Какая подписка за $20 даёт БОЛЬШЕ? Google vs Claude vs ChatGPT — Transcript

Full transcript

1. 0:00Claude Pro, ChatGPT Plus, or Google
2. 0:01Gemini Advanced. Each one costs 20 bucks
3. 0:04a month. I don’t need to pay for all
4. 0:06three. I need to figure out which
5. 0:08service is actually worth the money and
6. 0:09which one just nicely charges me twenty
7. 0:11bucks. I gave all three the same task on
8. 0:13a large document. One handled it in 50
9. 0:15seconds and didn’t even notice the load
10. 0:17. The second one thought for 2 and a
11. 0:19half minutes and burned through almost
12. 0:20its entire five-hour limit in just three
13. 0:22questions. And when I switched to its
14. 0:23fast model, it started faking dates in
15. 0:25the text to hide the fact that it wasn’t
16. 0:27reading the whole file.
17. 0:28And the weirdest thing is, based on
18. 0:30official pricing, the first service is
19. 0:32one of the most expensive, but within
20. 0:33the $ 20 subscription, it provides 30
21. 0:35times more resources than the others.
22. 0:37But just when I was about to write off
23. 0:39the loser and cancel the subscription,
24. 0:41this same service brought my photo to
25. 0:43life in a minute and generated this
26. 0:44video with sound. It wrote the
27. 0:46background music you’re hearing right
28. 0:48now and tossed in another 5 TB of cloud
29. 0:50storage.
30. 0:51In this video, I’ll show how many real
31. 0:53tasks fit into each subscription, which
32. 0:55model saves your limit four times over,
33. 0:57and why the service that failed
34. 0:59miserably on documents turned out to be
35. 1:01the best tool for content creators.
36. 1:03With numbers in hand, we’ll find out who
37. 1:05gives you the most for 20 bucks a month
38. 1:06. Let’s go.
39. 1:09I’ll start with what’s in every $ 20 box
40. 1:11, as there is quite a bit in common. All
41. 1:14three have two types of models inside
42. 1:15that we’ll pit against each other today
43. 1:17.
44. 1:18These are the heavy flagships and the
45. 1:20fast, everyday workhorse models.
46. 1:22For Claude Pro, that’s Opus 3.5 and the
47. 1:24workhorse Sonnet.
48. 1:26The top-tier Opus is only available on
49. 1:28more expensive plans, but Claude Pro is
50. 1:30cheaper if you pay for the whole year
51. 1:32upfront. It comes out to 17 dollars a
52. 1:34month, which means 40 bucks of savings
53. 1:36in your pocket. The competitors don't
54. 1:38offer such a discount. ChatGPT Plus has
55. 1:40the flagship GPT-4o and the work model
56. 1:43GPT-4o mini.
57. 1:44And Google Gemini Advanced has the
58. 1:46newest flagship Gemini 1.5 Pro and the
59. 1:48high-speed Gemini 1.5 Flash. At first
60. 1:50glance, all the services offer roughly
61. 1:52the same features in their boxes.
62. 1:54Everyone has a code interpreter, deep
63. 1:56web research, voice mode, and
64. 1:57connectivity to other services. ChatGPT
65. 2:00and Google can generate images. Claude
66. 2:03cannot. Memory for long documents.
67. 2:04Claude and Google have a million tokens
68. 2:06. That’s roughly like a whole book.
69. 2:08ChatGPT, in a standard chat, has from
70. 2:1132,000 to 128,000 tokens depending on
71. 2:14the mode. That’s just like a couple of
72. 2:16chapters. ChatGPT and Google have free
73. 2:18chat options. You don't even need to
74. 2:19touch your subscription for simple
75. 2:21questions. And ChatGPT often resets its
76. 2:23limits. Well, at least for now, you burn
77. 2:26through them, and bam, they’re reset.
78. 2:28But will it stay that way? That’s the
79. 2:30question. Can ChatGPT control a computer
80. 2:32, too?
81. 2:33Google cannot yet. But Google’s box has
82. 2:35something that the other two services
83. 2:37don't even come close to. Well, more on
84. 2:39that a little later. For now, it's
85. 2:40roughly a draw on the box. But what
86. 2:42matters isn't a pretty box, it's how
87. 2:44much real work fits into your limit.
88. 2:48Let me clarify exactly what I’m
89. 2:50comparing.
90. 2:51I won’t be comparing model intelligence
91. 2:53, but how much work fits into the same
92. 2:5420 dollars for different subscriptions.
