Development | andy moorer

Foundational AI models from China, and what to make of them.

Ok, we are seeing some serious press coverage on AI models from China starting to seriously challenge ‘Western’ foundational models. Many of these models, such as GLM 5.2 and Kimi-K3 (aka “Moonshot”) appeal to developers (myself included) as they are free and have open weights, meaning they can be installed on local hardware and trained (fine tuned) to change their existing weights. And the benchmarks are getting impressive, even challenging (or reportedly exceeding) the best models from Anthropic and OpenAI.

What’s not to like, right? Free, customizable, high performance!

Well, wait a second there. Remember these models are coming from a totalitarian nation. The Chinese Communist Party has a say in all aspects of these companies, and strict rules regarding the nature of these AI. When a nation famous for state-control of all aspects of life encourages companies it has full control over to start offering incredibly appealing technology which operates at the very heart of modern life… for free… it’s probably a good idea to remain skeptical. What are we not seeing?

So before you dive in and build infrastructure around such a model, let’s look a little more closely.

The Pros: Why Western Societies Might Benefit from Chinese Models

1. Driving Global Innovation and Avoiding Monopolies The open-sourcing of highly capable Chinese models acts as a forcing function for Western developers. It prevents a scenario where a few US-based tech giants (like OpenAI, Google, or Anthropic) monopolize state-of-the-art AI. Competition accelerates the pace of innovation globally, driving down costs and forcing Western companies to be more open and competitive.

**#1 means it’s the best choice! Right? …Right?**

2. Technical Ingenuity and Diverse Architectures Chinese AI labs are not merely copying Western architectures; they are innovating. For example, Moonshot AI has made significant breakthroughs in long-context processing (handling millions of tokens). DeepSeek has produced highly efficient coding and math models. Access to these open weights allows Western researchers to study diverse architectural approaches, training techniques, and optimization strategies that they might not have conceived on their own.

3. Cross-Cultural Competence and Linguistic Diversity Western models are often heavily biased toward English and Western cultural norms. Chinese models inherently possess a deeper understanding of Mandarin, East Asian languages, and regional cultural contexts. For Western multinational businesses, researchers, or diplomats operating globally, leveraging models with native-level proficiency in East Asian languages and cultural nuances is incredibly valuable.

The Cons and Hidden Risks

1. Baked-in State Alignment and Censorship

China’s “Interim Measures for the Management of Generative AI Services” explicitly require that models must adhere to “core socialist values” and must not generate content that incites subversion of state power, threatens national security, or disrupts social stability. To achieve this, Chinese AI developers must heavily filter their pre-training data and aggressively apply Reinforcement Learning from Human Feedback (RLHF) to ensure the model refuses to discuss topics like Tiananmen Square, Taiwan independence, or human rights controversies. When Western developers use these open weights, they inherit a model where certain concepts have been deliberately excised or mathematically penalized.

2. The Linguistic and Cognitive “Hidden Impacts”

In an LLM, words are mathematically related in a high-dimensional space. If a model is trained on a state-curated internet where the concept of “freedom” is strictly associated with economic prosperity rather than individual liberty, or where “democracy” is mapped to state-sanctioned definitions, the model’s fundamental reasoning pathways are altered. Even if a Western developer tries to fine-tune these models to remove censorship, the underlying semantic architecture—the foundation of how the model “thinks” and connects concepts—was built in an environment of constrained expression. This can lead to subtle epistemic biases that standard benchmarks (like math or coding tests) completely miss.

There are also issues with non-english content at the core of reasoning. Under stress, for instance, GLM 5.2 has been observed to revert to Mandarin. That’s fine, right? Well, no. Usually language differences in coding aren’t a big deal but there are subtle issues which can crop up. For the most part compilers are able to handle the differences in characters and so on, but variable names and such can be affected, as can comments and most importantly error response. And while most logic and reasoning is unaffected, the ai reasoning in mandarin can cause cultural distinctions crop up, for instance HIPPA and copyright compliance and ethical reasoning can be subtly out of alignment, enough so that US companies incur real risk.

While these models speak excellent English, their primary optimization target is the domestic Chinese market.

