I keep a running file of assistant comparisons, and the September 2026 digest of Russian-language chatbot tests earned a spot in it. It put 15 chatbots side by side — GPT, Claude, Gemini, DeepSeek among them — and scored them on speed, accuracy, and slang understanding. No overall winner was named, which is probably the most honest outcome a test like this can produce.
The numbers behind the comparison
The digest landed on 21 September 2026 and framed the ranking with a stat that is easy to scroll past: roughly 35 million people in Russia — about one in four — already use neural networks, which sits above the global average. That context matters. When a quarter of a country's population is already using these tools, a ranking stops being a curiosity and starts being a buying decision.
The three criteria were deliberately unglamorous:
| Criterion | What it tells you |
|---|---|
| Speed | Whether the model fits into a working session or breaks it |
| Accuracy | Whether the answer survives being used |
| Slang understanding | Whether it reads real human input, not textbook input |
No per-model scores were published, so I treat the list as a shortlist rather than a league table.
ChatGPT: GPT-6 splits into Sol and Luna
In late September 2026, OpenAI divided GPT-6 into two working models, Sol and Luna. Sol is the cheaper one, with prices reduced by 50%, positioned for routine text work — summarizing email threads, drafting copy, that kind of thing.
The material doesn't spell out what Luna is for, but the pattern is clear enough: instead of shipping one universal model, providers are slicing the lineup by budget and workload. For anyone planning spend, that changes the question from "which model?" to "which tier for which task?" — send the boring 80% to the cheap variant and keep the heavier model for the parts that actually need it.
Claude: the practical news came from GitHub
Claude was part of the 15-chatbot comparison, but the concrete Claude story came from a developer on GitHub rather than from the lab. ClaudePanel is a local desktop panel for AI conversations: it runs on your machine, needs no elaborate setup, and sits as a separate window above your other applications.
If your day is a loop of ChatGPT tab, Claude tab, back to ChatGPT tab, this is a small but real quality-of-life change. A dedicated window keeps the conversation visible while you work in something else, instead of buried at position fourteen in the tab strip.
Gemini, DeepSeek, and the long tail of 15
Gemini and DeepSeek both made the top-15 cut. The individual scores weren't published — only the criteria — so the useful takeaway isn't a placement, it's the shape of the market. There are now at least 15 serious chatbots in play, and these two are standard options rather than exotic alternatives you have to justify picking.
That also means the choice is no longer binary. If a model stalls on your particular workflow, you have more than one fallback that won't embarrass you.
The YandexGPT test that burned 300,000 tokens
Not every run ends well. In a real-task test, YandexGPT got stuck in a loop for 37 minutes and consumed 300,000 tokens without producing anything usable. The testers were explicit about their method: they don't grade a model on a static sample — they hand it a real task and watch what happens.
That is the part I'd underline. Speed and answer quality are the metrics everyone quotes, but stability and cost control belong on the same list, especially when every token has a price attached. A model that loops is not just slow; it is an invoice.
Jev and the "typed decisions" alternative
While the big assistants compete on conversation, TypeSafe AI came out of stealth with Jev, which it describes as the first "System One" model. Jev doesn't generate text at all. It returns typed decisions with probabilities.
That's a genuinely different philosophy from ChatGPT, Claude, Gemini, and DeepSeek. No chat, no prose — structured output you can wire into a system. It probably won't replace a conversational assistant for everyday work, but it hints at where a slice of the field is heading: models that answer in data structures instead of paragraphs.
Takeaway
What I'm carrying forward is a three-part check when I pick an assistant for a task: does it fit the work (speed, accuracy, slang and domain understanding), what does it cost per tier, and how does it behave when something goes wrong. The YandexGPT loop is the reminder that the third question is not optional — 37 minutes and 300,000 tokens is a failure mode with a bill attached.
This article was originally published on AI Release — https://ai-release.org , where I keep tracking these comparisons; the channel at https://t.me/ai_release1 is where the shorter notes land.
Top comments (0)