1 00:00:02,600 --> 00:00:06,206 So for about a fortnight, the two most talked about AI models on the planet 2 00:00:06,206 --> 00:00:07,167 were basically tied. 3 00:00:07,687 --> 00:00:11,904 Google's Gemini 4 Argon scored 53 on the Artificial Analysis Intelligence Index. 4 00:00:12,424 --> 00:00:14,364 Meta's Muse Wonder scored fifty-two. 5 00:00:16,034 --> 00:00:16,834 One point. 6 00:00:17,129 --> 00:00:20,331 which, honestly, is about the difference you get from running the same test 7 00:00:20,331 --> 00:00:21,131 twice. 8 00:00:21,107 --> 00:00:24,303 dasher, but one of them costs less than half as much to actually use. 9 00:00:25,973 --> 00:00:28,739 About one pound fifty a task, against three pounds forty. 10 00:00:29,259 --> 00:00:31,299 And that sounds like the end of the conversation. 11 00:00:31,819 --> 00:00:33,760 cheaper model, same brains, job done. 12 00:00:34,280 --> 00:00:35,080 It isn't. 13 00:00:36,584 --> 00:00:38,501 because Google's price has a sell-by date. 14 00:00:39,021 --> 00:00:41,730 Matta's biggest move this month isn't even the model you can buy. 15 00:00:42,250 --> 00:00:45,254 and the number that really decides your bill isn't the one on the pricing page. 16 00:00:46,924 --> 00:00:48,009 I'll get to all three. 17 00:00:48,529 --> 00:00:51,208 But first, let's be clear about who just walked into this fight. 18 00:00:54,478 --> 00:00:58,302 14 months ago, Meta basically tore up its AI effort and started again. 19 00:00:58,822 --> 00:01:03,035 New lab, new stack, and a restructure led by Alexandr Wang, the founder of Scale 20 00:01:03,035 --> 00:01:03,835 AI. 21 00:01:03,713 --> 00:01:07,278 Since then, they'd shipped Muse Spark, a lightweight model built to be quick 22 00:01:07,278 --> 00:01:08,078 and a Muse agent. 23 00:01:08,595 --> 00:01:11,086 Useful, but nothing that could sit at the top table. 24 00:01:11,606 --> 00:01:13,698 On the twelfth of October, that changed. 25 00:01:14,218 --> 00:01:16,908 Muse Wonder, and Muse Wonder Full. 26 00:01:17,428 --> 00:01:20,329 TechCrunch called it Matta's first real attempt at a frontier model. 27 00:01:20,849 --> 00:01:24,738 and it was trained at Prometheus, Meta's' Gigawatt Scale Data Center in Ohio. 28 00:01:25,258 --> 00:01:27,980 Now, Meta's pitch isn't really, we're cleverer than everyone. 29 00:01:28,500 --> 00:01:30,515 It's two features with very meta names. 30 00:01:31,035 --> 00:01:34,226 active memory, which is remembering what matters to you across conversations. 31 00:01:34,746 --> 00:01:38,358 and active focus, which is basically not giving up when a task gets hard. 32 00:01:38,878 --> 00:01:42,270 trying another route, recovering from errors, and checking in before it does 33 00:01:42,270 --> 00:01:43,118 anything important. 34 00:01:43,638 --> 00:01:45,842 and metas launch demos lean right into that 35 00:01:46,362 --> 00:01:49,881 A playable kitchen game built from a few photos. A running plan worked out across 36 00:01:49,881 --> 00:01:52,183 your apps. Custom merch designed from your Instagram. 37 00:01:52,703 --> 00:01:53,847 and a wedding seating chart. 38 00:01:54,367 --> 00:01:58,121 Which, if you've ever had to do one, is genuinely the hardest problem in artificial 39 00:01:58,121 --> 00:01:58,921 intelligence. 40 00:02:00,379 --> 00:02:02,414 One of their research scientists put it like this. 41 00:02:02,934 --> 00:02:06,340 We put a tremendous amount of effort into making sure that our model can remember 42 00:02:06,340 --> 00:02:08,863 what's important to you over days, weeks, months and beyond. 