93. 2:56That’s why this is a text-based test. I
94. 2:58gave all three services the exact same
95. 3:01task. I took a large work document—it’s
96. 3:04the changelog for my Telegram bot,
97. 3:06covering 85 versions and 228 KB. That’s
98. 3:09roughly 80 pages of full text with dates
99. 3:12, prices, file names, and build numbers
100. 3:14. It’s as close to real life as it gets
101. 3:17, like when you upload a large contract
102. 3:18, report, or project documentation and
103. 3:20ask questions about it. The answers can
104. 3:22be verified line by line. So, I can be
105. 3:24sure that the work will be practically
106. 3:26identical. And the difference in limits
107. 3:28is also fair. That wouldn't work with
108. 3:30writing code. Each model writes it in
109. 3:32its own way. Another important detail.
110. 3:34The default ChatGPT ID agent has nuances
111. 3:36you need to know.
112. 3:38ChatGPT cuts off large instruction files
113. 3:41without warning, so I manually raised
114. 3:43the file reading limit to 256 KB so it
115. 3:46would finish reading our document. And
116. 3:49Claude has a one-hour window during
117. 3:51which re-reading the context costs next
118. 3:53to nothing. If you take a break longer
119. 3:55than an hour, it will re-read your
120. 3:57entire context again, consuming your
121. 3:59limits significantly. So, the conditions
122. 4:01for everyone are equal. One document and
123. 4:03a series of identical questions for all
124. 4:05. I’m looking at how long each takes to
125. 4:07think, how much of the limit it uses,
126. 4:08and whether it answered correctly. I
127. 4:10kept the apps open side-by-side, timed
128. 4:12them with a stopwatch, recorded the
129. 4:14five-hour and weekly limit counters
130. 4:16after each question, and then
131. 4:17recalculated everything using token logs
132. 4:19.
133. 4:21Let’s start with the battle of the
134. 4:22flagships. Opus 3.5, Astra 6, and Argon
135. 4:254.
136. 4:26Let’s look at the speed and quality of
137. 4:28the answers. Claude Opus 3.5 answers the
138. 4:30fastest. 50-55 seconds for each large
139. 4:33report. It unfailingly finds the 26
140. 4:36versions from August all three times,
141. 4:38counts the 85 releases in the file, and
142. 4:40identifies the earliest version, 0.4
143. 4:42from May 30th.
144. 4:44Moreover, it was the only one to notice
145. 4:46on its own that the version order was
146. 4:48broken in one place in the file. ChatGPT
147. 4:50Astra 6 thinks a bit longer, 70-75
148. 4:53seconds per question, and also delivers
149. 4:55100%accuracy. Three for three and not a
150. 4:58single error. And what about Google's
151. 5:00flagship, Argon 4? It thinks for 130 to
152. 5:02150 seconds. Two and a half minutes per
153. 5:06question. That’s almost three times
154. 5:08slower. Claude answers the facts
155. 5:09perfectly and is even the most detailed
156. 5:11of all. Though, it ignored the request
157. 5:13to "write two or three lines" and rolled
158. 5:15out entire paragraphs instead. And now,
159. 5:18the most interesting part. Let's look at
160. 5:19the five-hour limit counters. Opus 3.5.
161. 5:22The first hour ate up 2%.
162. 5:24The second, one; the third, also one. A
163. 5:26series of short questions in a row, five
164. 5:28seconds each, didn't move the scale at
165. 5:30all—0%.
166. 5:31Total for the entire round: 4%of the
167. 5:33five-hour limit and 1%of the weekly. For
168. 5:36Astra 1.5, the first hour took 7%.
169. 5:39The second, six; the third, seven.
170. 5:41Another 3%went to five short questions.
171. 5:43Total for the round: 23%of the five-hour
172. 5:46window and 3%of the weekly. Now look at
173. 5:49what happened with Argon 1.5 Pro.
174. 5:51First question: -34%
175. 5:55of the five-hour limit. Second question
176. 5:57: another -33%.
177. 6:00Third question. -31%.
178. 6:03Three questions. Just three messages on
179. 6:06one document burned through 98%of Google
180. 6:09AI Pro's five-hour window.
181. 6:11Plus 2%from before the start, and the
182. 6:13counter simply hit 100%.