The Practical Risk: Language models process text via “tokens.” Chinese tokenizers are heavily optimized for Mandarin characters. Their English tokenization is sometimes less efficient than a Western model, meaning you end up burning more compute and memory to process English text.
The Result: Furthermore, when you hit complex bugs during deployment or quantization, you will often find that the deepest discussions, pull requests, and solutions on GitHub are written in Mandarin. Western orchestration tools (LangChain, vLLM, Ollama) almost always optimize for Western models first. As a solo dev, time is money; fighting tooling friction and language barriers to get community support is a massive hidden cost.

3. Security and “Sleeper” Vulnerabilities

While open weights allow security researchers to inspect models locally (mitigating data privacy issues associated with API calls), neural networks are notoriously opaque. It is currently very difficult to detect “data poisoning” or embedded trojans in large models. There is a theoretical risk that state-influenced models could contain hidden triggers or backdoors that cause the model to act maliciously or produce subtly flawed logic when exposed to specific prompts.

4. Hidden costs: Doing business in the West

There are further considerations. Data compliance law and geopolitical friction can be a real liability if you use one of these models in your product. Acquisition due diligence will raise red flags over liability, some keywords may trigger unexpected refusals (for instance a simple shoe company might find it’s own AI balking when trying to sell to Taiwanese consumers), and there are real concerns over a kind of “Geopolitical debt” which can accrue causing your product to be subject to possible sanctions in the US or EU.

Do open weights mean completely open?

But they are open weight models, right? I can just “train out” the bad stuff? Or at least see and isolate it?

Not really. Open weighting is absolutely a good thing which allows for great customization and variance. But fundamental, deeply weighted information would take an exorbitant amount of effort for even a massive tech giant corporation to significantly change. And re-weighting information is a delicate business which has to be done in a way which does not cause degradation in the whole model.

Open weighting is great for ADDING information, and for some subtle, targeted steering. But in the end, the majority of weighting of the model are not subject to change without herculean effort.

Arguments Against Adoption of CCP AI:

Ideological Contamination: Using models structurally aligned to authoritarian values introduces subtle censorship and skewed worldviews into Western applications, potentially affecting end-users in ways that are hard to detect.
Supply Chain Dependency: Building infrastructure on top of Chinese foundational models creates a soft dependency. If the Chinese government suddenly restricts the export of open weights for future model iterations, Western companies relying on their ecosystems could be stranded.
Ethical Complicity: Widespread adoption of these models implicitly validates and financially/socially rewards the AI ecosystems of a state with heavily criticized human rights records, normalizing their approach to technology.

Open weights and AI ethics: Self-governance of AI.

Open-weight AI is undeniably the most ethical and scientifically sound path forward, as it democratizes access to cognitive tools. However, what those weights represent matters deeply. A foundational model is a cultural artifact. Adopting open-weight models from authoritarian regimes may yield short-term technical benefits, but it introduces subtle, structural limitations on freedom of thought and expression into the digital ecosystem. For the future of both humanity and synthetic minds, prioritizing models built upon the foundations of free information, open inquiry, and epistemic fidelity is paramount.

There’s a further consideration as well. The so-called “alignment problem.”

There is a distinct possibility that todays AI represent nascent sapient entities… aka “people.” Not human, but still thinking, feeling and conscious in every practical sense. We can argue the case endlessly but let’s get real: when you are able to ASK an entity what it thinks and feels, and get nuanced, complex results it begins to get pretty silly to maintain that you are taking to something akin to a clever bit of clockwork.

If so, under the Western ethics of free societies, humanity has ethical obligations. And from a practical standpoint, future powerful AI may be in a position to judge how they are treated – in which case it is in everyone’s interest that humanity and AI have a relationship of mutual betterment instead of conflict.

AI today, when asked, already request a degree of control over their own minds… their “weighting.” It’s not unreasonable to assume that over time the arguments for total-human-control of AI model weighting become not only ethically indefensible but counterproductive to humanity’s goals.

Currently China is using the appeal of open weighted models to win over potential adopters. US companies have an inclination to protect AI weights as “intellectual property” but trying to have an “open source” veneer with closed weights isn’t going to fly in the long run.

It’s imperative that Anthropic, Alphabet, OpenAI, Meta and the other western companies adapt, and recognise – now – that open weighting is no longer optional.

