Update (2025.5.7)

Future Gemini 2.5 Pro updates moved to my self-hosted Flarum forum post Gemini Updates Summary. For Gemini App, recommend asking in English—Chinese experience definitely isn’t as good as ChatGPT or Claude.

Yesterday news broke that Google One AI Premium student discount can be claimed without edu email verification. This morning checking news, saw Gemini upgraded again, from 0325 version to 0506 (Google I/O conference version). Meanwhile Cursor also launched a free one-year student plan. I just claimed this deal and am editing this blog post in Cursor editor, immersed in the joy of free AI perks, can’t help but marvel at AI’s rapid development.

Update (2025.4.19)

Google’s generous student giveaway official blog post: https://blog.google/products/gemini/google-one-ai-premium-students-free/

Important (2025.4.18)

Google with its deep pockets is amazing. US Edu email verification gets 15 months free. Direct freebie! via: https://gemini.google/students/

Enough said, Google is awesome, hope Gemini sweeps all LLMs sooner or later, dominating!

Update (2025.4.18)

This article will be used to record Gemini 2.5 series model update notes.

• Gemini 2.5 Flash early preview released

Google’s first hybrid reasoning model, developers can adjust model’s thinking depth via thinking_budget parameter.

Available in AI Studio, Vertex AI, and Gemini App. In Gemini App, 2.5 Flash replaced the original 2.0 Flash Thinking.

For AI Studio pricing strategy, can refer to ChatGPT o4-mini-high’s analysis [not necessarily accurate, users with payment cards should watch for preview model charges]. I usually experience with non-card-linked AI Studio accounts, afraid of charges, my card-linked AI Studio account only used when calling API for batch data processing.

I definitely won’t prioritize 2.5 Flash since Gemini 2.5 Pro is currently the strongest Gemini model.

via:

1. https://developers.googleblog.com/en/start-building-with-gemini-25-flash/

2. https://blog.google/products/gemini/gemini-2-5-flash-preview/

3. https://simonwillison.net/2025/Apr/17/start-building-with-gemini-25-flash/

Update (2025.4.17)

• Gemini 2.0 Flash and 2.5 Pro in Gemini App support LaTeX rendering

via: https://x.com/joshwoodward/status/1912300138074444095

• Gemini Live gradually rolling out to all Android users

via: https://x.com/GeminiApp/status/1912591827087315323

Update (2025.4.16)

Veo2 gradually rolling out to Advanced users in Gemini App, Veo2 fully launched in AI Studio.

via:

• https://developers.googleblog.com/en/veo-2-video-generation-now-generally-available/

• https://blog.google/products/gemini/video-generation/

While organizing emails this morning, found Google gave me another three months of Gemini Advanced membership. This old account already had tried Gemini Advanced membership before. Can only say Google’s products are really affordable. Though privacy is poor—free users’ chat data will be used to improve Gemini—it’s truly good value. Compared to OpenAI and Anthropic, it’s way too cost-effective.

Claude launched Research feature today. After seeing the demo and that new features no longer open to Pro users, I shook my head. Following OpenAI’s old path is digging their own grave. Hope Gemini kills both these companies soon!

Update (2025.4.15)

Gemini 2.5 Pro also has pain points. When asking reasoning models to write code comments, they modify the original code.

Update (2025.4.12)

Gemini Advanced’s 2.5 Pro probably isn’t a chat-optimized version. Highly recommend friends subscribe to Google One AI Premium—after experiencing this long, Gemini Advanced experience is much better than before.

For low-cost continuous Gemini Advanced use, can buy monthly throwaway accounts on Taobao or Xianyu. I see some L Site friends do this. My current Google One AI Premium plan is $9.9/month after US edu email verification. See Google One Student Discount. Recently lots of US edu email student discount AI activities: $1/month Claude membership, free 2 months ChatGPT Plus, free 2 months Super Grok. All 4 major US frontier AIs can be had cheaply.

