2 days ago
8
Follow ZDNET: Add america arsenic a preferred source on Google.
ZDNET's cardinal takeaways
- Different AI models triumph astatine images, coding, and research.
- App integrations often adhd costly AI subscription layers.
- Obsessing implicit exemplary mentation matters little than workflow.
The gait of alteration successful the burgeoning generative AI satellite is blisteringly fast. It's often hard to support up with everything, adjacent if it's your full-time job. Readers archer maine that 1 country they find peculiarly confusing is the wide array of poorly-named AI models. What successful the heck is the quality betwixt GPT-5.1, Opus 4.5, Gemini 3, etc.? And wherefore would you usage 1 implicit the other?
Also: Claude Code made an astonishing $1B successful 6 months - and my ain AI-coded iPhone app shows why
To beryllium honest, trying to afloat recognize the elaborate differences betwixt each of the AI models volition nonstop astir anyone screaming into the woods. But it's reasonably easy, particularly with examples, to recognize which models to take for antithetic tasks. That's what we're going to bash successful this article.
Keep successful caput that apps and AI models are antithetic beasts. The exemplary is the underlying AI motor that does the quality processing. The exertion is the instrumentality you use. Here's a diagram, created successful Google's Nano Banana Pro, that gets it mostly right.
But successful the aforesaid mode that antithetic vehicles usage antithetic kinds and brands of engines, antithetic applications usage antithetic models. For example, I gave the nonstop aforesaid punctual that produced the supra diagram to ChatGPT's representation generator (on the left) and Midjourney's representation generator (on the right), and we got these results:
As you tin see, ChatGPT's effect is elemental and reasonably clear, whereas Midjourney went some overboard portion besides simultaneously not really pursuing the prompt. Midjourney is large for conceptual images (like my bunny extracurricular a edifice instauration shown successful this article), but it's reasonably unspeakable astatine diagrams.
Also: Gemini vs. Copilot: I tested the AI tools connected 7 mundane tasks, and it wasn't adjacent close
That said, ever support successful caput my person Vinnie Moscaritolo's description: "AI isn't the sharpest instrumentality successful the box. Sometimes it gets it perfect, and sometimes it dives consecutive into the rabbit spread of anserine similar it packed a luncheon for the trip." Rabbit spread of stupid. Gotta retrieve that phrase.
(Disclosure: Ziff Davis, ZDNET's genitor company, filed an April 2025 suit against OpenAI, alleging it infringed Ziff Davis copyrights successful grooming and operating its AI systems.)
The cost/profit origin
Most AI companies merchantability their models some arsenic APIs (which programmers tin incorporated into their applications) and arsenic chatbot applications. Keep successful caput that an AI chatbot isn't an AI connected its own. An AI chatbot similar ChatGPT is truly a chat exertion (like a messaging app oregon Slack) that calls into an AI exemplary for quality work.
Also: The champion AI chatbots of 2025: I tested ChatGPT, Copilot, and others to find the apical tools now
While determination are immoderate really bully escaped AI chatbots, erstwhile you commencement utilizing them a lot, you're apt to bargain into 1 oregon much of the astir $20/month plans each offers.
You've astir apt noticed however galore bundle vendors are present including AI capabilities. I've been a QuickBooks Online idiosyncratic for galore years, but present QuickBooks is each astir pitching "upgrades" to merchantability you AI cause subscriptions.
Some vendors let you to usage the AI capabilities from the paid chatbot plans you're already paying for wrong their apps. But astir exertion vendors privation that further gross stream. Even if you already wage for AI models similar GPT-5.1 oregon Opus 4.5 done your chatbot plan, astir bundle vendors incorporating AI volition complaint you their ain interest for entree to those models.
Also: 4 ways I prevention wealth connected my favourite AI instrumentality subscriptions - and you tin too
How you determine what to usage should beryllium based connected what you need. To assistance retired with that decision, I'll spell implicit the models I usage for galore of the antithetic tasks I bash for work. Keep successful caput that I seldom take the exemplary first. I take the task and application, and the exemplary usually comes arsenic portion of that application.
Creating explainers
- Models used: Gemini 3 and Gemini 2.5
A cleanable illustration of this is Google's NotebookLM, a instrumentality I added to my kit astir six months agone and present usage astatine slightest weekly.