43 00:02:09,383 --> 00:02:11,467 Fine, every lab says something like that. 44 00:02:11,987 --> 00:02:13,040 So, receipts. 45 00:02:13,560 --> 00:02:17,315 Artificial Analysis, who are independent, scored Wonder Full at fifty-six. 46 00:02:17,835 --> 00:02:21,423 That's 8 points up on Moose Spark 1.3, which came out just 6 weeks earlier. 47 00:02:23,093 --> 00:02:26,140 The biggest jump was on terminal bench, which tests command line coding. 48 00:02:26,660 --> 00:02:28,524 33% to 57%. 49 00:02:30,194 --> 00:02:31,113 That's not a tweak. 50 00:02:31,633 --> 00:02:35,756 On coding agents, Wonder Full scored sixty-eight, level with Claude Code running 51 00:02:35,756 --> 00:02:36,556 Sonnet 5.5. 52 00:02:37,993 --> 00:02:42,091 Meta's own testing also puts Wonder ahead of Claude, GPT-6 and Fable on its 53 00:02:42,091 --> 00:02:43,020 memory benchmark. 54 00:02:43,540 --> 00:02:45,332 Vendor numbers, so pinch of salt. 55 00:02:45,852 --> 00:02:47,287 But here's the bit I didn't expect. 56 00:02:47,807 --> 00:02:49,938 Knowledge is still a relative weakness for Wonder. 57 00:02:50,458 --> 00:02:53,023 Yet it pushed its score up mostly by making things up less. 58 00:02:53,543 --> 00:02:56,697 Its hallucination rate dropped from 33% to 23%. 59 00:03:01,117 --> 00:03:01,989 Now, Google. 60 00:03:02,509 --> 00:03:04,621 Argon was announced on the thirtieth of September. 61 00:03:05,141 --> 00:03:06,987 And then, well, you couldn't use it. 62 00:03:07,507 --> 00:03:10,521 For two weeks it's up behind something called the Fairwind Program, a gated 63 00:03:10,521 --> 00:03:13,253 rollout for approved partners, plus a government pre-release review. 64 00:03:13,773 --> 00:03:17,570 So for a Fortnite, the best value model on the leaderboard was, as one router 65 00:03:17,570 --> 00:03:20,134 company put it, a set of numbers rather than an API. 66 00:03:20,654 --> 00:03:22,773 On the thirteenth of October, the door opened. 67 00:03:23,293 --> 00:03:28,132 the Gemini API, AI Studio, the Gemini app, and Pro and Ultra subscribers. 68 00:03:28,652 --> 00:03:32,726 Google's pitching it at the heavy stuff. Complex coding, knowledge work, cybersecurity 69 00:03:32,726 --> 00:03:33,526 and science. 70 00:03:33,815 --> 00:03:34,814 and the numbers are good. 71 00:03:35,334 --> 00:03:39,001 53 on the Intelligence Index, 8 out of 216 models. 72 00:03:39,521 --> 00:03:43,279 And first on the Vals Index, which weights finance, legal, tax and coding work 73 00:03:43,279 --> 00:03:45,447 by how much of the US economy they represent. 74 00:03:45,967 --> 00:03:47,368 68.9% 75 00:03:49,038 --> 00:03:50,242 And this one I genuinely like. 76 00:03:50,762 --> 00:03:54,605 On a test of how often a model makes things up, Argon's hallucination rate is 77 00:03:54,605 --> 00:03:55,405 fifteen percent. 78 00:03:57,074 --> 00:03:59,475 The figure for GPT-6 Astra is fifty-one. 79 00:03:59,995 --> 00:04:03,450 Artificial analysis called it the lowest of any model in that top tier. 80 00:04:03,970 --> 00:04:05,629 Now, not everything's a win. 81 00:04:06,149 --> 00:04:10,961 On terminal bench, the command line test where meta just jumped, argon scores 82 00:04:10,961 --> 00:04:11,761 57.4 83 00:04:11,731 --> 00:04:14,662 Claude Opus 5.5 scores 66.4 84 00:04:15,182 --> 00:04:17,903 and that's from Google's own comparison table, so fair play to them. 85 00:04:18,423 --> 00:04:20,136 Google's internal claims are bigger again. 86 00:04:20,656 --> 00:04:24,307 more than a third more code submissions in their anti-gravity tool and help 87 00:04:24,307 --> 00:04:26,498 refactoring around 6 million lines of chrome. 88 00:04:27,018 --> 00:04:28,470 Interesting. Unverified. 