Gianluca Iannotta

Geek, wonk, nerd at heart. Repeat entrepreneur, DevOps and AI engineer. CEO at Axiom & Tublat.

[Gianluca Iannotta @jannotix]

@morganlinton @VulcanBench @zcode_ai is great but I used glm 5.3 in Trae work and it generated superior code than zcode.

[Gianluca Iannotta @jannotix]

@tridentasia @bindureddy I understood correctly, DeepSeek V4 Flash Vision basically keeps the same text capabilities as DeepSeek V4 Flash, but adds only the ability to see and understand images, so it doesn't improve.

[Gianluca Iannotta @jannotix]

@jp54362 Fable 5
Gpt 5.6 sol
Glm 5.3
Kimi K3
Grok 4.6

[Gianluca Iannotta @jannotix]

The biggest threat to the AI market may not be @OpenAI releasing a smarter model.

It may be OpenAI making intelligence dramatically cheaper.

If its custom inference stack really gets anywhere close to the 70–80% cost reductions people are discussing for specialized AI hardware, the impact could be brutal.

Not because API prices automatically drop 80%.

But because OpenAI suddenly has options:

lower prices, raise usage limits, improve margins, or do all three at once.

That’s much harder for competitors to answer than a benchmark win.

A better model can be matched.

A structural cost advantage across models, inference, hardware and distribution is a different game.

The next AI war may be less about who has the smartest model and more about who can deliver the most useful intelligence per dollar.

[Gianluca Iannotta @jannotix]

After spending a lot of time with almost every major AI model on the market, my current top tier is:

1. Claude Fable 5
2. GPT-5.6 Sol
3. GLM-5.3
4. Kimi K3

This is also shaping the Cycle setup I’m testing right now:

Fable 5 - Architect
Grok 4.6 - Executor
GLM-5.3 - Architecture Reviewer
Kimi K3 - Implementation Reviewer
GPT-5.6 Sol - Final Judge

The important part is that Cycle isn’t designed to throw five expensive models at the same problem and burn tokens.

It’s designed to preserve useful project memory, keep context bounded, give each model a specific job, and avoid repeatedly rediscovering the same codebase.

So the goal isn’t “more agents = more cost.”

The goal is to spend intelligence where it matters and finish with code that has actually been implemented, challenged, tested and verified before production.

This is just my current configuration.

Cycle itself is model-agnostic. Use whatever models you trust.

[Gianluca Iannotta @jannotix]

@louszbd Yeaaa 🚀

[Gianluca Iannotta @jannotix]

@MiniMax_AI just emailed me that subscription prices are going up.

This is exactly what I was worried about.

AI subscriptions are slowly converging toward the same price bands across providers.

If everyone ends up charging roughly the same, price stops being the differentiator.

Then the real competition becomes:

usage limits, context, model quality, reliability, openness, and how much actual work you can get done before hitting a wall.

The cheap-model era may be ending faster than expected.

[Gianluca Iannotta @jannotix]

GLM-5.3 Max by @Zai_org in @zcode_ai just surprised me.

I gave it a codebase to analyze with no background on its history.

It didn't just explain the architecture.

It inferred which model had likely produced/copied parts of it and reconstructed the actual evolution of the codebase closely enough to match the history I already knew.

I tested the same codebase with Fable 5, GPT-5.6 Sol Max, Grok 4.6 xhigh, Gemini 3.7 Flash and MiniMax M3 Thinking.

None of them noticed it.

That's the kind of thing benchmarks rarely capture.

GLM-5.3 is seriously underrated.

[Gianluca Iannotta @jannotix]

I've tested a lot of coding agents while building Cycle.

If I rank them purely by code writing and overall code quality, my current top four are:

1. Cursor
2. Claude Code
3. TRAE Work
4. ZCode

Cursor is the one that impressed me the most recently. It feels mature, fast, and very strong when you combine its native models with third-party ones.

Claude Code is still extremely reliable for serious engineering work.

TRAE Work has a lot of potential, especially because of its more open approach.

ZCode is improving quickly and GLM makes it especially interesting.

This is based on real daily use, not benchmarks.

