What's new in
LumiChats.

A big reliability release: document uploads work again in chat, long conversations remember far more, your messages stay put when you switch chats, and Agent mode recovers when a model stalls. New accounts now sign up with Google, and we strengthened protection against abuse.

A quick follow-up to multi-agent orchestration: when you ask for a specific number of helpers ('use 3 agents'), the assistant now always uses exactly that many. Previously it could decide a job was small enough to handle alone, even after you'd asked for several.

LumiChats Desktop can now split a task across multiple helpers working at the same time — ask for several and they genuinely run in parallel, each starting fresh, with a live status line of its own, and the assistant combines their findings into one answer. Helpers can also run in the background while you keep working elsewhere in the app, a new built-in Researcher helper looks things up on the web and reports back with sources, and read-only mode gained its own read-only helpers. Two or more helpers editing the same file can no longer overwrite each other, every safety limit you've set still applies to every helper no matter how it was started, and model routing was refreshed for current prices and availability.

LumiChats Desktop keeps getting faster and more resilient. The Originality Report now scores roughly 8.5x faster, its PDF export finally works on every document, and flagged passages are highlighted right inside the text instead of in separate boxes. Web research is far more reliable — it quietly routes around search-engine blocks using a browser already on your machine, and gives you a clear message instead of a dead end when a search fails or every model is busy. A stalled or confused model is now automatically skipped and retried, so a bad response never wastes your time for nothing. New sign-ins default to Lumi Quasar, a fast and inexpensive tier, alongside several small file-handling fixes.

LumiChats Desktop — our native Windows app — gets a dedicated spot in this changelog for the first time, catching up on ten releases of real fixes and one genuinely new feature. The headline is the Originality Report: score any document for how much of it reads as AI-generated, with a graded meter, flagged paragraphs, and a branded downloadable PDF, now running as its own dedicated tab instead of a chat request. Image understanding in chat, which never actually worked in any prior release, is now fixed and faster. Sign-in, PDF reading, and Word document support are all more reliable, and a run of security hardening closes several file-access edge cases.

Study Mode just got more reliable and more secure. We fixed a rare timing issue where switching PDFs while one was still building its search index could carry the wrong document's content into the new session, closed a gap where a failed quiz generation could leave you on a blank screen with no way back, and made document search fall back to a full scan on longer PDFs so questions about content outside the initial index are never missed. We also patched a security gap in how session usage was verified, and Study Mode's error messages no longer expose internal details.

We nearly doubled the AI model comparison library (435 → 903 side-by-side pages) with 13 in-demand models, gave comparison, glossary, feature and use-case links their own clean share-preview cards, fixed three pages that had no preview at all, broadened the blog's image variety, and made this changelog load smoothly as you scroll.

Claude Sonnet 5 is now available on LumiChats — Anthropic's newest and most agentic Sonnet, with near-Opus quality at Sonnet pricing — and you can pick it from the model menu today. We also published four new guides on the new model lineup, expanded the model-comparison library by 110 head-to-head pages (now 435 total), and refreshed the homepage to reflect the 45,000+ developers and students on the platform.