The argument in one line.
Opus 4.8 is a flat-price, modestly-improved refresh whose real signal is the Claude Mythos preview Anthropic is staging behind cybersecurity review for release in the coming weeks.
Read if. Skip if.
READ IF YOU ARE…
- A non-technical Claude user who pays for claude.ai or Cowork and wants to know whether anything practical changed with this release.
- Someone deciding whether to switch a workflow from Opus 4.7 to Opus 4.8 and wondering if the migration is worth the attention.
- An operator who already uses Claude Code and wants to know what dynamic workflows and extra-high effort mean for their setup.
- A reader who tracks Anthropic's roadmap and wants the read-between-the-lines on Project Glassuiing and Claude Mythos timing.
SKIP IF…
- You only care about raw benchmark deltas - the video explicitly refuses to dig into those numbers.
- You want a hands-on evaluation of whether Opus 4.8 is genuinely less prone to hallucination - the host punts that test to the future.
- You are looking for a tutorial on Claude Code's dynamic-workflow sub-agents; this video names the feature but never demos it.
The full version, fast.
Opus 4.8 ships at the same price as 4.7 ($5 per million input, $25 per million output) and lands everywhere Claude already runs. The hands-on upgrade most users will feel is the new effort selector (Low / Medium / High / Extra / Max) appearing inside claude.ai and Cowork, with Opus 4.8 defaulting to High; Claude Code gains an extra-high tier and higher rate limits to match. Anthropic frames honesty as the headline improvement, citing an internal four-times-fewer-fake-passes claim on coding tasks, but the host treats that as unproven until he can test it. The real story buried in the announcement is Claude Mythos - a next-generation model held back behind cybersecurity review with Fortune-500 testers and slated for claude.ai in the coming weeks, signaling that 4.8 is groundwork rather than the milestone.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:24
01 · Cold open and no-BS promise
Names the release, promises no benchmark talk, frames the video as honest analysis instead of hype.
00:24 – 01:14
02 · Benchmarks, dismissed
Shows the benchmark table and waves it off; only surviving claim is that Opus 4.8 beats GPT-5.5 and Gemini 3.1 Pro everywhere except agentic terminal coding.
01:14 – 01:45
03 · Availability and pricing
Same price as 4.7 ($5/M input, $25/M output). Available in claude.ai, Claude Code, Cowork, and via the API.
01:45 – 03:14
04 · The new effort mode
Effort selector (Low / Medium / High / Extra / Max) now lives inside the consumer Claude apps. Opus 4.8 defaults to High. Live demo of the dropdown in Cowork.
03:14 – 04:43
05 · Honesty as headline feature
Anthropic claims Opus 4.8 is roughly four times less likely to fake-pass coding tasks. Host stays skeptical and reserves judgement until he can test it himself.
04:43 – 06:03
06 · What is coming next: Claude Mythos
Reads Anthropic's own framing (modest but tangible) as the tell. Project Glassuiing is staging Claude Mythos with Fortune 500 cybersecurity testers; expected in claude.ai in the coming weeks.
06:03 – 06:51
07 · Also launching today
Dynamic workflows in Claude Code (hundreds of parallel sub-agents, can run longer on 4.8). Messages API gains system entries inside the messages array for developers.
06:51 – 08:00
08 · Four-takeaway recap
Same price, focus on honest responses, new effort levels, but ultimately small marginal improvements. The real story is Mythos.
Lines worth screenshotting.
- Opus 4.8 ships at the same price as Opus 4.7: $5 per million input tokens and $25 per million output tokens.
- Opus 4.8 defaults to High effort, which Anthropic calls the best overall balance of quality and user experience.
- The effort selector (Low, Medium, High, Extra, Max) now lives inside claude.ai and Cowork, not just Claude Code.
- Claude Code gains an Extra High effort tier above Max, plus higher rate limits to absorb the heavier token usage.
- Anthropic claims Opus 4.8 is roughly four times less likely than its predecessor to allow flaws in code it has written to pass unremarked.
- Anthropic's own framing - 'modest but tangible improvement' - is the strongest tell that 4.8 is staging ground, not the milestone.