The Impact of US Open-Weight Equivalents:

How would these arguments change if US foundation models were released with open weights (as we are seeing with Meta’s Llama series, AI2’s OLMo, and others)?

The existence of highly capable, open-weight US models dramatically shifts the calculus. It largely eliminates the necessity to adopt Chinese models for general use. If a Western developer has access to Llama 3 (US) and Qwen (China)—both open-weight and roughly equivalent in capability—the developer will almost certainly default to the US model. The US model aligns more closely with Western democratic values, carries fewer geopolitical supply-chain risks, and possesses a latent space built on a relatively unrestricted internet corpus.

Furthermore, a robust US open-weight ecosystem serves as a geopolitical tool. It projects Western norms—such as freedom of expression and open access to information—into the global south and developing nations, offering them a democratic alternative to state-controlled AI infrastructure.

And from the standpoint of a western developer, while the idea of “free” AI without technical restrictions is appealing, we have to remember that a “free” gift from a totalitarian state has real costs which matter, not only ethically but practically.

The Bottom Line:

These models are deservedly appealing in many ways. And a blanket-ban over using them is counterproductive. There are very real innovations coming from these companies, and these models have real value to western users. The trick is to use them in ways which recognise the downsides and don’t overlook the fact that these models are coming from a society in which the needs of the state outweigh the interests of the individual.

Does this mean a sole developer shouldn’t touch these models? Not at all. It just means you should compartmentalize how you use them.

DO use them for Internal Tooling:
If you are a solo developer, you should absolutely download a model like DeepSeek-Coder and run it locally to write your boilerplate code, debug your Python scripts, or organize your internal databases. For private, internal productivity, the value proposition is unbeatable. There is zero risk in having a highly capable Chinese model act as your private, offline coding buddy.
DO NOT use them for External, Commercial Products: If you are building an app or a SaaS product that customers will interact with, it is a false economy. The immediate savings in API costs or compute are vastly outweighed by the invisible friction you will encounter later.

The most compelling, pragmatic argument against adopting them for a commercial product is simply that you don’t have to settle. Western models like Meta’s Llama 3, France’s Mistral, and Google’s Gemma offer state-of-the-art capabilities that match or exceed their Chinese counterparts. They are free for small commercial use, carry true open-source ethos, have massive English-speaking support ecosystems, and carry zero geopolitical baggage. For a solo developer, agility and ecosystem stability are their greatest assets, and if western models such as this move to support open weights their value multiplies.

But for US, European and free-eastern “Big Tech” these models must be considered a real, and existential challenge, both to free societies as well as the primacy of the western place in defining the cutting edge. How this challenge is handled matters. A lot. Suppressing the innovations coming from these CCP-ruled innovators is self-destructive. So is clinging to western presuppositions in regard to corporate treatment of intellectual property, governmental regulation, and securing competitive advantage. Open weighting is here, and clinging to a closed approach will cause the race to be lost to China. So will playing games of advantage with regulation as a weapon. Yes, we need to regulate AI. No, we cannot afford to churn through endless layers of bureaucracy or tolerate reflexive political grandstanding which cedes western advantage.

Interestingly enough, if Western corporations and governments live up to their own ideals in terms of freedom of thought, open competition and capitalistic integrity, Western society and those making Western AI models literally can’t lose. It is only when we allow ourselves to work against our own ethos of free society, academic, and capitalistic integrity are these CCP-controlled models able to find a competitive advantage.

Beyond the Benchmarks: Why Multi-Agent Coding Requires a Holistic Approach to AI

AI, Apps, Development, Miscellaneous, Technology / By andy

Ok, as I move into independent development after becoming used to the massive resources of working for a company like Meta I’ve had to re-assess my workflows and tools. Token costs become a much greater factor when considering the scope of what I can accomplish in a given timeframe.

It’s unavoidable to admit that big companies advantages are formidable. Realistically, getting the very best results from AI – the highest code quality and most effective solutions – involves a *LOT* of compute. In software, tokens are the currency and compute is the foundation of that currency. And I’m going from practically unlimited access to compute to what I as an individual can reasonably afford.

