Anthropic can't stop winning — Transcript

Full transcript

1. 0:00I'm not going to lie, it kind of feels
2. 0:01like Anthropic just can't stop winning.
3. 0:03Recently, they've been on quite a tear
4. 0:05with all of the recent model releases
5. 0:06blowing us away, price changes that make
6. 0:08a lot of sense, and generally just not
7. 0:11being as annoying to work with. Today is
8. 0:13another step in that direction because
9. 0:14Anthropic has just changed their plans.
10. 0:16They've just introduced the ultra tier
11. 0:18to their subscriptions. And don't worry,
12. 0:20they're not here to charge a whole bunch
13. 0:21more money. It's the exact same plan as
14. 0:23the 20X was before, that is the $200 a
15. 0:26month plan, but they made some subtle
16. 0:27changes to it that are really worth
17. 0:29noting. In particular, they removed the
18. 0:315-hour limit. This is just a great
19. 0:33change all around, right? Like there
20. 0:35isn't much to cover here. I should be
21. 0:36able to just wrap this video up fast.
22. 0:39Oh, for verified ultra accounts.
23. 0:43>> [sighs]
24. 0:43>> You never make this easy for me, do you,
25. 0:45Anthropic? It wouldn't be an Anthropic
26. 0:47release without at least a little bit of
27. 0:48weirdness, right? That verified callout
28. 0:51is part of it. The way that they enforce
29. 0:53the 5-hour limits is another part of it
30. 0:55because remember, they're only removing
31. 0:56this from the $200 plan, even though
32. 0:58those 5-hour limits tend to hurt a lot
33. 1:00more on the cheaper plans. Ah, there's a
34. 1:03lot to dive into here, and I'm going to
35. 1:05do my best to cover it all and also talk
36. 1:07about how this affects us with T3 Code,
37. 1:09which sadly, this does directly affect
38. 1:11us with T3 Code. Don't worry too much,
39. 1:13though. We're going to be good. Either
40. 1:15way, we got bills to pay, though, so
41. 1:16let's take a quick break for today's
42. 1:17sponsor. If you love bugs and buggy
43. 1:19software, you should skip this ad. But
44. 1:21for everybody else, you should really
45. 1:22consider using Grappl. Grappl's AI code
46. 1:24review platform is a lot more than just
47. 1:26code review. If I'm being honest, it
48. 1:27feels like they're the only ones who
49. 1:28really understand what we want as
50. 1:30developers. If you've ever had to pull
51. 1:31down a pull request just to make sure
52. 1:33the changes worked, you'll understand
53. 1:34why T-Rex is so cool. It allows their
54. 1:37review agents to run your code in a
55. 1:38sandbox to make sure the changes are
56. 1:40actually valid and working. If you've
57. 1:41ever had to keep your docs and code base
58. 1:42in sync, you'll love their new knowledge
59. 1:44base product that allows you to do that
60. 1:46automatically. As your code changes, the
61. 1:48knowledge base does, too, and it's all
62. 1:49fully editable if you go in and want to
63. 1:51make changes. It's also exposed over
64. 1:52MCP, so your code agents can do better
65. 1:55at making changes because they have all
66. 1:56of the knowledge available to them
67. 1:57through the MCP exposed by Grappl. And
68. 2:00this one is small, but as I'm watching
69. 2:01my cost balloon, I really appreciate it.
70. 2:03They have three different tiers of
71. 2:05review in Graptile. So, if you want to
72. 2:07just quickly get some feedback, you
73. 2:08could use their base tier, which is
74. 2:10relatively cheap through their credit
75. 2:11system. Or, if you want super thorough
76. 2:13reviews for those big scary 20-plus file
77. 2:16PRs, that's where the Apex tier comes
78. 2:18in. You get to pick what level of review
79. 2:20you want when you're working with
80. 2:21Graptile, which allows you to right-size
81. 2:23the effort based on what changes you're
82. 2:25making. Everybody shipping faster
83. 2:26nowadays, do it with fewer bugs and more
84. 2:28practically at swadev.link/graptile.