NotebookLM is simply a notebook instrumentality integrated with a connection exemplary (hence, NotebookLM). But I don't usage it arsenic a notebook. While immoderate folks similar that capability, NotebookLM has a superpower: it creates audio explainers based connected the root worldly it is provided with.
Also: I recovered an open-source NotebookLM alternate that's powerful, backstage - and free
You tin take whether you privation consecutive audio oregon an audio way accompanied by slides. Mostly, I conscionable spell for the consecutive audio. Whenever I get a peculiarly dense document, property release, oregon method paper, I provender it into NotebookLM. About 15 minutes later, the app produces a treatment exploring the cardinal points and main issues of immoderate I've fed to it.
To beryllium clear, I ne'er usage that output straight successful my work. But it's a large mode to get immoderate speedy triangulation connected the large issues related to immoderate I'm moving on.
Google doesn't explicitly picture the exemplary underlying NotebookLM, but it's surely a Gemini variant. Since Gemini 3 conscionable came out, astir of the enactment I've done with NotebookLM was apt generated with Gemini 2.5.
Identifying keywords
- Models used: GPT-4o, GPT-5, GPT-5.1
Notice that I listed 3 models above. That's due to the fact that OpenAI updated them. As they were updated, the exemplary I utilized automatically upgraded arsenic well. That's wherefore I don't privation you to obsess connected exemplary versions. If you work this successful 3 months, each the models volition beryllium different.
In this case, I'm utilizing a self-hosted web archiving exertion called Karakeep. Karakeep replaced Pocket arsenic my mode of storing articles I've recovered for aboriginal usage successful research. One happening that Karakeep does that Pocket ne'er could is make truly large automatic keywords.
Karakeep does this utilizing the API from OpenAI. Because the self-hosted Karakeep is unfastened root and doesn't impact immoderate subscription fee, its doesn't instrumentality its AI taxation retired of fees. Instead, I usage an API cardinal provided by OpenAI, and wage OpenAI based connected my usage.
I person 24,594 items indexed successful Karakeep. Moving my listings from Pocket to Karakeep took astir 2 months, due to the fact that Karakeep had to rescan each article. It besides called retired to OpenAI truthful the AI could make keywords. That process outgo maine astir $40.
Since the archetypal conversion, I person been billed $5 each mates of months. Well worthy it.
Coding
- Models used: ChatGPT 5.1, GPT-5.1-Max, Claude Opus 4.5
I usage AI models successful 2 precise antithetic ways erstwhile coding. For projects wherever I'm dropping a chunk of codification into the chatbot and asking what it does, what's incorrect with it, oregon wherefore I'm getting an mistake message, I haven't recovered thing amended than ChatGPT Plus utilizing GPT-5.1.
When I tested the escaped chatbots for coding prowess, Microsoft Copilot really bushed ChatGPT. And erstwhile I aboriginal tested the paid mentation of Claude successful a chatbot, Opus 4.5, it failed fractional my tests. But the $20/month ChatGPT Plus, utilizing the latest GPT, has reliably done a large job.
Also: The champion escaped AI for coding successful 2025 - lone 3 marque the chopped now
However, erstwhile I bash agentic coding, I've recovered that OpenAI's Codex utilizing GPT-5.1-Max and Claude Code utilizing Opus 4.5 are astonishingly great. Agentic AI coding is erstwhile I hook up the AIs to my improvement environment, fto the AIs work my full codebase, and past bash substantial, multi-step tasks.
For example, I utilized Codex to constitute four WordPress plugin products for maine successful 4 days. Just recently, I've been using Claude Code with Opus 4.5 to physique an full analyzable and blase iPhone app, which it helped maine bash successful small sprints implicit the people of astir fractional a month. I spent $200 for the month's usage of Codex and $100 for the month's usage of Claude Code.
It does astonish maine that Opus 4.5 did truthful poorly successful the chatbot experience, but was a superstar successful the agentic coding experience, but that's portion of wherefore we're looking astatine antithetic models. AI vendors are inactive moving retired the kinks from this nascent technology.
Notion databases
- Models used: Claude, ChatGPT, Gemini
Despite my annoyance with Notion's prioritization of AI upsells implicit basal merchandise functionality, I yet broke down and paid Notion's AI taxation for the year. It doesn't telephone it that, of course. Instead, it offers precise constricted AI proceedings capabilities with its $12/month Plus plan. You person to upgrade to the $24/mo concern program to get the AI features for real. I saved a fewer bucks by buying the twelvemonth astatine once, which made my effectual monthly interest $20/month.
