I have been using Gemini 2.5 Pro for article editing and research work, primarily utilizing its ability to generate prompts tailored to the specific themes of each research topic or article, aiding in both research and article polishing.
I have been using Gemini 2.5 Pro for this work for two months, from June to the present.
My workflow is as follows:
I write a paragraph outlining my understanding of the research topic or article, then provide Gemini with a main prompt that includes a segment for generating research prompts and another for article polishing prompts, instructing it to specialize the corresponding System Prompt; subsequently, I use these two System Prompts to drive Gemini 2.5 Pro for research and polishing tasks respectively.
Issues:
Recently, Gemini 2.5 Pro has been prone to forgetting content, performing poorly with lengthy prompts, and struggling to understand the preceding content beyond 30k characters. (This is even with Media resolution set to Medium.)
Editing quality has deteriorated significantly. My instructions for the large model were to âpreserve the original text as much as possible while ensuring logical coherence in the article.â However, the results showed that large sections of arguments and points were deleted from the article.
The large modelâs thinking and response times have significantly decreased. This may be related to the modelâs poor performance.
I can understand Googleâs generosity in the face of increasing pressure from its growing global user base, and the measures it has taken in response to this pressure, such as halving its quota.
However, I really miss the excellent text processing capabilities of Gemini 2.5 Pro, the best large model.
In fact, as a user with multiple accounts and a loyal Google member, I donât mind Google charging a reasonable fee for its services.
Thank you for your feedback. We appreciate you taking the time to share your thoughts. To help us investigate the issue effectively, any additional details you can provide would be very helpful.
Could you share a complete, self-contained prompt that previously yielded a high-quality result but now produces a poor one? Ideally, please provide both the old response (if saved) and the new response.
Regarding the âforgetting contextâ issue, could you provide an example of a prompt and pinpoint the specific part of the context the model appears to be ignoring? What was the approximate character or token count of that prompt?
Lastly, for the slower response times, could you give us an estimate of the change?
YES, those are my observations as well. Gemini 2.5 Pro and, in parts, also 2.5 Flash have been having massive quality problems for some time now. Initially, Gemini 2.5 Pro was great, but currently, it is largely unusable.
Apart from the numerous API errors, Gemini 2.5 very often loses context within the conversation when using the WebApp or AI Studio, like a person with acute dementia. For instance, Gemini often no longer knows what was discussed in a previous message or what it was actually supposed to do. Very often, instructions that worked perfectly a few days ago are no longer followed. Even with simple instructions, Gemini now fails miserably.
Itâs no secret that Google has been having acute performance problems with Gemini for a while. You can read about that everywhere. But for a few days now, it has been a complete disaster. Gemini is no longer a help.
Either the development team made a mistake during optimization and unintentionally messed up the originally good performance and quality, or it is a corporate decision to intentionally and massively reduce performance to save costs. I would bet on the latter, knowing Google.
In any case, if this pathetic state of affairs doesnât change, I will turn my back on Google Gemini. I still donât understand why you would turn what was formerly the best model into such a terrible one.
Apart from that, Google doesnât give a shit about complaints and reports of acute problems. Therefore, we can write whatever we want here; it will neither be read by those responsible, nor will they act on itâŠmy opinionâŠ
As shown in the figure, for example, when using the Google search function, it was originally possible to search more than 50 web pages, but now only about 5 pages can be searched.
The issue I have repeatedly raised regarding the infrequent use of Google Search is once again evident in the latest model.
The number of characters in the response now seriously fails to meet the requirements. I requested 3,000 characters or more, but only 1,000-2,000 characters are output.
Specifically, when editing articles, the most common strategy is to delete content rather than improve it
In addition, I also use Gemini 2.5 Pro to provide creative ideas for my AI image creation.
I trained Gemini 2.5 Pro with some books and set some rules. I can clearly feel that after Google announced that it would cut the quota in half, Geminiâs performance declined significantly. Not only did it forget the rules I had trained it with, but the depth and breadth of its creativity and logic were also greatly impaired.
I asked Gemini 2.5 Pro whether there were any factual issues with the legal content described in this image, but it told me there were almost no issues. As shown in the figure
However, in fact,
The Securities and Futures Ordinance (SFO, Cap. 571)
1) The â2023 Revised Editionâ does not directly correspond to the virtual asset licensing system. The license for virtual asset trading platforms (VATP) under the revised Anti-Money Laundering and Counter-Terrorist Financing Ordinance (AMLO, Cap. 615) was implemented in June 2023, not the SFO revised edition.
2) Section 103 essentially prohibits âunauthorized investment advertising,â not âthe unauthorized issuance of prospectuses to the public.â The maximum penalty is HK$500,000 and three years' imprisonment, with additional daily fines for continued violations, not âHK$10 million/10 years.â
3) The criminal liability for âconducting business without an SFC licenseâ is stipulated in Section 114(8), with a maximum fine of HK$5 million and seven years' imprisonment; it is not covered by Section 116.