The ranking will probably change again. That’s how fast this space is moving.

[Gianluca Iannotta @jannotix]

I don't think the future of smartphones and PCs is simply "better hardware + more AI features."

The architecture itself should change.

AI should become the operating layer.

Local models running on dedicated NPUs.
Persistent agents that understand intent and operate software for us.
CPU/GPU for traditional workloads.
Cloud compute when more scale is needed.
And eventually quantum systems as specialized remote coprocessors for the problems where they actually provide an advantage.

I don't expect a quantum computer inside an iPhone anytime soon.

But I do expect our devices to become quantum-ready, AI-native computers.

The biggest change may be the interface.

Today we open apps and tell computers what to do.

Tomorrow we tell an agent what we want, and the computer decides which model, tool, application or compute architecture should do it.

The future OS may look much less like a collection of apps and much more like an orchestration layer for intelligence.

[Gianluca Iannotta @jannotix]

It's much more comprehensive than Grok Web in my opinion, and the Grok-Composer model pool is much more generous than the Grok subscription itself. Cursor is also part of SpaceX. My only recommendation is to use third-party models as little as possible, as the third-party model pool fills up very quickly. Additionally, in some cursor plans you also have access to Grok bot.

[Gianluca Iannotta @jannotix]

@thisisdimm In @cursor_ai you can organize chats by project, I don't understand what your problem is, can you explain it ?

[Gianluca Iannotta @jannotix]

Just shipped two more Cycle integrations:

Cycle for @cursor_ai
https://t.co/ucRrWW4vBQ

Cycle for TRAE Work
https://t.co/bPpgOS9GkC

The goal with Cycle has always been simple: make the engineering workflow portable instead of tying it to one coding agent or one model family.

Plan, implement, verify, review, challenge, then ship.

Both plugins are still early, so if you try them and something breaks, open an issue on @github. Real usage is the fastest way to make them better.

More integrations coming.

[Gianluca Iannotta @jannotix]

@cursor_ai became a much more interesting product after the @SpaceXAI acquisition and the integration of @grok.

After testing it again, I think the combination of Composer + Grok + third-party frontier models makes Cursor probably the strongest coding product on the market right now.

The main bottleneck I hit was third-party model usage.

That pool can disappear surprisingly fast if you're doing serious development all day.

Cursor already supports BYOK for some providers, but I think they should go much further.

Let developers plug in @Zai_org, @MiniMax_AI, @Alibaba_Qwen, @deepseek_ai, @OpenRouter and basically any OpenAI-compatible provider they want.

That openness is one of the reasons tools like @opencode and TRAE remain so interesting to me.

Now I'm waiting for @ByteDance_CN's response.

If the international version of TRAE gets closer to the model ecosystem and capabilities available on the Chinese side, I think it could become a very serious competitor.

For me, right now, the race is becoming pretty clear:

Cursor is leading. OpenCode has the openness. TRAE has the potential.

And those are probably the three products I'd watch most closely.

[Gianluca Iannotta @jannotix]

I’d rank this quite differently.

I’ve been using most of these models for months while building and testing Cycle, mostly on large, real codebases rather than isolated benchmark tasks.

My current list:

S: Fable 5, Opus 5, GPT-5.6 Sol
A+: Kimi K3, Grok 4.6, GLM-5.3, Qwen 3.8 Max
A: Sonnet 5, DeepSeek V4 Pro, Composer 2.5
B+: MiniMax M3, DeepSeek V4 Flash, Ox Alpha, Muse Spark 1.2
B: MiMo 2.5 Pro, Gemini 3.7 Flash, GPT-5.6 Luna
C+: Gemini 3.1 Pro

The biggest difference for me is what happens once you put these models on large codebases for hours instead of judging them from benchmarks or short prompts.

Some models are much better than their reputation suggests. Others look great in isolated tests but become less reliable as the task and context grow.

I also don’t think there is one “best coding model” anymore. The best architect, executor, reviewer and debugger are often different models.

That’s one of the main reasons I built Cycle.

[Gianluca Iannotta @jannotix]

@buildwithrajath Opus 5 bro

[Gianluca Iannotta @jannotix]

@omkarships Claude code