85. 2:30It is worth noting that part of this
86. 2:32name change is to clean up some
87. 2:34misleading things that existed before.
88. 2:37Specifically, that the 5x and the 20x
89. 2:40aren't actually reflective of your
90. 2:41overall limits. These are just for the
91. 2:445-hour limit. Because, believe it or
92. 2:46not, back in the day when these plans
93. 2:48were introduced, the 5-hour limit is all
94. 2:49they really cared about. And that was
95. 2:51what people were hitting all the time
96. 2:53and why they wanted upgrades. The weekly
97. 2:54limits were only introduced as our usage
98. 2:56patterns changed fundamentally, and they
99. 2:58wanted to keep up the super generous
100. 3:01usage that we got when we were used to
101. 3:03using these with like Sonic 4 in the OG
102. 3:05Claude code CLI. The $20 plan was meant
103. 3:08to be the like all you can eat buffet
104. 3:09style, like super, super generous plan.
105. 3:12It still is very generous, like absurdly
106. 3:14so. But, that generosity is because they
107. 3:17set a certain bar for how much work you
108. 3:18could do, and people wanted to still
109. 3:20have that without feeling restricted
110. 3:22under the same limits. Mike, who's the
111. 3:24head of product for Claude subs, made a
112. 3:26really good follow-up post here about
113. 3:28this. He also went on about how when
114. 3:29they introduced the $20 plan in April
115. 3:31last year, April 2025, it only had the
116. 3:345-hour limit. And the point of this plan
117. 3:36was that it gave you 20 times the amount
118. 3:38in those 5 hours that the Pro plan did.
119. 3:40So, they called it Max 20x. And when
120. 3:42they added a weekly limit to Max 20x,
121. 3:44they set them to be 10 times the limit
122. 3:46of Pro. Many people assumed that the 20x
123. 3:48applied to both limits. They should have
124. 3:50done better, and he personally
125. 3:51apologized. Actually really nice to see.
126. 3:53This it it just feels much more human
127. 3:55than I'm used to from Anthropic. Good
128. 3:57work, guys. This is an improvement. The
129. 3:59new naming is much clearer. They are
130. 4:01very specific about these limits in all
131. 4:03of their pages and everything now. Like
132. 4:05it just directly says it's 10 times more
133. 4:07than Pro, which is what it is. Awesome.
134. 4:09Cool. So, what's next?
135. 4:11There are four subscription plans for
136. 4:12Claude: free, Pro, Max, and Ultra. Free
137. 4:14is of course free, Pro is 20 bucks a
138. 4:16month, Max is 100, and Ultra is 200.
139. 4:18They all offer different limits for
140. 4:20people who use Claude different amounts.
141. 4:22And none of them are enough for super
142. 4:23heavy power users like myself balancing
143. 4:26seven different Claude accounts. But for
144. 4:27most people, the $200 account will be
145. 4:29more than enough for day-to-day
146. 4:30realistic dev work. Hell, even some
147. 4:32heavier stuff would be totally fine in
148. 4:34Ultra. When I turned off all my hill
149. 4:36climbs, I struggled to even do half an
150. 4:37account with Ultra right now, if I'm
151. 4:38being real. Max is also quite generous
152. 4:40at the $100 offering there, but it does
153. 4:42still have the 5-hour limits. We should
154. 4:44talk about those really briefly. Claude
155. 4:46subs have two limits by default. They
156. 4:48have the weekly limit, which is how much
157. 4:50Claude you can do in a given week before
158. 4:52you have to start paying API prices. And
159. 4:54separately, they've a 5-hour limit that
160. 4:55keeps you from using the whole thing up
161. 4:57in one burst. Why do they have that
162. 4:595-hour limit that's separate from the
163. 5:017-day limit? The reason is because of
164. 5:03compute availability. The level of usage
165. 5:05that the subscriptions give you is
166. 5:07absurd when you consider how expensive
167. 5:09the models can be over API. The $200
168. 5:12plan gets you as much as 12,000 dollars
169. 5:14a month in inference, which means if you
170. 5:16burn that all at a specific time, it can
171. 5:18cost Anthropic a lot of money. Not
172. 5:20because it's just like melting money in
173. 5:22their bank accounts. It's not like every
174. 5:23time you send a prompt, it costs them a
175. 5:25certain amount of money. Their compute
176. 5:26already exists, and it's either being
177. 5:28used or isn't. So, as long as our usage
178. 5:30in the subscriptions does not take away
179. 5:32from availability that they could sell
180. 5:34to enterprises, all is good and dandy.
