Tech • AI • Robotics • Game
Gemini 4 Argon hits 1M+ tokens as Base44 and Claude ship apps
Briefing
Gemini 4 Argon targets long builds
Google is positioning Gemini 4 Argon as a reasoning-heavy model for coding, document analysis and multi-step technical work. Its headline feature is a context window of more than 1 million tokens, intended to keep large codebases, long PDFs and extended planning sessions in a single thread. Argon also exposes low, medium and high reasoning modes, alongside function calling, code execution, structured outputs and search grounding. The pitch is less about chat and more about sustained execution inside developer and API workflows.
Google pushes Argon benchmark claims
Google says Gemini 4 Argon ranks first on the Valves Index, a benchmark spanning finance, coding, legal and tax tasks. That framing suggests the company wants Argon seen as a professional knowledge-work model rather than a consumer assistant. But the result remains a vendor-reported claim and still needs broader independent validation before it can be treated as settled. For buyers, the practical question is whether those scores translate into reliable performance in production agents and internal tools.
Base44 stress-tests Argon with Lumen
A more concrete test came from building Lumen, a journaling app created in Base44 with Gemini 4 Argon. Instead of judging a single polished prompt, the project expanded across six prompts to see whether the model could preserve earlier decisions while adding features. The app grew from a stylized editor into a fuller product with search, tags, favorites, mood filters, a calendar and streak tracking. That makes Lumen a more useful case study in iterative product building than one-shot UI generation.
Lumen shows agent memory under pressure
Lumen also tested whether a model can sustain coherence across design, database and feature changes without collapsing into inconsistency. The app kept an Apple-style liquid glass interface while adding AI-assisted writing, authentication, subscriptions and mobile use. That matters because many coding agents still excel at first drafts but drift as requirements accumulate. In that sense, Base44 became a proxy benchmark for product memory, not just frontend polish.
Claude assembles a SaaS replacement stack
Claude was used to outline and build a custom online training platform intended to replace subscription tools such as Podia or Systeme.io. The stack centers on GitHub for version history, Cloudflare for hosting and DNS, Convex for the database and Tella for video hosting. The resulting product scope includes a sales page, member area, lessons and admin functions. The broader claim is that AI coding can now substitute for a meaningful slice of small-business software spend.
Cloudflare and GitHub anchor no-code ops
The training-platform workflow is aimed squarely at non-technical operators, but its infrastructure choices reveal where manual discipline still matters. GitHub is treated as essential because it provides rollback and auditability when an AI agent breaks working code. Cloudflare becomes the public deployment layer, while Convex handles application data such as users, courses, lessons and progress. The message is that AI lowers build friction, but dependable operations still depend on conventional developer tooling.
Reverse Engineer Anything shifts desktop cloning
Reverse Engineer Anything stood out for a more ambitious claim: helping AI replicate or adapt desktop applications by inspecting compiled software rather than public webpages. The tool installs as a CLI and can analyze native binaries, Go, Rust, JavaScript, SQLite and some mobile targets to extract interface and runtime clues. That is a materially different proposition from ordinary browser-based site copying. If it works well, it could give coding agents a new path into legacy software modernization and competitive feature reconstruction.
Website clone hype meets Presentify reality
The sharpest editorial takeaway is that reverse-engineering hype looks overstated when applied to websites. Modern coding agents can already inspect and recreate public sites like tella.tv without a specialized binary-analysis layer, making flashy demos less novel than advertised. The more meaningful test was a Mac app, Presentify, where source code is unavailable and behavior has to be inferred from the shipped product. That comparison suggests the real value of Reverse Engineer Anything lies in local software analysis, not in website cloning theater.
Videos covered
J'ai testé le repo GitHub viral qui reverse engineer n'importe quoi
- •What the tool is built for
- •Why website demos are misleading
- •A practical test on Presentify
Gemini 4 Argon Full Tutorial for Beginners: How to Use (Like a Pro)
- •Designed for larger, multi-step work
- •Large context window and reasoning controls
- •Benchmark claims and caveats
Gemini 4 Argon - Full Tutorial: How To Build an App In Minutes
- •Testing beyond the first prompt
- •Base build and visual system
- •Organization features added next
Formation : Comment créer ton "SaaS personnel" en moins de 30 minutes (plateforme de formation)
- •AI-built software instead of subscriptions
- •Four main services form the stack
- •Non-technical users are the target