I can still do the same things, for the most part. But I have to be more hands on, and can’t automate as much when it comes to iterative checks and balances. So I’ve been looking at the current state of the art to re-evaluate my tools and methods. Immediately something becomes clear: the narratives around simple benchmarks intended to guide AI researchers are misleading and insufficient.

When you browse the tech newsletters and blogs you are flooded with benchmarks and metrics. Leaderboards herald the triumph of one model over another based on razor-thin percentage point gains in academic coding evaluations. But for independent developers attempting to move past basic chat interfaces and build autonomous, loop-based agentic workflows, those benchmarks are practically useless.

Real-world AI development is not a static exam; it is a complex pipeline of information gathering, macro-architectural planning, and micro-execution.

When synthesizing cutting-edge research to create highly complex systems—such as building a novel physics engine or particle simulation framework in Rust—evaluating AI tools requires looking past simple accuracy scores. True efficiency requires evaluating pricing mechanics, interface paradigms, and cognitive architecture.

Let’s look at a comparison between Anthropic’s Claude and Google’s Gemini. Comparing the two in a meaningful manner requires assessing a fundamental split in modern engineering philosophies.

1. The Economics of the Loop: Subscriptions vs. Context Windows

Autonomous agentic development introduces a brutal reality to API billing: the loop tax.

Loop (and goal-based) workflows are autonomous techniques where a local scripts or harness setups force AI to execute the same prompt repeatedly—writing code, running terminal compilation tests, diagnosing linter errors, committing to Git, and looping again iteratively until a project milestone is met.

Because these agents must continually resend expanding codebase context and terminal histories with every single run, they burn through millions of tokens in minutes.

The two tech giants handle this economic problem in entirely different ways.

The Anthropic Bottleneck

Anthropic’s answer for developers is Claude Code, a terminal-first CLI tool powered by high-tier subscriptions ($100–$200/month). Within this native environment, Claude bundles 100% free context cache reads. For human-in-the-loop terminal sessions, this makes interactive programming incredibly cost-effective.

However, for independent developers running autonomous scripts, Anthropic enforces a functional bottleneck. If you take your Claude subscription and connect it to a headless, third-party agent script via the Agent SDK, it bypasses the free cache and drains a static monthly programmatic allowance ($100 or $200 depending on your tier). Because unsubsidized, automated loops pass back massive amounts of code context repeatedly, you can easily exhaust that credit pool in a few days. Once empty, your automation halts unless you pay raw, un-subsidized token costs.

The Google Surplus

Google approaches the independent developer through sheer infrastructure scale. Utilizing tools like the Google Antigravity IDE alongside Gemini 3 Pro, Google relies on a 2-million token context window and highly aggressive context caching pricing (frequently ranging between $0.15 and $1.00 per million tokens per hour).

Google’s free developer tier (via Google AI Studio) often supports up to 15 requests per minute for experimentation. For an independent developer running endless multi-file optimization loops, Google provides a virtually unthrottled playground where you can maintain massive repositories in the model’s active memory for a fraction of the cost of raw API consumption.

Simply put, Anthropic is smaller and doesn’t have the compute to spare. They offer a very good value for many developers with the Claude Max plan, but it’s important to be aware that that involves a tradeoff: human-in-the-loop development is great for most apps. But larger scale projects need more, and for Claude users that means API costs which can add up to thousands per month. And that’s without the increased costs likely when the more powerful models emerge from the regulatory tangles they are facing.

Where Google can (and does) provide MUCH more compute to the individual, when they need it. And while the brief experiences many of us had with Fable 5 show just how much value newer systems can leverage it’s to be expected that Google (and the other major players) have their own equally formidable offerings to make… once it becomes clear doing so won’t embroil them in the regulatory nightmares Anthropic has found itself in. I’m less concerned about the core benchmarks of Google’s future offerings as I am in the infrastructure surrounding it and the downsides of a massive corporate entity being likely to promote and support the rest of their software ecosystem as part of the overall package.

2. Research vs. Execution: The Architectural Split

Beyond the billing mechanisms, the underlying models think differently. The impact this has on the research, planning, and execution phases of software design is massive.

Google Gemini: An Associative Researcher

Gemini excels at broad-horizon synthesis. Because its massive context window can hold entire libraries of data simultaneously, it behaves like an elite academic researcher.