183. 6:15There wasn't enough limit left for the
184. 6:17remaining short questions in this window
185. 6:18. So, one of our work rounds requires
186. 6:21about 125%of the five-hour limit from
187. 6:23Google's flagship. It doesn't even fit
188. 6:26into a single five-hour window. By the
189. 6:28way, why is there such a gap between
190. 6:29Claude and ChatGPT? 4%versus 23%.
191. 6:33When I looked at the token logs,
192. 6:35everything fell into place. Claude
193. 6:37caches the large file between new chats
194. 6:39. In the second and third chats, 118,000
195. 6:42tokens are taken from the cheap cache,
196. 6:45while Gemini reads all 48,000 document
197. 6:47tokens anew in each new chat. It turns
198. 6:50out that in 5 hours, Opus can manage
199. 6:52about 75 such tasks. Astra 15, Argon
200. 6:55just three.
201. 6:58It's logical to ask here. But every
202. 7:01subscription has lighter, everyday work
203. 7:03models. Maybe in Google AI Pro, you just
204. 7:06need to switch from the ravenous Argon
205. 7:081.5 Pro to the fast Gemini 1.5 Flash. I
206. 7:11ran all three work models along the
207. 7:13exact same route. I found two things
208. 7:15that will save you a lot of nerves and
209. 7:17limits. The first finding is for those
210. 7:20on ChatGPT Plus. The work model, GPT-4o
211. 7:22, passed the whole test with a result of
212. 7:24three out of three on long questions and
213. 7:26five out of five on short ones. Not a
214. 7:28single loss of facts. The accuracy is
215. 7:30one-to-one with the top-tier Astra. At
216. 7:32the same time, it answered faster—60-65
217. 7:34seconds—and used almost four times less
218. 7:36of the limit for the entire round. Just
219. 7:396%of the five-hour window versus 23%for
220. 7:42Astra. So, for working with large
221. 7:44documents, it makes absolutely no sense
222. 7:46to use the hungry Astra. GPT-4o will do
223. 7:49everything just as well and save your
224. 7:51limits fourfold. Claude Sonnet 3.5
225. 7:54consumption is even lower—just 3%of the
226. 7:56five-hour window and 0%of the weekly for
227. 7:58the entire round at a speed of 40-45
228. 8:00seconds. It wrote out all 26 August
229. 8:03versions perfectly, down to the minutes
230. 8:05for starting all blocks, but in the
231. 8:07third chat, it made a one-digit error in
232. 8:10a line once. It wrote 84 versions
233. 8:12instead of 85.
234. 8:14And now, the second finding. How about
235. 8:16Gemini 1.5 Flash?
236. 8:18It is truly something else. We send the
237. 8:20same prompt to Flash. Boom. The response
238. 8:23pops out in 5 seconds. If you recall,
239. 8:25previous models took 50 or 70 seconds
240. 8:28for this. Usage of the limit is zero.
241. 8:31Well, meaning it's free. But I start
242. 8:33verifying the answer against the file,
243. 8:35and I see a complete failure. Zero out
244. 8:37of three in every chat. In the first
245. 8:40chat, it counted 63 versions in the
246. 8:42document, in the second with the same
247. 8:44question 103, in the third 122.
248. 8:47And there are 85 of them, by the way.
249. 8:49The earliest version it named 1.0.
250. 8:52Makes sense. It simply threw out the
251. 8:54whole chunk of files from May. And the
252. 8:56most brazen part: to fulfill the task of
253. 8:58listing August versions, Flash took
254. 8:59releases from September and October and
255. 9:01stupidly changed the month in the dates
256. 9:03to August. The September version from
257. 9:0617.09 turned into a version from 17.08,
258. 9:08and the October one from 11.10 into
259. 9:1111.08.
260. 9:14Well, you get the idea. However, for
261. 9:16short, specific questions, the very same
262. 9:18Flash answered perfectly—five out of
263. 9:19five. It even instantly found the very
264. 9:22first version 0.4, which it completely
265. 9:24failed to see a second ago. Why is that
266. 9:27? When you drop a large file into a
267. 9:29standard Gemini chat and enable the
268. 9:31lightweight Flash, the app saves
269. 9:32resources.
270. 9:34It doesn't give the model all 80 pages
271. 9:35of the document, but clips out separate
272. 9:37chunks.
273. 9:38That’s why for a specific question, like
274. 9:40"find one digit in the contract," Flash
275. 9:42works great and doesn't drain the limit
276. 9:44at all.
277. 9:45But you cannot ask it to compile a
278. 9:46report across the entire large file. It
279. 9:48will silently throw out half the text
280. 9:50and doctor the dates to fit your prompt
281. 9:52. A tip for Google Gemini Pro owners.
