Latest / AI Fire Daily / #522 Neil: Claude Sonnet 5 Just Dropped and It Nearly Matches Opus 4.8
Transcript
- 0:00Picture this scene for just a moment. It's July
- 0:021st, 2026. Anthropic posts a simple number 5
- 0:06on X. It's just made of these delicate flowers
- 0:09and green leaves. Yeah, super minimalist. Right.
- 0:12And there are just two lines of text below it.
- 0:15Beat. But those two lines are quietly cutting
- 0:18enterprise AI bills in half. It really was a
- 0:21massive moment. I mean, that post hit over 300
- 0:24,000 views in hours. Welcome to this deep dive
- 0:27into the source material. Today, we're unpacking
- 0:30Anthropic's massive silent efficiency revolution.
- 0:34We are looking very closely at Claude Sonnet
- 0:365. We've definitely got a lot of ground to cover
- 0:38today. We do. Our mission is to really understand
- 0:40how this alters your workflow. Exactly. So we're
- 0:43going to look at the raw benchmark numbers first.
- 0:45We'll stack Sonnet 5 right against the flagship,
- 0:47Opus 4 .8. Then we need to break down the tricky
- 0:50new pricing math. Right. And after that, we analyze
- 0:52real -world live tests and finally figure out
- 0:55exactly when you should actually trust this cheaper
- 0:57model. So before we look at the cost, we have
- 1:00to understand the engine. What exactly did Anthropic
- 1:03change under the hood? Well, on June 30th, 2026,
- 1:07it completely replaced Sonnet 4 .6. Right. Sonnet
- 1:105 is now the default mid -tier model for free
- 1:13and pro users. And the major shift here is really
- 1:15eponymy, isn't it? Totally. It handles these
- 1:18longer, complex, multi -step jobs on its own.
- 1:22It doesn't stop halfway to ask you for directions.
- 1:24I mean, I have to admit, I still wrestle with
- 1:26prompt drift myself. We all do. It's so frustrating.
- 1:30It really is. It's like finally hiring an assistant
- 1:32who finishes the whole project, instead of them
- 1:34knocking on your door every five minutes for
- 1:36help. Yeah, that mental overhead is just exhausting.
- 1:39But Sonnet 5 removes that friction when tool
- 1:43calls fail in Claude code. It self -corrects.
- 1:46It just tries again silently. Exactly. It reads
- 1:48its own error and rewrites the code. Which makes
- 1:51me wonder, is the real upgrade here just raw
- 1:54intelligence or pure persistence? It's definitely
- 1:56about context retention and self -correction
- 1:58over long task chains. It finishes the whole
- 2:01gob without needing constant hand -holding from
- 2:03you. You nailed it. That's the biggest practical
- 2:05difference. So it feels more autonomous, but
- 2:07how does that translate to the raw benchmark
- 2:10data against Opus 4 .8? The numbers are fascinating,
- 2:14honestly. Let's look at reasoning. Without any
- 2:16tools, Opus wins by a decent margin. 49 .8 %
- 2:20versus Sonnet's 43 .2%. But that completely changes
- 2:23when you add tools into the mix. Oh, completely.
- 2:25You have a browser and a calculator, and Sonnet
- 2:275 jumps up to 57 .4%. It virtually ties Opus
- 2:314 .8's 57 .9%. Right. And for standard knowledge
- 2:35work, Sonnet 5 actually beats Opus. Wait, really?
- 2:39Yeah. On the GDPVAL AAV2 test, it's 1 ,618 to
- 2:431 ,615. That's a mid -tier model beating a flagship
- 2:47at daily office work. What about software development?
- 2:50Well, Opus wins the SWB Bench protest. That's
- 2:5369 .2 % to 63 .2%. But for everyday coding on
- 2:56Terminal Bench 2 .1, they are almost perfectly
- 2:58identical. Yeah, 80 .4 % versus 82 .7%. Which
- 3:02brings up agentic coding. That's AI writing and
- 3:05testing software across multiple files independently.
- 3:07Exactly. It is crazy that a mid -tier model is
- 3:10practically tying a flagship here. It's wild,
- 3:12and honest world computer use tells the same
- 3:14story. The human expert baseline there is 72
- 3:18.4%. Yep, and both miles crush it. Sonnet 5 hits
- 3:2181 .2 % and Opus is at 83 .4%. So why does Sonnet
- 3:265 suddenly catch up when tools are introduced?
- 3:29Well, tool use just levels the playing field.
- 3:31It makes raw internal reasoning slightly less
- 3:34critical. External tools bridge the intelligence
- 3:36gap between the mid -tier and flagship models?
- 3:39That's exactly it. So the benchmarks prove that
- 3:41Sona 5 holds its own. Now let's look at the actual
- 3:43cost. Right, because the pricing math gets slightly
- 3:45complicated. Let's start with standard pricing.
- 3:48From September 1st, 2026, it's $3 in, $15 out
- 3:52per million tokens. And Opus 4 .8 is a steep
- 3:55$5 in, $25 out. But there is an introductory
- 3:58price right now until August 31st, 2026. Yeah,
- 4:02it's just $2 in and $10 out. Two secs silence.
