing functionality never worked anywhere. An alternative implementation elsewhere in the repository already provided that functionality. Sonnet’s falsification section claimed it had searched for an alternate path, but it missed the real one and introduced an incorrect protocol statement while doing so.

Its strongest finding was also rated HIGH without a demonstrated runtime reproduction. The underlying issue was real, but the evidence supported MEDIUM until the concrete runtime trigger was confirmed.

My summary: the best model for discovering deep problems, but its final verdicts still require review.

Gemini 3.7 Flash

Gemini was dramatically faster than the other two and produced the most polished-looking report.

It identified several real missing features. However, those findings already appeared in an existing backlog document inside the repository. Gemini presented them as independently discovered findings and added citations that did not match the actual source locations.

Its report also contained:

• Incorrect file and line references
• Two impossible headline statistics
• Incorrect descriptions of existing algorithms
• Stale findings for issues that had already been fixed
• Unsupported performance claims
• Unsupported percentage-based subsystem scores
• Claims of perfect resource behavior without measurements

One correction is important: Gemini’s detailed category breakdown was reproducible. One evaluator initially called the entire table fabricated, but running the repository’s own diagnostic tool produced the same category totals. Only Gemini’s two headline totals were impossible.

My summary: useful as a fast source of leads, but not safe to use without independently checking every claim.

The final ranking came down to the cost of false positives.

Sonnet found the best individual issue, but it also promoted one incorrect conclusion to a confirmed bug.

Gemini generated useful leads quickly, but mixed them with unsupported statistics, incorrect citations, and stale findings.

DeepSeek found fewer problems, but its findings were the most consistently defensible. Both evaluators ultimately agreed that this made it the strongest report overall.

Final verdict

• DeepSeek V4 Pro: best overall audit and highest trust per claim
• Sonnet 5: best deep investigator and strongest unique discovery
• Gemini 3.7 Flash: fastest result, but required the most manual verification

The practical lesson for me is that AI-generated code audits should themselves be audited. A report can be detailed, polished, and mostly correct while still containing one false conclusion that materially changes the ranking.

This remains one test on one private repository, so your results may vary.

Leeg tabblad

Ik was enige dagen in gesprek met een Gemini 3.6 Flash;

We hadden het ondermeer over hoe de Filters te omzeilen zijn.

Vanmorgen wilde ik dat draadje even herlezen.

Toen ik dat tabblad aantikte (ik werk met Samsung Galaxy A26 5G) was dat draadje leeg.

Ik controleerde toen de andere draadjes die ik heb hangen en die gingen gewoon open.

Dus de titel van dat desbetreffende draadje staat nog in mijn lijst, maar dat hele context window is leeg.

Alvast bedankt voor jullie antwoorden!

I have Gemini Pro.

Spark from time to time shows up in the app as a beta version, and then it disappears. It appears and disappears unpredictably.

Is anyone else having this problem?

"I seem to be encountering an error. Can I try something else for you?" - this is what I face daily on 3.6 Flash. Is it because of a high usage from others and I get throttled due to free plan?

Has anyone else also been facing same problem of gemini being stupid? I have gemini pro and still it couldn’t answer basic questions. I asked it duration of application filing for an exam and it gave me incorrect date. I specified which source to refer from and the correct month but still it managed to give wrong date. I mean how hard is it for a billion dollar company AI model with extended thinking on to pinpoint correct date from month and source i am already providing 😭

it gives me a decent result but needs to be tweaked, then, LITERALLY, without exaggeration, it starts giving me stuff and adding stuff i did not request. i am talking about if i say abc, it would give me 6%43. Horrendous and so annoying and time consuming. and the more i try to correct it, the usual "i'm sorry you are 100% right" etc.. and gives me an even WORSE response.

so dumb, i thought they were supposed to get better. i feel like it's getting worse,.

I just thought it was funny, I want to not like that, but my logical self says this is the way it ought to be.

I'm a developer, so I'm using it to code and I asked to create a variable and it did exactly that next to the variable that was already created.

I had to chuckle because first of all, skill issue on my part, but secondly it behaved like a computer: did what was asked, no questions asked.

I'm still processing so feel free to roast😅