You know that feeling when you ask an AI a question and it answers with such confidence you almost believe it, only for you to realize five minutes later that it just lied to your face? That's the "AI hallucination" headache most of us are tired of. Today, the developers behind the Claude series dropped version 4.8, claiming they've finally taught the model how to admit when it doesn't know something.
This isn't just a minor update hidden in the fine print. The new model comes with a feature called "fast mode," which runs at 2.5 times the speed of the older versions. Better still, it costs three times less than it used to. For someone in Nigeria trying to build a startup or crunch data without burning their entire budget on API tokens, that price drop is the real headline.
The Price of Intelligence
If you're using the regular version of Opus 4.8, the cost remains at $5 for every million input tokens and $25 per million output tokens. However, for power users who need the "fast mode," you'll be looking at $10 for input and $50 for output. The developers are clearly betting that by making the model more affordable, they can get more people using it for heavy-duty coding and complex tasks. They've also updated the "Claude Code" tool with dynamic workflows, essentially letting the bot chew through massive, complicated projects that would usually make older models crash or loop endlessly.
Early testers report that Opus 4.8 is more likely to flag uncertainties about its work and less likely to make unsupported claims.
This "honesty" metric is a big deal in the world of machine learning. The system card shows that the model is now four times less likely to leave mistakes in the code it writes for you. It's not perfect, but it's a step away from that annoying habit where the bot pretends to be a genius when it's actually guessing.
Control Your Own AI
One of the coolest features they've added to the website is the effort slider. Usually, you're stuck with whatever "thought process" the AI decides to give you. Now, you can choose whether you want the model to put in normal effort, "extra," or "max" effort. If you're stuck on a particularly nasty piece of programming, flipping that switch to "max" tells the AI to spend more time and "thinking energy" to get it right, although it will cost you more in tokens.
They're also testing a much more powerful system called Claude Mythos. Currently, only a select group of organisations are using this for cybersecurity work, where mistakes could literally be the difference between a secure server and a massive hack. They're holding off on a full public release until they can make sure the safety measures are rock solid. They expect to roll out this higher intelligence class to everyone in the next few weeks.
For those of us watching the AI race, this move shows a shift in strategy. Instead of just chasing the biggest, most expensive model, these companies are finally listening to people who want faster, cheaper, and—most importantly—less "confidently wrong" machines. If you're a developer or just someone who uses AI to get through your workday, you can access the new model through the API right now, or jump on the website to play with the new settings.