Also: Dreading AI occupation cuts? 5 ways to future-proof your vocation - earlier it's excessively late
There are 2 things I usage Notion AI for. The archetypal is searching and summarizing my nonfiction drafts. ZDNET doesn't let extracurricular AIs to entree its articles, but I often request to excavation backmost into my erstwhile articles to reappraisal what I've written.
I tin constituent Notion AI to my draught library, and it volition scan the full thing, not conscionable searching for articles, but summarizing what I've antecedently said. This is large for assignments similar "Summarize what I've written astir location labs by looking successful the blogs database."
Note: I don't usage this AI-generated contented successful my articles. It's solely for readying purposes.
As you tin spot from the screenshot above, I tin popular unfastened the Notion AI model and inquire questions of it, close portion I'm writing. The nonfiction you spot successful the screenshot is this article, inactive being drafted.
The different happening Notion AI does for maine is to crook elephantine lists into categorized databases. I did a task wherever I gathered unneurotic a ample database of AI-related programming tools, but I wanted to beryllium capable to prime and presumption tools by vendors and categories. I gave Notion the list, and it created a database for me. It wasn't perfect, but it is usable.
Notion doesn't usage conscionable 1 model. It tends to leap betwixt Claude, ChatGPT, and Gemini (and their assorted models) depending connected what it thinks volition bash the occupation champion and which exemplary tin bash it for the slightest cost.
Speech designation
- Model: Parakeet
I usage the Mac's dictation diagnostic rather often. My small pup likes to slumber connected my near shoulder. That's adorable, but it makes it somewhat hard to benignant with 2 hands. So I usage my rodent with my close hand, and often dictate tiny blocks of substance utilizing the Mac's autochthonal dictation feature.
Recently, I've begun exploring dictation programs that connection much precocious code designation and investigation capabilities. One specified instrumentality is Paraspeech. Paraspeech offers 2 features I find rather compelling.
First, you wage conscionable once. There's nary subscription fee. This is imaginable owed to its 2nd feature: the AI runs connected your ain machine. When you instal Paraspeech, it downloads a saltation of the Nvidia Parakeet code designation model. The payment of this is that your code doesn't get sent to the cloud. It conscionable stays connected your machine.
This illustration shows however an AI exemplary tin disagree considerably from different models. This exemplary runs close connected my ain machine. It's designed specifically to bash code recognition. And it's included wrong the application, alternatively than called by it. That allows the developer to complaint a one-time interest without having to wage usage fees backmost to the AI vendor. It whitethorn besides springiness the developer a income advantage, due to the fact that paying erstwhile is thing rather charismatic to astir consumers. We each wage subscription fees erstwhile we person to, but cipher likes them.
Deep probe
- Models: GPT-5.1 Thinking (in ChatGPT Pro)
Deep probe is the process by which the AI spends a batch of clip figuring retired answers and providing information. It often uses a ton of AI resources, truthful it costs much to access. Back erstwhile I had the $200/mo Pro subscription for my programming, I besides utilized Deep Research to bash immoderate reasonably astonishing things.
My favourite illustration was erstwhile I asked Deep Research to make a selling briefing papers from root code. I utilized a alternatively agelong prompt, truthful I'm conscionable going to summarize. What I told it was that there's a reasonably modular process wherever engineering managers little selling and merchandise managers connected what a portion of exertion does.
Marketing managers past usage that briefing, which includes features, tech tidbits (think Apple presentation) and usage cases, arsenic a instauration to assistance them nutrient each their selling materials and activities.
But I didn't person an engineering manager. Instead, I told Deep Research to work each 12,000+ lines of codification from the root codification files, and from them, intuit what the products did, and from that, constitute the briefing documents. The results were astonishing. The AI got a fewer insignificant things wrong, but they were casual to correct.
I past upped the ante by taking those briefing documents and feeding them to NotebookLM, which produced compelling, voice-annotated descent shows describing the products. It was amazing, fascinating, and terrifying, each astatine the aforesaid time.
Net-net-net though, since I'm my ain engineering and merchandise manager, the 2 AIs unneurotic astir apt saved maine 60-80 hours of merchandise absorption work.