89 00:04:28,990 --> 00:04:30,161 File under nice if true. 90 00:04:31,831 --> 00:04:36,216 And, according to Business Insider, Google already has more Gemini 4 variants 91 00:04:36,216 --> 00:04:38,950 in testing, apparently called Barium and Carbon. 92 00:04:39,470 --> 00:04:42,060 Because of course they do, the periodic table is a long list. 93 00:04:46,480 --> 00:04:47,280 Right, money. 94 00:04:47,781 --> 00:04:50,034 This is where most of the coverage stops reading. 95 00:04:50,554 --> 00:04:51,856 price per token first. 96 00:04:52,376 --> 00:04:56,006 Argon, two dollars per million tokens in, and ten dollars out. About one pound 97 00:04:56,006 --> 00:04:57,402 fifty, and seven pounds fifty. 98 00:04:57,922 --> 00:05:00,636 Stalls. Muse Wonder, 4 in and 20 out. 99 00:05:01,156 --> 00:05:03,608 Exactly double, the same as Claude Opus 5.5. 100 00:05:04,128 --> 00:05:05,575 but nobody really pays per token. 101 00:05:06,095 --> 00:05:06,940 You pay per job. 102 00:05:07,460 --> 00:05:11,276 So artificial analysis ran both models through their entire index and added 103 00:05:11,276 --> 00:05:12,076 up the bill. 104 00:05:13,557 --> 00:05:16,118 Argon, about one dollar ninety-nine per task. 105 00:05:17,788 --> 00:05:19,951 And for Muse Wonder, it's four dollars fifty-two. 106 00:05:21,621 --> 00:05:24,830 That's 2.3 times the cost for 1 point less on the index. 107 00:05:25,350 --> 00:05:26,632 So why's the gap so big? 108 00:05:27,152 --> 00:05:28,945 Here's the bit that genuinely surprised me. 109 00:05:29,465 --> 00:05:31,523 Wonder is actually quite thrifty with its answers. 110 00:05:32,043 --> 00:05:36,414 About 54,000 output tokens per task, fewer than Argon 62,000. 111 00:05:36,934 --> 00:05:37,746 It's the reading. 112 00:05:38,266 --> 00:05:41,681 Wonder pulls in about seven point seven million input tokens per task. 113 00:05:42,201 --> 00:05:46,926 That's two to seven times what Argon and OpenAI's latest models use per task. 114 00:05:47,446 --> 00:05:49,393 76% of its bill is input. 115 00:05:51,063 --> 00:05:54,322 Now, most of that reading is cached, which means it's billed at a big discount. 116 00:05:54,842 --> 00:05:57,771 Both models knock about 95% off cached input. 117 00:05:58,291 --> 00:06:01,347 But even at a discount, seven and a half million of anything adds up. 118 00:06:01,867 --> 00:06:03,827 So it writes less, but it reads a lot more. 119 00:06:04,347 --> 00:06:07,161 Which, if you've ever worked with someone who reads every single attachment 120 00:06:07,161 --> 00:06:08,887 before replying to an email, you'll recognise. 121 00:06:09,407 --> 00:06:10,833 And now, the sell-by date. 122 00:06:11,353 --> 00:06:14,505 Argon's price is introductory. It runs until the thirtieth of November. 123 00:06:15,025 --> 00:06:16,866 On the first of December, Google doubles it. 124 00:06:18,536 --> 00:06:20,112 Four dollars in, twenty out. 125 00:06:20,632 --> 00:06:22,433 Exactly. Muse Wonders price. 126 00:06:24,103 --> 00:06:28,679 So run the same maths on cost per task and argon goes from about $2 to about 127 00:06:28,679 --> 00:06:29,479 4. 128 00:06:30,469 --> 00:06:31,932 Wonder is at 452. 129 00:06:32,452 --> 00:06:34,781 The gap that looked enormous shrinks to about 10%. 130 00:06:36,451 --> 00:06:39,237 Now, that's my back-of-the-envelope, not a re-run of the benchmark. 131 00:06:39,757 --> 00:06:43,428 Pashing changes things and Argon was measured at high effort while Wonder was 132 00:06:43,428 --> 00:06:44,228 at max. 133 00:06:44,282 --> 00:06:45,606 but the direction is pretty clear. 