AI Release DeepSeek did the task in 1.5 minutes. YandexGPT Pro 5 — 37 minutes, 3,246 ₽, and zero results. We set both models up as agents on the same routine task — checking a website's indexing. Connected via the OpenAI-compatible API. For Yandex we took their most expensive model — Pro 5. DeepSeek V3: — 1.5 minutes — pennies in tokens — task solved YandexGPT Pro 5: — 37 minutes — 295 consecutive tool calls — 296 context rebuilds — 5.6 million cached tokens — 3,246 ₽ charged (with a discount; without it — 7,246 ₽) — task NOT solved The model went into a loop: tool call → context rebuild → call again. It couldn't exit on its own. We stopped it manually. Now about the money: At list price, YandexGPT Pro 5 is $12 per 1M tokens. DeepSeek V3 — $0.80. A 15x difference. But that's list price. In reality, the difference is infinite, because DeepSeek solved the task and Yandex didn't. Moreover, 93% of Yandex's costs were cached tokens. In other words, the model just kept re-reading the same context in circles. The cache that's supposed to save money became the mechanism draining it. The most disgusting part: Yandex has no protection against looping. No "stop after N identical calls," no repeat detector — only your manual abort. If the model enters a loop — that's your money, not their problem. But attaching a card and charging for every step? Sure, that works perfectly. What to do: — Hard limit on agent steps. Always. — Watch the billing: a spending spike with no progress = a loop. — Set a cost ceiling before launching. Yourself. The vendor won't do it. — Keep cached tokens on your radar. They're exactly what silently eats the budget. — Calculate the price per task, not per million tokens. We've gone back to models that don't loop. We left YandexGPT for short queries where looping is impossible in principle. Caught YandexGPT looping? Write in the comments or DM — we'll compile a collection. Full breakdown with numbers — on our website. Subscribe: @ai_release1 #YandexGPT #DeepSeek #LLM #agents #tokens #billing 5 views edited 21:14
Source: t.me · post in Telegram