181. 5:36The problem is the way that usage occurs
182. 5:38over the day. I just ran the numbers
183. 5:40against our analytics for T3 Code to
184. 5:42give you an idea of how much this varies
185. 5:44throughout the day. On a given weekday,
186. 5:46there's a lot of variance, but it really
187. 5:47starts to spike around
188. 5:50say like 4:00 a.m. Pacific time during
189. 5:52the weekdays. Absolute peak of prompts
190. 5:55being sent is between 7:00 and 9:00 a.m.
191. 5:58Pacific time. And then, as we approach
192. 6:01the evening from 6:00 p.m. up until like
193. 6:03midnight or so, you have a huge lull.
194. 6:06There is a 2x gap in how much prompting
195. 6:09is being done from the lowest lows and
196. 6:12the highest highs just in normal
197. 6:14variance of a given weekday. I bring
198. 6:16this up because uh what happens if you
199. 6:18spend whole of your usage for your week
200. 6:21right here in the peak? Eventually,
201. 6:24you're going to hit a point where the
202. 6:25amount of available computing Prophetic
203. 6:27has is just not enough for them to serve
204. 6:29these customers and then everybody sees
205. 6:32these not enough compute available
206. 6:33errors and starts getting upset. So, the
207. 6:365-hour limits were introduced as a way
208. 6:38to prevent a subscription user from
209. 6:41taking more than a certain percentage of
210. 6:43their weekly sub in any one of these
211. 6:45work windows. If you broke up this chart
212. 6:48into 5-hour chunks, the most a user
213. 6:51could do would be I don't know, like 10
214. 6:52to 15% of their weekly limit in any one
215. 6:54of those chunks, which massively impacts
216. 6:57the amount of available compute that
217. 6:59Anthropic has when they are trying to
218. 7:01sell it during the workdays. Back in the
219. 7:03day, Anthropic even went as far as
220. 7:04giving you worse limits during these
221. 7:06windows and more generous limits outside
222. 7:09of these windows because of how bad the
223. 7:11compute scarcity was affecting them. And
224. 7:13the other harsh reality is that during
225. 7:15these late-night hours, there's probably
226. 7:17a lot of GPUs they're paying for that
227. 7:18are sitting not doing much. They don't
228. 7:20mind when you burn during these windows.