134 00:06:46,126 --> 00:06:49,437 If you're choosing on price, you're choosing on a price that expires in six 135 00:06:49,437 --> 00:06:50,237 weeks. 136 00:06:54,122 --> 00:06:57,094 and then Mehta did something that, to me, is the bigger story. 137 00:06:57,614 --> 00:07:01,207 Three days after Wonder, on the fifteenth of October, they released the weights 138 00:07:01,207 --> 00:07:02,117 for Muse Spark Open. 139 00:07:03,787 --> 00:07:05,866 816 billion parameters. 140 00:07:07,536 --> 00:07:10,787 free to download, free to fine-tune, free to sell products built on it. 141 00:07:11,307 --> 00:07:12,361 with a couple of strings. 142 00:07:12,881 --> 00:07:16,105 You have to show built with new spark open somewhere in your product. 143 00:07:16,625 --> 00:07:20,187 And if you run it as a paid API business turning over more than $20 million 144 00:07:20,187 --> 00:07:22,229 a year, you need a separate deal with Meta. 145 00:07:22,749 --> 00:07:23,880 Here's the clever part. 146 00:07:24,400 --> 00:07:26,137 It's what's called a mixture of experts. 147 00:07:26,657 --> 00:07:31,088 There are 256 specialists inside it, and for each word, only 8 of them wake 148 00:07:31,088 --> 00:07:31,888 up. 149 00:07:31,785 --> 00:07:36,288 So of those 816 billion parameters, only about 42 billion are working at any 150 00:07:36,288 --> 00:07:37,088 one moment. 151 00:07:38,610 --> 00:07:40,964 Huge model, small active core. 152 00:07:41,484 --> 00:07:45,091 And people jumped on it. Within days, Hugging Face was already listing twenty 153 00:07:45,091 --> 00:07:47,011 quantised versions, and seven fine-tunes. 154 00:07:48,681 --> 00:07:52,044 Matter also says this open version had extra safety training, because once it's 155 00:07:52,044 --> 00:07:54,342 out there there's no system around it to catch things. 156 00:07:54,862 --> 00:07:58,255 They recommend adding your own guardrails on top, which is fair. 157 00:07:58,775 --> 00:08:01,192 Which is why Meta says it runs on 2 Mac Studios. 158 00:08:02,862 --> 00:08:04,296 Two. Not a data centre. 159 00:08:04,816 --> 00:08:08,298 Though, to be fair, they mean the M5 Ultra ones with a frankly silly amount 160 00:08:08,298 --> 00:08:09,098 of memory. 161 00:08:09,282 --> 00:08:12,715 and the full precision version still wants about 2TB of graphics memory. 162 00:08:13,235 --> 00:08:15,715 So no, you're not running it on the laptop you're watching this on. 163 00:08:16,235 --> 00:08:18,259 But think about what that does to the whole price argument. 164 00:08:18,779 --> 00:08:19,965 Google is charging rent. 165 00:08:20,485 --> 00:08:22,941 Meta just handed out the keys to a very large house. 166 00:08:23,461 --> 00:08:26,709 And for a lot of businesses, the question stops being, which API is cheaper? 167 00:08:27,229 --> 00:08:29,039 It becomes, do we need an API at all? 168 00:08:33,459 --> 00:08:36,612 So, if you're actually choosing, and this is the sort of thing we test at AI 169 00:08:36,612 --> 00:08:38,355 Tools Review, here's how I think about it. 170 00:08:38,875 --> 00:08:42,640 If you're building something that reads a lot, long documents, big codebases, 171 00:08:42,640 --> 00:08:44,792 research, Argon, today, is the obvious pick. 172 00:08:45,312 --> 00:08:47,917 Lower hallucination, much lower cost per task. 173 00:08:48,437 --> 00:08:51,011 Just put a reminder in your calendar for the first of December. 