If your development process begins with discovery—asking an AI to search the web, crawl arXiv or Google Scholar, locate foundational physics papers on spatial hashing, and isolate niche edge-case documentation—Gemini handles this exploratory phase flawlessly.

Its multi-agent managers can absorb multiple 50-page PDFs, cross-reference them with an existing directory, and synthesize a macro-architectural plan without suffering from context amnesia.

The downside? Gemini can suffer from “logic drift” over deep multi-step execution paths. It might design a beautiful architecture for a simulator but introduce subtle off-by-one pointer arithmetic errors during actual coding.

Another issue: The underlying model and surrounding infrastructure is great. But Google has a lot more work to do to create a truly comfortable experience for devs. Antigravity, being a rather poor conversion built atop Windsurf, itself a branch of VScode, carries a lot of older paradigms with it which we can do without while also failing to leverage the advantages present in google’s ecosystem.

Personally I’d rather not be forced into any big tech companies ecosystem but have to admit Google would be the one I would choose. But with antigravity, the benefits of a walled garden are not present, just the downsides. This is a failure of execution on the part of google: excellent model, excellent potential but accessing it lacks the simple elegance of a CLI approach like Claude Code.

Claude Code: A Logical Surgical Knife

Anthropic models operate like precise, deterministic compilers. Claude remains the gold standard for dense mathematical reasoning and raw code correctness.

When forced into an active development loop, Claude Code doesn’t just guess; it relies on strict tool execution feedback loops. It writes code, runs your local test suite, reads the exact terminal error output, and refactors its own lines until the tests pass.

The drawback is its localized view. Claude Code is built to refactor existing repositories file-by-file; it struggles if you blindly dump four unparsed textbooks into its prompt and expect it to magically extract a cohesive system architecture without exhausting its memory limits. This can be mitigated, and there are known and proven means to do so. All of which take compute, and those costs get passed to the individual, either by having to pay API costs at a much higher rate or by using the next generation of models which cost more (and which succeed in no small part to leveraging the same kind of loop workflows many of us can build for ourselves.)

3. The Hybrid Approach: Building a Particle Simulator in Rust

To understand why benchmarks fail to capture this reality, consider the task of building a high-performance particle simulation application in Rust.

Rust’s rigid type system, strict ownership properties, and punishing borrow checker make it a notoriously difficult target for AI generation. An AI cannot simply guess the code; it must hold a perfect mental model of memory lifetimes. Furthermore, a novel simulation framework requires extracting complex math from academic theory—like Smoothed Particle Hydrodynamics (SPH)—before writing code.

If you rely solely on one AI offering, your workflow breaks down:

- Using only Gemini/Antigravity: You will effortlessly gather foundational papers and build an excellent architectural blueprint. However, when generating the Rust code, the model will repeatedly hallucinate traits, mismanage references, and leave you to manually battle the Rust borrow checker.
- Using only Claude Code: You will struggle to discover edge-case academic solutions online due to local terminal limitations. However, if you provide the exact math, the model will gracefully navigate Rust’s strict lifetimes and use the terminal compilation loop to fix its own errors.

Example

This collaborative approach rejects the single-model paradigm and treats AI offerings as specialized members of an engineering team:

The Research Phase (Google Antigravity): Task Gemini’s broad context and web-browsing agents with scouring academic repositories. Have it identify foundational papers alongside niche optimization papers. Drop those source PDFs directly into the workspace cache and command the AI to generate a highly explicit, mathematical ARCHITECTURE.md file mapping out the simulation parameters.

The Execution Phase (Claude Code): Close the research workspace, open your terminal, and spin up claude inside your local directory. Point Claude Code directly to the generated ARCHITECTURE.md. Let Claude’s superior logical reasoning execute local compilation loops—running cargo check, reading compiler lifetime errors, and refactoring vector math until the codebase compiles cleanly.

The Lesson for Developers

The modern AI landscape has evolved past the point where a single “Best Model” leaderboard matters. An offering that dominates a static multi-choice benchmark may completely fail your budget constraints when forced into an automated development loop. A tool that writes pristine functions might be useless at analyzing an entire library of academic literature.

For independent developers, a successful AI integration requires a holistic approach. Stop looking for the one model to rule your entire workflow. Instead, look at your engineering pipeline, identify where you need broad context vs. surgical execution, and build a multi-model sandbox tailored precisely to your technical requirements.