- Claude Mythos Preview is a next-class model currently restricted to a small number of Fortune 500 organizations for cybersecurity testing.
- Dynamic workflows let Claude Code plan a task and run hundreds of parallel sub-agents in a single session, and on Opus 4.8 those agents can run longer.
- Opus 4.8 beats GPT-5.5 and Gemini 3.1 Pro across the benchmark suite except agentic terminal coding, where GPT-5.5 still leads.
- The Messages API now accepts system entries inside the messages array, letting developers update instructions mid-task without breaking the prompt cache.
- Higher effort settings burn more tokens and respond more slowly but give better responses; lower effort settings respond faster and use fewer tokens.
- Anthropic positions the next post-Opus tier (Mythos-class) as requiring stronger cybersecurity safeguards before general release.
Read the model card like a release strategist.
WHAT TO LEARN
Opus 4.8 is the same price and a modest jump - and the most useful thing in the post is the next-model tell Anthropic buried under a headline upgrade.
- Treat Opus 4.8 as a free upgrade: identical pricing to 4.7 ($5/M input, $25/M output) means there is no reason to keep prompting against the older model.
- Use the new effort selector inside claude.ai and Cowork to spend tokens on hard problems and save them on easy ones; High is the safe default Anthropic ships.
- When a model card calls its own release 'modest but tangible' and names a stronger model held back for cybersecurity review, the strategic move is to plan around the next release window, not this one.
- Anthropic's four-times-fewer-fake-passes honesty claim is unverified until you stress-test the model on the kind of code-review or research task that previously hallucinated; do not adopt the claim on faith.
- If you live in Claude Code, dynamic workflows and the new Extra-High effort tier are the only items that materially change your day; everything else is a UX polish.
- Watch the Claude Mythos preview as the real next inflection - Anthropic stating it requires stronger cybersecurity safeguards is a signal it will land with usage limits and access friction, not a clean rollout.
Terms worth knowing.
- Claude Cowork
- Anthropic's consumer task-orchestration surface inside claude.ai where you queue and run agentic tasks against your accounts; the video uses it to demo the new effort selector.
- Effort level
- A per-prompt setting (Low, Medium, High, Extra, Max) that controls how many tokens Claude spends thinking before answering; higher levels are slower but better.
- Dynamic workflows
- A new Claude Code feature in research preview that lets the model plan a task and spawn hundreds of parallel sub-agents inside a single session.
- Claude Mythos
- Anthropic's next-class model, more capable than Opus, currently restricted to Fortune 500 cybersecurity testers and expected to surface in claude.ai in the coming weeks.
- Project Glassuiing
- Anthropic's program name for the controlled release of next-generation models like Mythos, in which a small set of customers stress-test capabilities before public availability.
- Agentic terminal coding
- Benchmark category measuring how well a model performs autonomous, multi-step coding tasks via a terminal interface, where GPT-5.5 still outperforms Opus 4.8.
- System entries in messages array
- A developer-only API change that lets new system instructions be injected mid-conversation without invalidating the prompt cache, useful for long agent runs.
Things they pointed at.
05:22productClaude Mythos Preview (Project Glassuiing)
Lines you could clip.
00:14
“I'm not gonna talk benchmarks and complicate this. I'm just gonna talk about tangibly what is different.”
03:46
“One of the most prominent improvements inside of Opus 4.8 is its honesty, which to me, honestly, sounds like a no brainer. You want AI to give you honest responses.”
04:55
“When they are coming out and saying this is just a modest update, I'm not expecting this to be a huge leap in capabilities.”
05:22
“Claude Mythos - this is the AI model that apparently is too good for them to release to the public, so they've actually given this to some of the Fortune 500 companies to run internal testing.”
07:26
“This is definitely not a game changing release. Of course, it is a new level up from Claude Opus 4.7, but I think this is kind of laying the groundwork for a bigger model release in the Claude Mythos preview.”
Word for word.
metaphor
00:00Anthropic just released Claude Opus 4.8. I dissected the entire article breaking down the new features as well as I've been testing it out myself. So I'm gonna make a on exactly what's different.