Analysis, SEO, and wide usage
- Model: GPT-5.1 Auto
My go-to instrumentality for wide concern enactment AI operations has been the Plus tier of ChatGPT. While it tin occasionally beryllium incredibly uncooperative, it's besides mostly useful. Even if I didn't constitute astir AI for a living, I'd astir apt support my $20/month subscription. I usually support the exemplary acceptable connected Auto mode, which lets the AI take the processing level for each prompt.
My woman precocious bought a Plus subscription. She has been going large guns with ChatGPT, discussing the nuances of spiritual arcana. She's some reasonably devout and reasonably academically inclined, truthful ChatGPT has been hugely amusive for her. I peculiarly bask the moments erstwhile she calls ChatGPT retired successful a fib oregon fabrication, due to the fact that that behaviour some amuses and annoys her astatine the aforesaid time.
In presumption of applicable use, I find the Plus tier to beryllium peculiarly adjuvant doing information analysis. I provender it large spreadsheets of antithetic kinds of data, similar sentiment analysis, and person it crunch it, digest it, find the cardinal issues, and springiness maine backmost results. I besides usage it to assistance take SEO keywords for my YouTube videos.
Up until now, nary different chatbot has been arsenic broadly useful. However, Gemini 3 looks similar it mightiness springiness ChatGPT a tally for its money. Gemini 3 has lone been retired for a week oregon so, which is wherefore I don't person capable acquisition to comparison them. But, who knows, successful six months this class mightiness database Gemini 3 arsenic the favourite exemplary alternatively of GPT-5.1.
Tools and models I don't usage
Among the large names, I don't usage Perplexity, Copilot, oregon Grok. I cognize Perplexity besides uses GPT-5.1, but it's conscionable ne'er resonated with me. It's known for search, but the fewer times I've tried immoderate searches, its results person been meh. Also, I can't basal the information that you person to log successful via email.
Copilot is an absorbing case. It did good successful the escaped coding evaluations, but it's precise Microsoft-centric. Despite the information that I was erstwhile all-Redmond, each the time, I hardly ever usage Microsoft products now. I support my household subscription to Microsoft 365 Family due to the fact that I cognize Excel and PowerPoint similar the backmost of my hand. But I conscionable don't bash Excel oregon PowerPoint enactment overmuch anymore. Copilot is tightly integrated with those tools, truthful it conscionable hasn't proven to beryllium that compelling yet.
As for Grok, I was amazed however good it did successful my coding tests. But it doesn't clasp an agentic candle to either OpenAI's Codex oregon Claude Code. It besides did a large occupation connected my question itinerary trial successful my escaped chatbot tests. But it inactive seems inconsistent, and determination are amended options easy available. I volition proceed to trial it, and thing is formed successful concrete, but it's disconnected my database for now.
Also: Apple's iPhone App of the Year is an AI instrumentality for radical with ADHD - and it's free
You whitethorn besides person noticed that I didn't speech astir AI-generated video. I haven't truly had overmuch of an juncture to bash overmuch with AI video, different than my aboriginal Sora 2 evaluation for ZDNET. I'm definite I'll bash more. When I do, I'll constitute astir it here. Also successful the video category, I'm surely hoping that Apple adds much and amended AI capabilities to Final Cut Pro, which is my preferred video editor.
Oh yeah. Apple. Nope. Sorry. Nothing to spot here. Apple conscionable hasn't been connected my radar overmuch for AI. I tried its Coding Intelligence diagnostic successful Xcode, and it conscionable crashed each the time. Apple's implicit deficiency of AI practice successful articles similar this is atrocious quality for the company. The Cupertino elephantine truly needs to up its AI game, and fast.
What astir you?
What astir you? Do you actively take antithetic AI models depending connected the task, oregon bash you mostly instrumentality with 1 favourite tool? Have you recovered definite models to beryllium intelligibly amended for images, coding, research, oregon writing? Are determination immoderate tools mentioned present that amazed you, oregon ones you powerfully disagree with? Let america cognize successful the comments below.
You tin travel my day-to-day task updates connected societal media. Be definite to subscribe to my play update newsletter, and travel maine connected Twitter/X astatine @DavidGewirtz, connected Facebook astatine Facebook.com/DavidGewirtz, connected Instagram astatine Instagram.com/DavidGewirtz, connected Bluesky astatine @DavidGewirtz.com, and connected YouTube astatine YouTube.com/DavidGewirtzTV.
English (US) ·