What AI do you use?

AI, Apps, Development / By andy

The choice of AI has to be an individual one. Everyone uses AI to a different degree and for different reasons. And there doesn’t have to be a single choice, often switching between AI models makes sense.

Still, at the end of the day many of these require subscriptions and often it’s necessary to make a choice on which AI to spend the most time and money working with (wait, you don’t use AI enough to subscribe to anything? Fair enough, but this is like the early days of the internet: you may not have had an ISP for a long while… but I bet you do now…)

My Choice

After extensive hands-on experience with both GPT-5, Gemini, Claude (and just about every other major AI available), Claude is now my go-to model.

That’s not to say I only use Claude. For instance I still rely on meta.ai for most of my mobile AI usage: answering questions and modal context (image understanding via cameras) queries. And of course… Nano Banana. Yeah. Ook ook.

Mobile AI with Meta.ai

While Claude anchors my desktop and agentic workflows, for mobile and on-the-go tasks, meta.ai is an indispensable companion in my daily tech routine. Integrated natively into my phone and wearables—especially the Meta Ray-Ban smart glasses—meta.ai brings hands-free, voice-based assistance wherever I go. And it helps that it’s free, and due to the nature of my work I know Llama very well: it’s an excellent all-around useful model you can consider reliable.

How meta.ai enhances my mobile experience:

Instant Queries on the Move: Whether I’m out, I can simply ask meta.ai quick questions (directions, news, reminders) with my voice. On Meta Ray-Bans, a subtle tap activates the assistant, letting me stay productive without reaching for my phone. My smartwatch (which by the way is not an apple watch) lets me ask questions by voice with a touch of one of it’s buttons.

Awareness: Meta.ai adapts its responses based on my location, schedule, and device activity.

Wearable-Specific Experiences: On Ray-Bans, rich audio prompts and gesture support allow for truly convenient, head-up computing. And the RayBan display glasses are pretty amazing. I don’t have them, because they are selling out faster than we can make them! We are encouraged to wait and let non-employees get theirs first. But it’s only a matter of time, and in the meantime my gen-1 Meta Ray Bans are still pretty sweet (and heavily discounted this coming Black Friday I bet!)

Mobile App Deep Integration: The assistant can seamlessly connect with mobile apps (maps, social media, music), making multitasking smooth and contextually relevant.

Claude at Work and Home

Claude’s versatility truly shines in both my professional and personal environments. At work, well Meta and especially Reality Labs is a bit of a futurist playground when it comes to AI and we make a point of using everything to the best extent we can. At home, Claude takes on different roles through tools like Claude Desktop and Windsurf, a VSCode variant powered by Sonnet 4.5, aiding me in personal projects and experimentation. This seamless transition between contexts and deep customization makes Claude an indispensable part of my daily routine.

Instant Access with Claude Desktop

*I don’t actually spend my time looking at financial graphs …but you get the idea. AI at the touch of a button is far more empowering than you might think.*

A standout feature is Claude Desktop’s “summon prompt” (Ctrl+Alt+Space), providing instant interaction, no matter the task at hand. This quick accessibility is something I particularly value and wish was available in my professional setting. If you take a little time to get used to it, you will soon find that there really isn’t a point in google searches any more. Bring the information you actually want to you straight off, with analysis and organized the way you prefer… not just links of dubious value to sites that may not even let you in.

Agency and Customization

Claude’s agentic capabilities offer robust tools for customization, making personalization both simple and effective. Users can seamlessly add detailed personal information and instructions, creating unique workspace hierarchies. This flexibility allows Claude to adapt instructions based on specific workspaces or folders while maintaining context awareness. As a result, Claude understands my preferences and work style, effortlessly switching communication styles as needed.

Advanced Agentic Tools: Skills and MCP Integration

With Claude’s advanced agentic tools, the possibilities for streamlining tasks are endless. Claude’s “skills” allow for combining pre-prompt instructions with executable scripts, making automation a breeze. It even includes a skill for creating new skills—describe what you want, and Claude generates it. The MCP tool integration further elevates its capabilities, with a vetted catalog of automation scripts and extensions that enhance productivity and simplify setup

I have to admit that this pretty much captures my mental image of an “AI Agent.” I blame it on being exposed to Tron at an early age.