229. 7:23In fact, if you take the time to go
230. 7:24through all of the resets that we've
231. 7:25been given for Codex subs, you'll notice
232. 7:28they tend to be either late at night or
233. 7:31on weekends because that is when they
234. 7:33have the most compute available, where
235. 7:35the compute isn't being used heavily by
236. 7:37the companies that they sell to. Because
237. 7:38it's also worth noting that these $200
238. 7:40subs don't make the labs much, if any,
239. 7:42money. The point's to get us addicted to
240. 7:44using 12 grand of tokens a month so that
241. 7:46when we go to our employer, we're
242. 7:48expecting to spend a similar amount
243. 7:49there and they'll have to pay those API
244. 7:51prices. There is a small benefit to the
245. 7:535-hour limit to us as users. Makes it
246. 7:55slightly harder to accidentally overrun
247. 7:57your whole subscription. I know I and
248. 7:59others have had this silly mistake that
249. 8:01like, let's be real, any us can make,
250. 8:03where we send a prompt we think will be
251. 8:04resolved quickly, leave, come back, and
252. 8:07it's just burned an insane amount of
253. 8:08tokens that we never thought it would
254. 8:10have. The 5-hour limit helps a little
255. 8:11bit with preventing that, which is nice,
256. 8:13but overall, I'm happy to see it gone,
257. 8:16at least from the verified Ultra
258. 8:18accounts. Uh
259. 8:20yeah, we'll talk about the verification
260. 8:22thing in just a sec, but first we have
261. 8:24to time travel a little bit all the way
262. 8:26back to May. Feels like years ago that
263. 8:29this happened. It's crazy it was only a
264. 8:30few months. This was when Anthropic
265. 8:32announced their partnership with SpaceX
266. 8:35for compute. This was a massive
267. 8:36multi-billion dollar deal that allowed
268. 8:39for Anthropic to borrow SpaceX's GPUs
269. 8:41from their old data center that they
270. 8:43weren't using as heavily, and I don't
271. 8:44think they're using it at all. I'm
272. 8:45pretty sure Anthropic just runs that
273. 8:46whole data center now. And when they did
274. 8:48this deal, suddenly Anthropic had access
275. 8:50to more GPUs and more compute, so they
276. 8:51could run more Claude. And with that,
277. 8:53they made a huge change. They doubled
278. 8:55the Claude code 5-hour rate limits, and
279. 8:58they removed the peak hour limit
280. 9:00reductions for Claude code. The reason
281. 9:02they could do this is, again, compute
282. 9:04was available finally. They had to have
283. 9:06these limits to restrict how much we
284. 9:09could abuse Claude in a given hour. To
285. 9:11be clear, this wasn't restricting how
286. 9:12many dollars we could spend in terms of
287. 9:14API prices. This was limiting how much
288. 9:16GPU we could use during those work hours
289. 9:20where they needed those GPUs to sell
290. 9:22enterprise inference to these bigger
291. 9:23companies. They don't care about how
292. 9:25much money our usage costs. They care
293. 9:27about how much GPU they have available
294. 9:28for enterprises. So, when they had more
295. 9:30GPU available, they got way more
296. 9:31generous with their limits.
297. 9:33Note the wording here though, cuz we're
298. 9:34going to come back to this. Doubling
299. 9:36Claude code's 5-hour limits. We'll be
300. 9:39back. Because now we have to talk about
301. 9:41the verified aspect here. Why are they
302. 9:44limiting this to verified Ultra
303. 9:46accounts, and what even is a verified
304. 9:48account? They specify a little bit here,
305. 9:50saying that this starts today with most
306. 9:51existing subscribers. Newer accounts
307. 9:53keep a temporary limit until they're
308. 9:56verified, which helps us prevent fraud
309. 9:58and abuse. You'll get an email when it's
310. 10:00removed. Yeah, I don't love this. I
311. 10:04specifically don't love the lack of
312. 10:06transparency here. They have been truly
313. 10:08like legendarily vague about what
314. 10:11verification even is. They say they
315. 10:13remove the limit automatically when your
316. 10:14account is verified, so there's nothing
317. 10:16you need to do. Verifying your phone
318. 10:18number at checkout is a separate step
319. 10:19and does not remove this limit. To
320. 10:21check, go to settings usage. If you
321. 10:23don't see a current session bar, your
322. 10:25account is verified. Yeah, not great. It
323. 10:28would be really nice for them to give a
324. 10:29little more detail on how this works.