174 00:08:51,531 --> 00:08:54,988 If you want a model that remembers you and keeps going on long, messy tasks, 175 00:08:54,988 --> 00:08:56,853 that's exactly what MetaBills Wonder for. 176 00:08:57,373 --> 00:09:01,206 It also takes speech and video in directly, which Argon's published inputs don't. 177 00:09:01,726 --> 00:09:05,850 If you're a coder, Wonder Full inside Muse Code is now level with Claude Code. 178 00:09:06,370 --> 00:09:09,714 But it's subscription only. You can't call Wonder Full from the API at all. 179 00:09:10,234 --> 00:09:12,398 and a tip if you do use Wonder through the API. 180 00:09:12,918 --> 00:09:16,419 It has an effort dial, from minimal right up to max, and every bit of thinking 181 00:09:16,419 --> 00:09:17,317 is billed as output. 182 00:09:17,837 --> 00:09:21,643 So run the simple stuff, extraction, sorting, tagging at minimal, and save Max 183 00:09:21,643 --> 00:09:23,399 for the jobs that genuinely need it. 184 00:09:23,919 --> 00:09:26,448 That alone can move your bill more than the choice of model. 185 00:09:28,118 --> 00:09:31,672 If you've got data you can't send to anyone, or volumes that make API bills 186 00:09:31,672 --> 00:09:33,851 scary, Mu's Spark Open is the interesting one. 187 00:09:34,371 --> 00:09:37,409 Not cheap to host, but nobody can double the price on you in December. 188 00:09:37,929 --> 00:09:42,087 And if the job is heavy terminal work, command line agents, devops, the honest 189 00:09:42,087 --> 00:09:43,579 answer this week is neither. 190 00:09:44,099 --> 00:09:46,401 Claude Opus 5.5 still leads that test. 191 00:09:46,921 --> 00:09:47,721 One more thing. 192 00:09:48,073 --> 00:09:50,150 A one-point difference on an index is noise. 193 00:09:50,670 --> 00:09:53,388 Meta actually makes this point themselves, which I didn't expect. 194 00:09:53,908 --> 00:09:57,432 Their blog talks about two models with similar scores, where one just handles 195 00:09:57,432 --> 00:09:58,714 unfamiliar tasks far better. 196 00:09:59,234 --> 00:10:03,256 They even use the internet's words for it. Benchmaxxing, versus big model smell. 197 00:10:04,926 --> 00:10:08,386 When the AI lab start quoting Reddit at you, you know the leaderboard isn't 198 00:10:08,386 --> 00:10:09,186 the whole story. 199 00:10:09,644 --> 00:10:10,672 Test on your own work. 200 00:10:11,192 --> 00:10:14,809 Run 50 real tasks through each one and look at the bill, not the benchmark. 201 00:10:16,729 --> 00:10:19,390 Zoom out and the shape of this month is pretty clear. 202 00:10:19,910 --> 00:10:21,549 The top of the leaderboard is a crowd now. 203 00:10:22,069 --> 00:10:26,041 Google, Meta, OpenAI, Antropic, all within a few points of each other. 204 00:10:26,561 --> 00:10:30,795 Both of these models read a million tokens at a time. Both cache, both code 205 00:10:30,795 --> 00:10:31,595 well. 206 00:10:31,597 --> 00:10:34,123 On raw ability, the differences are getting hard to see. 207 00:10:34,643 --> 00:10:35,894 which means the fight has moved. 208 00:10:36,414 --> 00:10:39,620 It's not really about who's smartest anymore. It's about price, memory, and 209 00:10:39,620 --> 00:10:41,031 who lets you take the model home. 210 00:10:41,551 --> 00:10:43,435 Google's betting you'll pay rent once you've settled in. 211 00:10:43,955 --> 00:10:45,980 Meta's betting you'd rather own the place. 212 00:10:46,500 --> 00:10:49,619 And we'll find out who read that right around the first of December, when that 213 00:10:49,619 --> 00:10:50,419 price doubles. 214 00:10:51,849 --> 00:10:53,077 Anyway, that's it for this one. 215 00:10:53,597 --> 00:10:55,514 Tell me in the comments which one you're actually using. 216 00:10:56,034 --> 00:10:59,717 Full comparisons and pricing are on aitoolsreview.co.uk. 217 00:11:00,237 --> 00:11:01,172 See you in the next one.