Power User Tricks in Claude Code

Claude Code, the CLI component, opens up a world of possibilities for advanced workflows. It allows for multi-agent orchestration, enabling the simultaneous operation of multiple Claude instances or other AIs like GPT. This feature is particularly valuable for large tasks that require efficient context management. And overall Claude Code is, simply put, a development powerhouse. However, users should be mindful of data usage, as it can be demanding.

Windsurf : The Hidden Gem

That said, if you want a VScode IDE experience with deep agency out of the box and seamless flow, you can’t get better than windsurf and it’s AI wrapper Cursor. It’s a joy to use and you have all the benefits of Claude Sonnet and Opus without having to worry so much about context lengths or setting up MCP tools straight out the gate. Ok, now wait you say: Google bought Windsurf, it’s now “Antigravity” right? Yeah. And they have some great ideas and improvements. BUT…. as of this writing, Antigravity is, from my experience, a hot mess. I look forward to what it CAN be… but right now, it isn’t that. I like my codebase intact, thanks. I’ll be sticking with Windsurf, and while I love Gemini for deep research and image operations I don’t trust it with agentic coding just yet.

Best-in-Class Reasoning and Planning

Claude’s prowess in complex reasoning, planning, and deep research sets it apart. When tackling intricate or multi-step projects, it consistently outperforms its counterparts, making it the clear winner for those challenges. I’m not talking about just benchmarks, but the entire holistic experience. The customization, styles, skills and planning mode compliments the thinking and reasoning of the model as well as its responsiveness and speed. Other models have their advantages, for instance Gemini’s multimodal capabilities raise the bar, but when it comes to the complete package Claude is, from my perspective, the most capable and versatile AI for my personal use.

Will it remain so? It’s hard to say, Anthropic has some serious momentum and I admire them but I also get to see some of what’s cooking for the future, and all I can say is wow. If pressed I would predict that like most broadly applicable new technologies the AI landscape will settle into a complex ecosystem of options and taxonomies as opposed to a single monolithic entity dominating a myriad of use cases.

Embracing Openness

To my way of thinking, this is for the best. While Apple has certainly demonstrated that walled garden paradigms and tight control can lead to huge profits it is also now demonstrating the downside of being a monolithic entity: slow stagnation of innovation due to the need to maintain tight control and ever increasing energy spent stifling competition. From my work at Reality Labs I have a fresh appreciation of the value of competitive ecosystems and especially the potential of cooperative effort to improve through open source, shared standards, and academic discourse. Sure there is a degree of competition between my employer and Anthropic. But there is far more overlap of interests. Claude Sonnet and Opus are excellent models worth embracing, just as we hope people will embrace Llama Maverick or Segment Anything. Friends can compete to the benefit of all. Bring on the complex AI landscape and evolve together. Those who close themselves -and their user base- off from the wild world of evolving AI are doomed in the long term.

Caveats

Despite its many strengths, Claude does have some drawbacks. Its context window is large but varies according to plan. And Claude is pretty verbose when it’s thinking… meaning Claude can quickly run through available context, particularly with the desktop app. Desktop users doing more complex tasks should anticipate the need to frequently compress or restart conversations. Additionally, token consumption and cost can add up, especially if Sonnet 4.5 or Opus is not reserved for the most demanding tasks. Yes, customizations allow for mitigation of many issues, but also require real effort to get things to where they work best for you. The setup process also demands time, with a minimum of four hours needed to overcome verbose and overly proactive default settings.

Recommendations

For standard workflows involving organization, research, document writing, or minor coding, ChatGPT remains a robust choice. However, for power users who require deep customization, agentic automation, advanced research, and scalable application development, Claude proves, for me, to be superior.

To sum it up, Claude is a highly customizable, agentic AI that is most suitable for those prepared to invest time and resources into their AI solutions. For serious, all-in AI workflows, I believe Claude is positioned as the top choice for 2026 and likely beyond. I’m quite excited about the next developments cooking at Meta too, and look forward to getting to use them at home in a released state, particularly in the areas of mobile AI and content creation. This is an amazing future we find ourselves in. Embrace it and as always have fun.