- 4:05Whoa, whoa. Imagine scaling to a billion queries.
- 4:08A $1 ,000 daily opus bill just drops to $400.
- 4:11The savings are staggering, but there's a catch.
- 4:14The new tokenizer. Right. Sonnet 5 produces 1
- 4:17.0 to 1 .35 times more tokens for the exact same
- 4:20text compared to 4 .6. So it's slightly inflating
- 4:23the token count. Does the new tokenizer wipe
- 4:26out the... promised cost savings. No, even with
- 4:29token inflation, the 40 to 60 percent price drop
- 4:33keeps it drastically cheaper. No, even using
- 4:36more tokens, your final bill remains substantially
- 4:39lower. Definitely. It's like buying a much more
- 4:41fuel -efficient car. even if they change how
- 4:43gallons are measured. Exactly. Sponsor. So we've
- 4:46looked at theoretical pricing and benchmarks.
- 4:49Let's talk about the launch day live tests. Yeah.
- 4:51Day one tests are always the most honest. Test
- 4:53one was having both models build interactive
- 4:56HTML5 physics demos. Right. And Sonnet 5 cost
- 4:5915 cents. Opus 4 .8 cost 58 cents. Sonnet used
- 5:04more tokens, about 15k versus 23k for Opus. But
- 5:07it was still roughly 53 % of the final Opus bill.
- 5:10And they matched on two out of the three tests.
- 5:12Though, Opus definitely had better visual graphics.
- 5:15Test 2 is the one that really caught my eye.
- 5:17A fully autonomous landing page build via the
- 5:19command line. Oh man, the speed difference here
- 5:21was insane. Sonnet 5 finished in 2 minutes and
- 5:2411 seconds, cost $3 .36. And Opus. Over 20 minutes,
- 5:29costing $20 .66. So Sonnet was one tenth the
- 5:33time, one sixth. The cost? Opus did have better
- 5:37final polish on the page. Wait, I want to push
- 5:39back on this a bit. Opus won on output quality
- 5:41there. Are we just advocating for faster, cheaper,
- 5:44but ultimately worse work? Not worse, just less
- 5:48polished on draft one. In scale heavy environments,
- 5:51rapid iteration matters more than first draft
- 5:53perfection. Is the massive speed trade off worth
- 5:56the slight drop in final polish? Absolutely.
- 5:59You can always refine a fast, cheap draft later.
- 6:02For most businesses, being 10 times faster easily
- 6:05beats perfect initial polish. 100%. And test
- 6:07three confirmed this exact pattern across three
- 6:09complex prompts. Sonnet cost $0 .10, Opus cost
- 6:12$0 .39. Again, Opus was... Functionally slightly
- 6:15more complex, but Sonnet was way faster. Which
- 6:18brings us to the most practical question. When
- 6:20should you actually use which model? Okay, so
- 6:22Sonnet 5 is your daily driver. Daily coding,
- 6:25agent workflows, repeated automation. Customer
- 6:28chat bots and high -scale content generation
- 6:30too. But Opus 4 .8 is for heavy refactoring and
- 6:33complex code reviews. And deep math or logic.
- 6:37But we have to talk about the safety gap. Yes,
- 6:39the misaligned behavior chart. Lower is better
- 6:41here. Opus 4 .8 scores 2 .0. Sonnet 5 is at 2
- 6:47.53. But the cybersecurity test is what really
- 6:50stands out to me. Yeah, the exploit development
- 6:52test. Sonnet 5 scored exactly 0 % on developing
- 6:56working exploits. While Opus 4 .8 scored 8 .8%.
- 7:01Anthropic explicitly recommends Opus for security.
- 7:05Beat. This is the safety net. You don't go bargain
- 7:08hunting for a parachute. You really don't. An
- 7:108 .8 % success rate in offensive security is
- 7:13highly significant. So Opus is essentially becoming
- 7:16a specialized high -stakes consulting tool. Pretty
- 7:18much. You deploy it when a small drop in accuracy
- 7:21has real dangerous consequences. Exactly. Reserve
- 7:23Opus for when absolute precision and safety are
- 7:26completely non -negotiable. That's the smartest
- 7:28deployment strategy. We've covered a massive
- 7:30amount of ground today. We really have. But the
- 7:32big idea is just how close Sonnet 5 gets to Opus
- 7:354 .8. It redefines the good enough threshold
- 7:37for everyday AI tasks. Paying double for the
- 7:40flagship for daily work just makes zero sense
- 7:43now. You really should take advantage of the
- 7:44introductory pricing before August 31st, 2026.
- 7:48Definitely. Test Sonnet 5 on your heaviest workflows.
- 7:51I want to leave you with a final thought to mull
- 7:53over. We have a mid -tier model beating human
- 7:56expert baselines in computer navigation for 13
- 7:59cents a task. If this level of intelligence is
- 8:02now the floor rather than the ceiling, what happens
- 8:04to the global digital economy next? Outtero music.