00:12If you're nontechnical, I'm not gonna talk benchmarks and complicate this. I'm just gonna talk about tangibly what is different, what you can expect from this, and if this is a real game changer that I could almost guarantee many YouTubers out there are gonna say, I'm not gonna sit here and hype this up.
00:27So without further ado, let's dive into a real breakdown on this. First of all, I read this entire article right here by Anthropic, and I turned it into bite sized breakdowns on the main things and the four main takeaways from this. If you guys want to see the full article, there will be a link in the description so you could just dissect it yourself, but I wanted to condense this for you here.
00:47So first things first, here is a breakdown on the benchmarks. We're not gonna talk about this. Basically, all we need to know is, yes, this is better than Claude Opus 4.7 as well as GPT 5.5 and Gemini 3.1 Pro on basically everything here.
01:02The only thing that any of these other models are better at than OPUS 4.8 is GPT 5.5 for AgenTic terminal coding, but if you're not using Claude code, this doesn't necessarily pertain to you. Alright. So first of all, let's talk about availability.
01:15As of right now, Cloud Opus 4.8 is available everywhere today, so we could use this on the desktop app inside of Cloud Code, Cloud Cowork, all these different places we could use it. And then let's talk about pricing as well. So pricing for regular usage is unchanged from OPUS 4.7, which is great.
01:31That is $5 per million input tokens and $25 per million output tokens. So in terms of pricing, we don't need to worry about anything upgraded from the OPUS 4.7 model.
01:42It's gonna be exactly the same, as well as developers can now use Claude OPUS 4.8 via the Claude API. So if you're using Claude code or you're using it as a developer, that's available right now as well.
01:54Next up, this is one of the bigger things inside of this release. There is a new effort mode that's inside of the Claude dot a I and the Claude desktop app, meaning that if you've ever used Claude code, you can select which effort level Claude Opus 4.7 was gonna perform tasks for you.
02:11So if you're doing something a bit higher, you could put it on max or high mode, and you can now do this directly inside of the normal Claude app. It does it does say that OPUS 4.8 defaults to high effort, which we judge to be the best overall balance of quality and user experience.
02:26So, again, it's gonna automatically default to this. If you have specific tasks that are a bit easier for it to do, you could always toggle it down to a lower mode so you're not spending as much on your tokens. It also looks like if you are using Claude code, users can now choose extra high mode inside of Claude Code, or you could use the normal max mode that, you know, you already could use.
02:47Then it also looks like there is increased rate limits inside of Claude Code to accommodate the higher token usage of higher effort levels. Um, and, again, users can just select whatever they want for a particular project. In order to show you this live in action, here is Claude Cowork.
03:01As you could see, I have Opus 4.7 high selected here, and now I can go ahead and click on Opus 4.8. It automatically defaults to the high mode like I just mentioned before.
03:13And if we click, uh, on Opus 4.8, we can now see this little effort level, and we can just go and preselect what we want. So we could turn it on low, medium, high, extra, or max, as well as we can, you know, toggle on adaptive thinking.
03:27This was already there, but all of these new effort levels that we could see right here is new. So after dissecting this article, one of the main things that I've seen and one of the things that Anthropocos is saying, like, is the biggest part of this release is that one of the most prominent improvements inside of OPUS 4.8 is its honesty, which to me, honestly, sounds like a no brainer.
03:49You want AI to give you honest responses. And for me, personally, this is something I ran into with Opus 4.7 and any AI model for that matter. But a lot of the times, the information that it would give me just wouldn't be correct, and I'd always have to say, hey.
04:02Can you make sure that this is honest and this is the correct information? So, apparently, this is gonna be a bit more honest of a model, which I would hope so. It said that we train all of our models to be honest, for instance, to avoid making claims that they can't support, which, again, I've had this countless amount of times using other OPUS models before and any Claude model.
04:22They did mention that early testers report that OPUS 4.8 is more likely to flag uncertainties about its work and less likely to make unsupported claims. Obviously, we're gonna have to see this in practice, if it actually is more honest, them saying this on paper really doesn't mean much until we test it out. There are a couple of other things that are being released alongside this, so I'm gonna talk about that here in a second.
