Why Anthropic's Claude Sonnet 5 and the Advisor Tool are Changing the AI Game

I have been keeping a close eye on the latest moves in the artificial intelligence space, and Anthropic's newest release, Claude Sonnet 5, really caught my attention. To me, this launch is a major step forward because it makes "agentic" AI so much more efficient. It is not that this model suddenly does completely new things that older models could never do. Instead, it is about making these smart, independent tasks much faster and, more importantly, a lot cheaper to run.

In the past, if you wanted an AI to do all these tasks, you had to use massive, incredibly expensive models. What Anthropic has done here is change the math. They are offering these smart, independent skills at a much lower price. Of course, they are also keeping our expectations realistic. They have made it clear that Sonnet 5 is not their most independent model, as their deepest and most powerful skills are still kept for their Opus models. Even so, the way I see it, Sonnet 5 performs amazingly close to Claude Opus 4.8 without the giant price tag. This makes it a very attractive option for businesses trying to save money. People testing it early say it is much more independent than older versions, handles roadblocks without getting stuck, and even fixes its own coding mistakes as it goes.

What Stands Out in the New Model

Compared to Claude Sonnet 4.6, the new model brings major upgrades to reasoning, coding, tool use, and safety systems.

Historically, big AI models have struggled with two annoying problems. First, they "hallucinate," which means they make up fake facts and present them as true. Second, they suffer from "sycophancy," which is a fancy word for always politely agreeing with whatever the user says, even when the user is completely wrong. Sonnet 5 tackles both of these issues head-on. It does a much better job of sticking to the facts, and it will actually point out your mistakes instead of just nodding along.

Because this kind of AI can run code directly, safety is a huge concern. Anthropic has been very open about how the model handles tricky tasks. For example, when they tested it on hard hacking challenges—like trying to find security holes in the Firefox web browser—the model actually failed. This failure was simply because the model is not trained for advanced hacking, rather than a safety block stopping it in real-time. That said, they do ship Sonnet 5 with default safety guards to block harmful actions and keep computer systems secure.

I also want to clear up a common misunderstanding about the models. Some people talk about "Fable 5" as if it is just a secure version of Sonnet 5. That is not the case. Fable 5 is actually Anthropic's most powerful, top-tier model. It belongs to their "Mythos" class, which sits even higher than Opus 4.8. Fable 5 was launched back on June 9, but it has had a bumpy ride. Because of U.S. government export controls, both Fable 5 and Mythos 5 were taken offline from June 12 to June 30. Thankfully, access was finally restored on July 1.

For developers who need extra security, there is an optional setup they can choose to turn on. If they opt-in, any highly sensitive or risky questions sent to Fable 5 that trigger safety flags will automatically route over to Opus 4.8. But this the actual triggering of this fallback setup is supposedly quite rare.

Let's Talk About Costs and Privacy

Using these big models always requires balancing how much they cost with how they handle private data, especially with strict privacy laws like GDPR in Europe.

For example, Fable 5 keeps your prompts and responses for 30 days just to check for any bad behavior or abuse, and they promise this data is never used to train future AI models. Still, some companies have strict rules that do not allow any data storage at all, which means even temporary storage is a dealbreaker.

To help businesses budget, Anthropic has set up a clear pricing system. Claude Haiku 4.5 is still the cheapest option for fast, simple tasks. Sonnet 5 is now the default model for the Free and Pro accounts. Right now, Anthropic is running a discount price until August 31, 2026. During this promo, it costs $2.00 per million "input tokens" (the words you feed into the AI) and $10.00 per million "output tokens" (the words the AI writes back). After the promo ends, the price will go up to $3.00 for inputs and $15.00 for outputs. Even though Sonnet 5 uses a new system that might count words a bit differently, this cheap promo price helps offset any extra costs. Meanwhile, Fable 5 is priced much higher, at $10.00 per million inputs and $50.00 per million outputs, making it twice as expensive as Claude Opus 4.8.

To keep these bills under control, smart tech managers use special software to set spending limits, block access to the most expensive models, and automatically send simple questions to the cheaper, faster models.

What I really appreciate about Anthropic's approach is their focus on stability. While other AI models sometimes slow down or get worse when a lot of people are using them at the same time, Anthropic has worked hard to keep their speeds steady and reliable. In coding tests, their top models show incredible results. For example, according to the SWE-bench Pro leaderboard Fable 5 successfully solved 80.3% of the problems in a tough coding test, compared to 69.2% for Claude Opus 4.8.

The Advisor Tool

Usually, choosing an AI system is all about making tough trade-offs between cost and smarts. But Anthropic recently came out with a new beta feature that completely changes how we think about this: the advisor tool.

This is a brilliant concept. Instead of choosing between a cheap, fast model and a super-smart, expensive one, you can now combine them. Think of it as a team. You have a cheap "worker" (the executor) doing the bulk of the work. If the worker gets stuck on a hard problem, it can quickly ask a smarter "manager" (the advisor) for strategic advice, and then go back to finish the job.

This is where Sonnet 5 gets really exciting. It is the cheapest model that is smart enough to act as the advisor (with Haiku acting as the cheap worker). In more expensive setups, you can use Sonnet 5 as the worker, and let Opus or Fable 5 act as the advisor.

This architecture means we no longer have to make a hard choice between cheap or smart. By combining these models, we can get top-tier results while keeping our token costs incredibly low. To me, that is the real future of AI.