325. 10:33But when you understand why they did it,
326. 10:35I think you'll understand why they
327. 10:36can't. One of the biggest issues
328. 10:38Anthropic has to deal with right now is
329. 10:40distillation. In order to do
330. 10:42distillation, you need a lot of
331. 10:44real-world requests and responses from
332. 10:47real-world Claude usage. One of the best
333. 10:49ways to generate this is to get an API
334. 10:51key. And if you can't do those because,
335. 10:53I don't know, maybe your country is
336. 10:55sanctioned and you're not allowed to do
337. 10:56that, or you don't have the money and
338. 10:58you're trying to save because again,
339. 10:59$200 in, $12,000 out. It's a lot cheaper
340. 11:03to use subs. So, as you can probably
341. 11:05guess, there's a lot of interest from
342. 11:06these Chinese labs and other companies
343. 11:08that want to do the distillation to get
344. 11:10these subs, burn them as quick as
345. 11:12possible, generate a ton of responses,
346. 11:14and then use that for distillation. The
347. 11:16removal of the 5-hour limit does make it
348. 11:18slightly easier to do that, which is why
349. 11:20they want to enforce things more
350. 11:22strictly this way, but it's also just
351. 11:24weird and hits users that are doing
352. 11:27things the honest way a little bit too
353. 11:29hard. From my calculations, you can do
354. 11:31roughly 30% of your weekly limit through
355. 11:33one of those 5-hour windows, so you get
356. 11:35three and a half or so 5-hour limits in
357. 11:38a given week. But here's where we get to
358. 11:40the final catch, that hint I gave
359. 11:42earlier in the video. Remember when I
360. 11:44pointed out they specified that they
361. 11:45were doubling Claude codes' 5-hour rate
362. 11:48limits? They meant what they said.
363. 11:50And what they meant was only Claude
364. 11:52code. This has been incredibly annoying
365. 11:56for us to figure out because as people
366. 11:58have started using T3 Code more heavily,
367. 12:00they have identified the fact that their
368. 12:025-hour limits burn much faster in apps
369. 12:05like T3 Code. This user pointed out that
370. 12:07T3 Code uses around 1.8 times more of
371. 12:10the 5-hour window per token than the
372. 12:13interactive CLI does. Even {dash} P does
373. 12:16the same thing, by the way. Incredibly
374. 12:18annoying. This is an issue that's opened
375. 12:20on the Anthropic repo that I became
376. 12:22aware of because of how many people are
377. 12:24reporting it to us in T3 Code. I did my
378. 12:26own calculations to confirm this, and it
379. 12:28is indeed true. I've been working with
380. 12:30Anthropic on this because I was hoping
381. 12:33it would get fixed. I patiently waited,
382. 12:35and they did this change, which to be
383. 12:36fair, for a significant portion of our
384. 12:38users, does resolve the problem because
385. 12:40anybody on the $200 plans no longer has
386. 12:425-hour limits. So, you're good. We're
387. 12:45done. We're This is solved now, right?
388. 12:47Well, the $20 and $100 tier people are
389. 12:50extra screwed because their 5-hour
390. 12:52limits are worse, which already sucks.
391. 12:54And when you combine that with the 2x
392. 12:55decrease for agent SDK users, it's just
393. 12:58garbage. I am not fond of this. I've
394. 13:00been trying to get Anthropic to do
395. 13:01something about it for a minute, and uh
396. 13:04not had the best luck, if I'm being
397. 13:05real. I was hopeful. I've lost hope,
398. 13:07which is why I'm going to tell you guys
399. 13:09what I do. I already see the comment
400. 13:11happening, so I'm going to jump on it
401. 13:12now, but Theo Open Code has the Claude
402. 13:15sub. How are they doing it without agent
403. 13:18SDK? I was really excited about this
404. 13:19release. I thought this was Anthropic
405. 13:21finally being chill and letting Open
406. 13:23Code integrate Claude directly. I was
407. 13:25wrong. They're using agent SDK under the
408. 13:27hood, too. If you are using Claude subs
409. 13:29in your apps, you either have to have
410. 13:32your app just call the Claude Code CLI
411. 13:34directly with like the terminal visible,
412. 13:36or you have to use the agent SDK, which
413. 13:38means you're going to be hit with this
414. 13:3950% decrease in the 5-hour limits.