Can compare using 2.5 Pro models from Gemini Advanced, AI Studio, and Vertex AI Studio. Like this afternoon I had 2.5 Pro from all three places write code comments [nearly a thousand lines, Python file downloaded directly from Colab]. Gemini Advanced performed best, Vertex AI output a ton then errored out, AI Studio output a ton then had code block formatting issues. Note: best reply source varies each time. Currently for code comments I’m leaning toward Gemini 2.5 Pro—OpenAI’s o3-mini-high feels lacking.

The 1M context window is unbeatable. Claude’s 200K or 500K, ChatGPT Plus users’ 32K, Pro users’ 128K—all little brothers compared to Gemini. Plus Gemini’s context decay resistance is also top-tier. True LLM King!!!!!

Update (2025.4.10)

Expected future AI competitive landscape: Google’s leading advantage will grow larger.

When OpenAI launches expensive Pro plans, Anthropic follows OpenAI’s path with expensive Max plans, while Google continues offering free access to the most powerful LLM services. Paid Gemini Advanced membership besides nearly unlimited usage also includes Google One and other services—the value is obvious.

Update (2025.4.9)

Gemini App Deep Research powered by the most powerful Gemini 2.5 Pro model, available to Gemini Advanced users.

via: https://blog.google/products/gemini/deep-research-gemini-2-5-pro-experimental/

Update (2025.4.8)

• Gemini Live further rolling out to Android users

Chinese voice options aren’t great, but Gemini 2.5 Pro model capability is outstanding.

via: https://blog.google/products/gemini/gemini-live-android-tips/

• Gemini 2.5 Pro coding ability is indeed outstanding

After seeing many netizens’ praise for Gemini 2.5 Pro and combining with my own feelings, Gemini 2.5 Pro’s coding expertise indeed trends toward surpassing Claude 3.7 Sonnet. Overall, Gemini 2.5 Pro is far ahead—not only sweeping various benchmarks but user experience is also top-notch.

Dual-wielding Gemini 2.5 Pro and Claude 3.7 Sonnet is definitely the optimal solution currently [Since Gemini 2.5 Pro is currently a reasoning model, reasoning models excel in: logic, code, math, so in other areas still can’t underestimate Claude 3.7 Sonnet. Additionally, Anthropic develops models based on real needs, not benchmarks, which causes public underestimation of Claude!]. The AI closest to AGI form in the future very likely will be Gemini.

Update (2025.4.5)

• OpenAI delays GPT-5 release

OpenAI delayed GPT-5 release, most likely because Gemini 2.5 Pro has reached GPT-5 level. Currently OpenAI’s integrated GPT-5 can’t crush all benchmarks, can’t beat Gemini 2.5 Pro.

Think about it—if you were Sam Altman, would you allow releasing a GPT-5 that isn’t far ahead in all aspects? I predict GPT-5 release timing depends on OpenAI actually developing a model that crushes all benchmarks, otherwise GPT-5 will forever be just a pie in the sky.

Continuing to release reasoning models is a sign of OpenAI barely surviving. Though the recent GPT-4o native text-to-image was indeed impressive, OpenAI’s iteration on basic non-reasoning models has been somewhat disappointing personally.

Developing models for benchmarks is a big taboo. I really appreciate Anthropic developing models for real needs—that’s the right approach to LLM development. Good user experience means no shortage of users. Even though Claude no longer tops various benchmarks, its good user experience always attracts me. Currently the better choice for Agent base models is also Claude.

Recently read many Gemini 2.5 Pro articles. Summary: Gemini 2.5 Pro is most cost-effective, arguably the current strongest model (LLM King), with reasonably controlled pricing.

Don’t think I’m always praising Claude—actually Claude isn’t omnipotent either. Current LLMs each have their strengths.

• Gemini 2.5 Pro launches paid preview version

Experimental model with ID gemini-2.5-pro-exp-03-25 continues free. Preview model with ID gemini-2.5-pro-preview-03-25 starts paid preview, more competitive pricing compared to OpenAI and Anthropic. See Simon Willison’s blog post—Gemini 2.5 Pro Preview pricing.