282. 9:54Don't attach large documents via the
283. 9:56paperclip icon in a regular chat. Upload
284. 9:57them to Gemini NotebookLM. With a paid
285. 10:00subscription, you can fit up to 300
286. 10:02sources into one notebook. And there,
287. 10:04working with large texts doesn't burn
288. 10:06through your limit like that.
289. 10:09And now for the most interesting part.
290. 10:11What if we weren't paying $ 20 for a
291. 10:13subscription, but paid for each request
292. 10:15separately at the official developer
293. 10:16rates? Well, I mean via the API. I
294. 10:18extracted the exact number of tokens
295. 10:20from the logs and calculated the cost of
296. 10:22our task based on the official prices of
297. 10:23Anthropic, OpenAI, and Google. Look at
298. 10:25the paradox that was revealed. First.
299. 10:28The same Russian document weighs 1.75
300. 10:30times more in Claude than in ChatGPT and
301. 10:33Gemini. 85,000 tokens versus 48,000.
302. 10:36Second. At official API prices, one of
303. 10:40our work rounds on the flagships costs $
304. 10:431.57 for Claude Opus, $ 2.28 for ChatGPT
305. 10:46-4o, and about $ 1.20 for Google Gemini
306. 10:49Pro.
307. 10:50But a subscription is like a buffet at a
308. 10:52restaurant. It doesn't matter how much
309. 10:54one dish on the menu costs; what matters
310. 10:56is the size of the plate you're given at
311. 10:58the entrance for your $ 20.
312. 11:00You see, if you divide the API cost of a
313. 11:02task by the percentage used of your
314. 11:04five-hour limit, we see the real
315. 11:06measurement of this "plate" for each
316. 11:08service. In a five-hour window, Claude
317. 11:11Pro on Opus 3.5 offers about $ 39 worth
318. 11:14of work at API prices. You can complete
319. 11:1725 such full rounds in a single
320. 11:19five-hour window. The weekly limit holds
321. 11:22about $ 150 worth of value. In a
322. 11:24five-hour window, ChatGPT Plus on GPT-
323. 11:274o offers $ 10 worth of work at API
324. 11:29prices. That is just over four rounds on
325. 11:31the flagship model or 16 rounds on the
326. 11:33GPT-4o mini model.
327. 11:35And in a five-hour window, Google Gemini
328. 11:37Advanced with Gemini 1.5 Pro offers only
329. 11:40$ 1.50 worth of work via API.
330. 11:42Think about it: all three subscriptions
331. 11:44cost the same $ 20, but the plate for
332. 11:46the flagship Claude Pro holds nearly 30
333. 11:48times more work than Google Gemini
334. 11:50Advanced.
335. 11:51Thirty times.
336. 11:53So, returning to the buffet analogy,
337. 11:55Google's food is inexpensive. But the
338. 11:58plate you're given to put food on is
339. 12:00just very small. Subscription vs. two
340. 12:03different economies. For subscriptions,
341. 12:05Claude is the most profitable. API-wise
342. 12:07, it's Google and ChatGPT.
343. 12:11And here you might ask: "Wait a minute.
344. 12:14If Gemini burns through a third of its
345. 12:16limits in one request, and its fast
346. 12:18model is constantly hallucinating, why
347. 12:20even get the $ 20 Google subscription?"
348. 12:23Well, now look at it through the eyes of
349. 12:25someone who creates content, runs a
350. 12:27YouTube channel, shoots Reels, Shorts,
351. 12:29makes ads, presentations, or manages
352. 12:31social media.
353. 12:32In these tasks, Google makes a complete
354. 12:34comeback because, within that one $ 20
355. 12:36subscription, you get a set of powerful
356. 12:38services with separate balances that
357. 12:40don't touch your main limit at all.
358. 12:42First. Music and song generation with
359. 12:45MusicFX. Claude and ChatGPT have no
360. 12:47music generation capabilities whatsoever
361. 12:49. Like, none at all. But with the Google
362. 12:52Gemini Advanced subscription, you get
363. 12:54access to the music model, not sure how
364. 12:56it's pronounced, and a separate MusicFX
365. 12:58service with its own supply of credits.
366. 13:01I tested it for my videos, and I really
367. 13:04liked it. Here are three tracks I made
368. 13:06in just a couple of minutes. The first
369. 13:08track is that calm electronic background
370. 13:10music that played at the beginning of
371. 13:11our video.
