Fable 5 and Sonnet 5

Best Bang for Your Buck

To get the best bang for your buck, GPT-5.5 is still the best pick for anything it can handle, assuming you adjust the thinking level to match the task you are doing. For high-performance work, Fable 5, or Mythos 5 if you have access to it, is both the best model out there, and the best bang for your buck.

One of those two options gives you the best value at almost any price level: GPT-5.5 for normal low-to-medium reasoning, and Fable 5 for the hardest stuff.

TLDR of the News

Fable 5 is back globally. Mythos 5 access is also being restored where Anthropic allows. Along with Fable 5's return, Anthropic released a new Sonnet-class model.

Sonnet 5 is smarter than Sonnet 4.6. It's better at tool use, coding, browser use, terminal use, and agentic work. That being said, the model has issues. While it's cheaper per token than Sonnet 4.6, it uses significantly more tokens. This means it can cost even more than Opus 4.8 in some cases, and produce worse results.

Artificial Analysis found that it can cost more per completed benchmark task than Opus 4.8.

My take is that Sonnet 5 gives worse value than Opus 4.8 in a lot of real use cases, especially computer use. It'll most likely be useful for sub-agents, but overall this isn't a groundbreaking release.

Fable 5 Is Back. Good.

On June 12, Anthropic said the U.S. government issued an export control directive suspending access to Fable 5 and Mythos 5 for foreign nationals. Anthropic said it couldn't verify nationality in real time, so it suspended everyone's access.

Anthropic also said it disagreed with using a narrow jailbreak report as the reason to recall a deployed model.

On June 30, the U.S. Commerce Department lifted the export controls. Anthropic said Fable 5 would return globally on July 1. Anthropic also said it added stronger safeguards for the reported jailbreak technique. Reuters reported that the export controls were lifted after Anthropic implemented those safeguards.

In my opinion that's the right outcome to a situation that shouldn't have happened. I still think the original control was too broad, but Anthropic handled the response well. They worked with the government, fixed the specific issue, and got access restored globally.

Sonnet 5

Along with Fable 5's return, Anthropic launched Claude Sonnet 5, a brand new Sonnet-class model. Anthropic calls Sonnet 5 its "most agentic Sonnet model yet." It can use browsers and terminals. Anthropic says it improves on Sonnet 4.6 in reasoning, tool use, coding, and knowledge work. Anthropic also says it can match Opus 4.8 on some tasks at higher effort settings.

Sonnet is cheap per token but bad value in real use

Sonnet 5 is cheaper than Opus 4.8 per token. Sonnet 5 is $2 per million input tokens and $10 per million output tokens through August 31. After that, it goes to $3 and $15. Opus 4.8 is $5 and $25. (Pricing.)

That makes Sonnet 5 look cheaper on paper. The problem is that Anthropic says Sonnet 5 uses a newer tokenizer that produces about 30% more tokens for the same text compared to Sonnet 4.6 and earlier. Artificial Analysis also found that Sonnet 5 costs about $2.29 per Intelligence Index task, around 15% more than Opus 4.8 before promo pricing.

In real-world use, it seems that Sonnet 5 is both more expensive and less intelligent than Opus 4.8 at a lot of tasks.

It does score well on some benchmarks, but those numbers need context. SWE-Bench-style scores are optimized for, and they're starting to mean nothing. Despite this, I think some benchmarks can be positive indicators, like Terminal-Bench. In my opinion, Terminal-Bench matters a bit more because it tests real terminal agent work. Sonnet 5 is doing better there and that's a real positive, but it doesn't erase the value issue.

Sonnet 5 is not the Safety Story

Sonnet 5 isn't in the same class as Fable 5 or Mythos 5.

Anthropic says Sonnet 5 has lower undesirable behavior than Sonnet 4.6 and is better at refusing malicious requests. Anthropic also says Sonnet 5 has much lower cybersecurity capability than current Opus models. In Anthropic's own broader behavioral audit, Sonnet 5 scored below Opus 4.8 and Mythos Preview.

It seems like Sonnet 5 refuses a lot of benign requests that look malicious, even more than Mythos 5. This will have its downsides, but overall I don't think it will matter too much in everyday use.

Best use case: sub-agents

I don't think Sonnet 5 is best as a main model.

It makes more sense as a sub-agent. Give it a narrow job like exploring code, checking logs, inspecting a repo, or reviewing another model's work, and it will work wonders. It may also work well when it can spawn another Sonnet 5 sub-agent for a second pass.

That lines up with Anthropic's Claude Code guidance. Their help docs describe a common pattern where Opus plans and Sonnet executes.

My take is simple: use Sonnet 5 as a worker model. Do not build the whole workflow around it.

Note: There is no Opus 5 and no Haiku 5 right now in Anthropic's current public model lineup.