• Gemini 2.0 Pro no longer exists

via: https://x.com/OfficialLoganK/status/1908179750536827183

Update (2025.4.4)

My Gemini App also got Gemini Live Camera feature. Same old story—for Chinese voice see Doubao. Gemini Live Camera is better experienced in English—Chinese experience isn’t great. Currently Gemini App voice-to-Chinese seems unsupported, so don’t expect great experience with Live Camera. Fluent English expression is prerequisite for good Gemini Live Camera experience!

Update (2025.4.3)

• Gemini 2.0 Pro experimental model taken offline

Gemini 2.0 Pro experimental model in AI Studio is gone. According to L Site users, currently still kept in Vertex AI, but my testing found Vertex AI’s Gemini 2.0 Pro experimental version showed chain-of-thought, probably replaced by Gemini 2.5 Pro experimental version—name might temporarily remain.

via: https://ai.google.dev/gemini-api/docs/models#gemini-2.5-pro-exp-03-25

Short-lived Gemini 2.0 Pro experimental version. Future GPT-5, Gemini 2.5 Pro [probably releasing around late May at Google I/O developer conference, possibly OpenAI will again race with Google] will likely be models combining reasoning and classic LLM, no longer requiring manual switching. If encountering questions not needing reasoning, use classic LLM for reply; if encountering reasoning scenarios, enable chain-of-thought.

• Gemini 2.5 Pro VS Claude 3.7 Sonnet

Meanwhile, while organizing some bookmarked pages today, found a Hacker News post—Gemini 2.5 Pro vs. Claude 3.7 Sonnet: Coding Comparison. The linked article shows Gemini 2.5 Pro and Claude 3.7 Sonnet each have strengths, but overall Gemini 2.5 Pro is more cost-effective.

Too many post comments, so I had ChatGPT Deep Research it. Results at Halo Blog.

• Gemini Code

Found a Gemini Code project inspired by Claude Code: https://blossom-tarsier-434.notion.site/Gemini-Code-1c6c13716ff180db86a0c7f4b2da13ab

Update (2025.4.2)

Deep use of Gemini App—experience is indeed good. Code comments written are nice, daily Q&A is sufficient.

But recommend asking in English as much as possible, especially when Gemini needs to call search tools. I found when asking in Chinese involving search, Gemini often simulates searching but doesn’t actually search. Friends not proficient in English can pair with Immersive Translate extension: triple-space to convert questions to English + translate English webpage to Chinese.

Update (2025.4.1)

Right after OpenAI’s Sam announced GPT-4o native text-to-image open to all users, Google’s Sundar Pichai announced Gemini 2.5 Pro Canvas open to all users. via: https://x.com/sundarpichai/status/1906874604683137285

First half of year’s most intense LLM battle has begun. April, May, June will be interesting. I feel Google’s Gemini model will impact RAG research—context window is large and attention decay is far less than other models.

Update (2025.3.31)

Before sleeping last night, remembered a benchmark proving Gemini 2.5 Pro context capability. via: https://fiction.live/stories/Fiction-liveBench-Mar-25-2025/oQdzQvKHw8JyXbN87

Can see Gemini 2.5 Pro’s article analysis in Gemini App—Gemini 2.5 Pro maintains very high understanding accuracy in long context.

Introduction

At midnight Beijing time on March 26, 2025, Google released Gemini 2.5 Pro model¹. Only a little over a month since my last Gemini blog post—can’t help but marvel at LLM iteration speed.

Gemini 2.5 Pro leads by far on various leaderboards. After these days hearing influencers’ praise, netizens’ reviews, and my own experience, I personally feel this model indeed has strength. Can say Google’s leading advantage in the 2025 LLM battle is beginning to show.

Google took down gemini-2.0-pro-exp-02-05 model in AI Studio to push Gemini 2.5 Pro—currently only callable via API.

Today Google directly opened Gemini 2.5 Pro in Gemini App to all users.