372. 13:13Steady volume, soft synthesizer, fits
373. 13:14perfectly under a voice-over and motion
374. 13:16graphics. The second track is a lively,
375. 13:19energetic instrumental for dynamic
376. 13:20transitions and screen recording. And
377. 13:22the third track is a full three-minute
378. 13:25song with great female vocals and a
379. 13:27powerful dance chorus.
380. 13:29All of this downloads as an MP3 and can
381. 13:31be added to your video without fear of
382. 13:33copyright strikes or needing separate
383. 13:35paid subscriptions.
384. 13:36Second. After OpenAI removed Sora from
385. 13:39the $ 20 ChatGPT Plus, Google became the
386. 13:41only one of the Big Three that provides
387. 13:44video generation with audio in a $ 20
388. 13:46subscription. Just look at how it works
389. 13:48in tandem. I'm sitting at a desk
390. 13:50recording a video for you, I give the
391. 13:52neural network my photo, and now I'm
392. 13:54standing up and dancing to hip-hop.
393. 13:56Google's work for Shorts and vibrant
394. 13:58inserts that keep the viewer's attention
395. 14:00is just fire.
396. 14:01Third. Argon can watch videos directly.
397. 14:06I fed all three flagships my finished
398. 14:08ten-minute YouTube video and asked 10
399. 14:10tricky questions about details that
400. 14:12couldn't be answered from a transcript
401. 14:14alone.
402. 14:16You literally have to examine the frames
403. 14:17. Argon 4 swallowed the video file
404. 14:19directly and answered faster than
405. 14:20everyone else, in one minute. Asra and
406. 14:22Opus took 4 minutes each. I have to tell
407. 14:25you, the way Google sees videos is
408. 14:26something that I think no other model
409. 14:28even comes close to. And add to this 5
410. 14:30TB of cloud storage for Google Drive,
411. 14:33mail, and photos. And you get a very
412. 14:35competitive service. Anyone who shoots
413. 14:37video knows how fast raw footage fills
414. 14:39up memory.
415. 14:40A separate 5 TB cloud costs more than
416. 14:42half of this subscription.
417. 14:45Here, it's already included in your $ 20
418. 14:47, and it's also shared with the whole
419. 14:48family, as is this entire subscription.
420. 14:52Now, the fly in the ointment. Most of my
421. 14:54viewers are in Russia, and some Google
422. 14:56features are closed to us. Well, more
423. 14:59accurately, most of them. Google's
424. 15:00antigravity program for working with
425. 15:02code. Well, it's an analogue to Lootcode
426. 15:04, and the codec won't let you in with or
427. 15:06without a VPN. Flow studio for video.
428. 15:09The same thing. I generated videos in Y
429. 15:11Studio. YouTube Premium Lite, which
430. 15:13comes as a bonus, cannot be activated
431. 15:15here at all. But Gemini itself, although
432. 15:17via VPN, still works. And everything you
433. 15:19saw in this video—the video, music, and
434. 15:21the breakdown of my clip—was done
435. 15:23exactly there.
436. 15:26Let's summarize. If you need work,
437. 15:27documents, code, or a long conversation
438. 15:29with a neural network, use Claude Pro.
439. 15:31Its limits accommodate the most. It is
440. 15:33the fastest and makes no mistakes. For
441. 15:36daily tasks, take Sonnet, but don't
442. 15:37forget about the hourly window. If you
443. 15:40need images and a flagship at hand, use
444. 15:42ChatGPT Plus, but it's better to work on
445. 15:43CO. Keep Argon only for complex tasks.
446. 15:46Create content with Google AI Pro. Video
447. 15:48, music, video breakdown, and 5 TB for
448. 15:51raw footage. Working in it normally
449. 15:53won't be possible. The limit is enough
450. 15:54for three serious tasks, but you can
451. 15:56create as much as you want. I decided
452. 15:59for myself: I work in Claude. I create
453. 16:01in Google. I have compiled the entire
454. 16:03summary table with test measurements and
455. 16:05brief conclusions into one convenient
456. 16:07cheat sheet. You can get it for free in
457. 16:09my Telegram bot via the link in the
458. 16:10description. In the channel itself, I
459. 16:12post the results of my findings, such as
460. 16:14Claude's hourly window or the codec's
461. 16:16file size limit. Write in the comments
462. 16:18what subscription you use and what you
463. 16:20think of my setup.
464. 16:25J