04:43But first of all, let's talk about what Anthropic is saying is what's coming next. So in their own words, they're saying users will find Opus 4.8 to be a modest but tangible improvements on its predecessor. So when they are coming out and saying this is just a modest update, you know, I'm not expecting this to be a huge leap in capabilities.
05:02We're probably gonna notice a little upgrade, probably just a marginal one. It's not gonna be anything game changing like I said. Hopefully, it is more honest.
05:09You know, as they said, this is, like, a big part of their release. They wanna make this more honest. And they did come out and say that there is still a lot more work to be done.
05:17We're working on developing and releasing models that provide many of the same capabilities as Opus at a lower cost, and that is what's coming next. And if you've never heard of Claude Mythos, this is the AI model that apparently is to the public, so they've actually given this to some of the Fortune 500 companies to run internal testing to make sure this is actually safe to give to the public.
05:39They came out and said that not only that, but we plan to release a new class of models with even higher intelligence than Opus, and this is the new Claude Mythos preview model. Right here inside of the article, it says that we can expect this inside of Claude in the next coming weeks. It didn't really give an exact date, but just know that this is something that could be coming out in the next two, three, four weeks.
06:00Alright. So next up, we have dynamic workflows. This is a new feature available in research preview, allows Claude to take on even bigger tasks inside of Claude code.
06:09So if you don't use Claude code so let's say you're just an everyday Claude chat and Claude CoWork user, this won't necessarily pertain to you, but let me break it down just a little bit here. They are claiming that Claude can now plan the work and run hundreds of parallel sub agents in a single session, and it says that with OPUS 4.8, the agents can run for even longer.
06:29So, again, that only matters if you are using Claude code. Next up, one of the other changes that we did mention is the effort control inside of claude.ai and Claude Cowork, which is something I'm personally happy about.
06:41And then this is another small one, but the messages API now accepts system entries inside the messages array, and this only matters if you are a developer. Alright.
06:50So my main takeaways after reading this article is that this is the same price as OPUS 4.7. It's automatically gonna use the new default high effort mode, and so that is great that this isn't a more expensive model. Next, a big part of this release is focused on honest responses, which is a no brainer.
07:06Obviously, they should be focusing on that. Then we have the new effort levels. Again, I'm actually pretty excited for this.
07:12It's a very small feature, but I think this will be, you know, pretty practical for a lot of people. And then lastly, my main takeaway from this is that these are all just small marginal improvements. This is definitely not a game changing release.
07:24Of course, it is a new level up from Claude Opus 4.7, but I think this is kind of for a bigger model release in the Claude Mythos preview that should be coming out over the coming weeks. So if you guys do wanna test it out, it's available right now in the Claude desktop app as you can see right here.
07:41We could see it inside of Cowork. We could see it inside of normal Claude chat mode, as well as this is available inside of auto code as well. And there we have it.
07:49That's a quick breakdown. If you guys got some value from this video, leave a like, subscribe to this channel for more AI content for nontechnical people. I wanna make this video short and to the point, so hope you have a good day, I'll see you in the next video.
The Hook
The bait, then the rug-pull.
Anthropic shipped Opus 4.8 the same day, and most creator videos turned it into hype. This one does the opposite - reads the announcement aloud, dismisses the benchmarks, and treats the release as a tell about what is actually coming next from Anthropic.
Named ideas worth stealing.
06:55list
Four Takeaways
- Same price as Opus 4.7 (default high effort)
- Focus on honest responses
- New effort levels in claude.ai and Cowork
- Small marginal improvements - this release is groundwork, not the milestone
The host's own whiteboard distillation, used as the through-line in both the intro and the recap.
Steal forany 'read the release post for me' explainer that needs a memorable spine
03:15list
Effort Levels (Cowork / claude.ai)
- Low
- Medium
- High (default for Opus 4.8)
- Extra
- Max
The five effort tiers exposed in the consumer Claude apps for Opus 4.8.
Steal forany tutorial recommending a smart default plus an escape hatch
How they asked for the click.
VERBAL ASK
07:45subscribe
“If you guys got some value from this video, leave a like, subscribe to this channel for more AI content for nontechnical people.”
soft, 10-second outro with no funnel mention; relies on watch-time and topic affinity rather than a lead magnet
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