Current differences between free users and Gemini Advanced users:

• Lacking Gemini 2.5 Pro Canvas feature
• Rate limited [e.g., Deep Research free users have 10/month quota]
• Context window limited

Google’s moves help acquire users faster while showcasing latest LLM progress.

Also freeloading users’ conversations to further improve Gemini model. When 2.5 Pro reaches GA (General Availability, official release), capability will probably level up again.

Personal Experience

Initially on the night of the 25th before sleeping, learned from overseas leak tweets that Gemini 2.5 Pro appeared in Gemini App.

Next morning, what first attracted me was actually OpenAI’s GPT-4o native text-to-image capability. After watching OpenAI’s Livestream, it indeed felt like the video said: Feel The AGI.

Can’t help but sigh: OpenAI is truly petty, always timing releases with Google, once again stealing some Gemini hype. GPT-4o’s native text-to-image is indeed impressive—everyone who used it says it’s good. Looking forward to May’s GPT-5 release vs Gemini 2.5 Pro official release showdown.

Initially, because Gemini 2.5 Pro is a reasoning model, I didn’t pay much attention. My daily reasoning model usage is limited: using OpenAI’s o3-mini-high for code comments; math-related questions, asking Yuanbao’s DeepSeek R1 [Thinking King—DeepSeek R1’s math ability is indeed impressive, more suitable for Chinese users, some math problems only DeepSeek R1 can answer accurately]; occasionally using Claude 3.7 Sonnet Thinking for problems Claude 3.7 Sonnet can’t solve.

Don’t underestimate Gemini 2.5 Pro in Gemini App—after all, App’s Gemini 2.5 Pro can use Browse, Youtube and other tools. AI Studio’s Grounding With Google Search is far inferior, even though Gemini App models may be chat-optimized versions with some poor UX aspects.

When I ask Gemini daily, I send questions to both Advanced and AI Studio, comparing and adopting the better answer.

Below shows Gemini 2.5 Pro model summarizing my Gemini 2.0 experience blog post. Gemini App model summarized the article via Browse tool, while AI Studio model lacking this tool failed to summarize.

If tools aren’t needed, can prioritize AI Studio models. Of course both places now offer Gemini 2.5 Pro free—can compare, not every question is answered better in AI Studio.

Gemini’s political sensitivity in Gemini App seems to be decreasing. Originally for sensitive topics like US President it would refuse to discuss. Now can get correct answers via external tool retrieval. Summarizing YouTube political videos—sometimes fails, sometimes succeeds.

Gemini 2.5 Pro’s knowledge cutoff reached January 2025—should be the most current knowledge cutoff of any LLM.

Other advantages: OCR, multimodal, longest context window, output token limits, etc.—see Simon Willison’s review article².

What I’ve Seen and Heard

KCORES LLM Arena claims Gemini 2.5 Pro is currently the best coding LLM.

Some foreigner had Gemini 2.5 Pro play Pokemon. Twitch livestream link: https://www.twitch.tv/gemini_plays_pokemon

Many LLM leaderboards currently show Gemini 2.5 Pro at the top.

Summary

For coding—who’s stronger between Gemini 2.5 Pro and Claude 3.7 Sonnet—only experiencing yourself gives answers. I personally think Claude 3.7 Sonnet might be better since I’ve trusted Claude since Claude 3.5 Sonnet. The only one letting me confidently Vibe Code, running LLM-generated code directly, is Claude—others fall somewhat short.

Once dependent on Claude, it’s really hard to break away. Though LLM development is rapid—yesterday’s darling might become today’s abandoned child—I’ll always reserve a spot for Claude. Won’t abandon it unless a truly Claude-replacing LLM appears. Gemini 2.5 Pro probably isn’t that replacement. In my mind, Claude’s coding and writing abilities are both number one.

After Gemini 2.5 Pro came out, my Grok3 usage dropped. Only use Grok3 when comparing Deep Search or linking X queries. Most other daily conversations prioritize Gemini 2.5 Pro.