4) Including security tokens under the regulation of âcollective investment schemesâ â is an inaccurate statement. The SFO's definition of âsecurities/collective investment schemesâ is technically precise and does not include a specific amendment in 2023 to âinclude security tokens.â Whether a scheme falls under regulation depends on its specific structure and offering method, not on the addition of a âtoken clause." (Summarized from the SFO framework and SFC's public explanations)
4. Project Ensemble Sandbox Project Guidelines (HKMA)
1) The timeline is incorrect. The HKMA announced Project Ensemble in March 2024 and officially launched the Sandbox on August 28, 2024; there is no âguidelineâ document dated âNovember 2023.â
2) The so-called âSection 5.3/Section 7.2â terms in the table, such as âquarterly submission of risk control reportsâ and â24-month testing period,â have no official source for their section numbers or fixed deadlines. The HKMA's public documents are press releases and speeches, which are positioned as testing interoperability and four key themes, and do not contain the aforementioned terms.
5. Circular on the Sale and Distribution of Tokenized Products (HKMA, February 20, 2024)
1) The document and date are authentic.
2) The scope is misrepresented. This circular applies only to tokenized products and tokenized deposits that are not regulated by the Securities and Futures Ordinance (SFO), and explicitly does not apply to stablecoins or tokenized securities regulated by the SFO.
3) Due diligence and disclosure, risk management, customer suitability, and AML/KYC obligations are indeed regulatory requirements. However, the circular does not require a âblanket review of the issuer's solvency and liquidity for all token products,â but rather requires thorough due diligence and information disclosure on the structure and risks of the issuer and its key third parties.
Even the above content was told to me by GPT-5.
Gemini 2.5 Pro now requires the deep search function to be enabled in order to significantly reduce such errors.
I am very concerned about the current accuracy of Gemini 2.5 Pro. Because of this error, I may not be able to create anything with Gemini for the time being, whether it be articles, code, or images.
Geminiâs performance is now back to where it was three months ago.
I am having the same experience. I use it a lot for coding PHP, JS and SQL. A new fresh chat with a well crafted prompt is initially going quite well. But there are always inconsistencies in the code, typos even(!). The real problem starts when there is a bug or a feature hasnât been implemented well. Also when I make manual edits to code and re-upload it so sync, Gemini loses it 8 out of 10 times.
It is specifically very bad at debugging and bug fixing. It keeps forgetting what it made, it delivers incomplete files, doesnât follow instructions (at all), it sticks to rules for 10 minutes and after that itâs wild west again. It hallucinates a lot in a very short period of time. It just canât listen and follow guidelines or rules. Very frustrating.
Compared to other models and as a heavy user, I find Gemini an extremely frustrating experience. I am using Gemini 2.5 Pro on an Ultra subscription.
I have been using Gemini for several months. I was very impressed with its capabilities until roughly mid-August, when I noticed a drastic deterioration in performance.
While my workflow for codeâwhich involves small, granular requestsâstill works well, the performance for tasks requiring large text context has degraded from being the best among competitors to almost unusable.
Specifically, when editing text in the canvas and making a request with multiple points, Gemini frequently only executes the first point, ignoring the others. Even the executed edit is often poor in quality and distorts many of the requested details.
Furthermore, I have encountered issues with the user interface (UI):
Canvas Update Failure: After working for a period, edits acknowledged by the chat (e.g., âexecuted editâ) are not reflected in the canvas. To resolve this, I usually have to create a new chat, provide all the necessary context again, and start the editing process from scratch.
Highlighting Failure: Sometimes, even when edits are processed correctly, the interface fails to highlight the updated parts of the text.
At this point I cancel my subscription as any other competitor model provides significantly better results for me. Just letting you know that there are problems still present.
1 is for coding. In the last few weeks, probably since the beginning of September, I have found geminiâs code quality to be really bad, even when using gemini-2.5-pro, it doesnât check for linter errors after coding, or is unable to fix simple errors like wrong import path.
2 is that I use gemini-2.0-flash in some RPG/Storytelling webapps. Since the beginning of September, the metrics show that the average usage frequency has decreased by 10% across all apps. I havenât found out exactly where the error is, but it seems like there are some small changes that affect the role-playing experience.
These customer services arenât going to do anything. Theyâll just tell you âwe are looking into itâ or âcan you give us more evidence,â blah blah.
They know well themselves that theyâre using quantized models or smaller models when usage is high. But the part I hate the most is they never admit, so you might spend hours on things and get nothing useful, still wondering if something went wrong with my prompt or script.