415. 13:43Obnoxious. I really hope they decide to
416. 13:45fix this instead of causing more
417. 13:47problems. I will say before I go any
418. 13:49further, I cannot be held responsible
419. 13:51for any account bans. Okay, now that
420. 13:53said, there's some very easy solutions
421. 13:54here. Solution one is there is a field
422. 13:56called Claude Code entry point that we
423. 13:58currently have set to SDK-TS for
424. 14:01TypeScript for the official TypeScript
425. 14:03agent SDK. If you change this to
426. 14:05literally anything, the problem goes
427. 14:07away. There are lots of different ways
428. 14:08to set this. Hypothetically speaking,
429. 14:11you can make this problem go away pretty
430. 14:12easy in this certain JSON file that will
431. 14:15be loaded properly into T3 code. I think
432. 14:18we should be reporting this as T3-code
433. 14:20anyways because, let's be real, that is
434. 14:23the entry point. The entry point isn't
435. 14:25SDK. TS The entry point is T3 code. So,
436. 14:26I think this PR is an accuracy win, but
437. 14:29I'm scared if I merge it, might cause
438. 14:31problems when Anthropic sees a huge
439. 14:32spike in this new entry point. So,
440. 14:35do as you please here.
441. 14:37Well, good news. Things changed
442. 14:39literally while I was filming. You might
443. 14:41see this 2 minutes ago. Yeah, I got
444. 14:44semi-solid permission to merge this
445. 14:46which should, hypothetically, fingers
446. 14:48crossed, get around the 5-hour limit
447. 14:51changes that affect T3 code so that you
448. 14:54will get the same usage for the 5-hour
449. 14:56limit as well as for the longer-term
450. 14:58weekly limit in T3 code as you would in
451. 15:01Claude code directly. So, yeah. Thank
452. 15:03you to my friends at Anthropic for being
453. 15:05reasonable with all of this. I know it
454. 15:07is not trivial to convince that company
455. 15:09to let you do things, but yeah. I wanted
456. 15:12these limits to not affect our users,
457. 15:14and as of this moment, literally as I'm
458. 15:16recording, once this release goes out,
459. 15:18they no longer will. Huge win. That all
460. 15:20said, I didn't notice these problems
461. 15:22because I've been routing through CLI
462. 15:24proxy anyways, which already applied
463. 15:25things like this to prevent this issue.
464. 15:27So, again, if you wanted to have
465. 15:29multiple accounts or manage all of that,
466. 15:31I don't recommend doing it in T3 code
467. 15:32directly. I will not be held accountable
468. 15:34for any of these things getting your
469. 15:35account banned. I'm not telling you you
470. 15:37should go do this and that it's totally
471. 15:38safe and friendly and fine and Anthropic
472. 15:40loves it. I'm just saying what my
473. 15:42experience has been.
474. 15:43And from my experience, if you hunt
475. 15:45through the FAQ, you might just find a
476. 15:47guest somewhere that shows my exact
477. 15:50setup that your agents should be able to
478. 15:53take and understand and work against.
479. 15:56And this is the exact setup that I've
480. 15:57been using for all of my inference for
481. 15:59at least 4 months now, probably more,
482. 16:01and I haven't had any issues personally.
483. 16:02So, again,
484. 16:04take it as you will. Few things I have
485. 16:05to say about the 5-hour limits though.
486. 16:07That's all I have to talk about here. A
487. 16:08nice, simple, short video all about the
488. 16:10new Claude sub plan that isn't really
489. 16:12all that new. I'm sure there won't be
490. 16:14anything else I have to cover similar
491. 16:15like, I don't know, a new Codex plan
492. 16:18Thanks, Tibo. Coming up very soon will
493. 16:20be my video on the new Codex limits and
494. 16:22how they compare to what Claude offers
495. 16:24right now. Because before this change,
496. 16:26the Codex plans were just not a good
497. 16:28value in comparison to the Claude plans.
498. 16:30Now the math is very different and I
499. 16:31will go through all of that math with
500. 16:33you guys in a video coming very soon. So
501. 16:35subscribe if you haven't. Let me know
502. 16:36what you think. Hopefully this was a
503. 16:37useful breakdown, a bit different from
504. 16:39what I normally do, but just wanted to
505. 16:40get this news out as quickly as I could.
506. 16:42Let me know how y'all feel about it and
507. 16:43until next time,
508. 16:44peace nerds.