Sonnet 5: Agentic Power at Half the Cost

Claude Sonnet 5 is, according to Anthropic's official announcement from June 30, the most agentic Sonnet model ever. It can independently plan tasks, use tools like browsers and terminals, and operate autonomously at a level that just a few months ago required significantly more expensive Opus-class models.

In terms of performance, Sonnet 5 closely approaches Opus 4.8, which is approximately 2.5× more expensive. In both agentic search (BrowseComp benchmark) and computer use (OSWorld-Verified), Sonnet 5 offers a dramatically better price-to-performance ratio than its predecessor, Sonnet 4.6. At higher effort levels, it even matches Opus 4.8 on some tasks.

From a safety perspective, Sonnet 5 is overall safer than the previous generation — it better refuses malicious requests and resists abuse attempts. Its cybersecurity capabilities are substantially lower than Opus and Mythos models — in tests for developing Firefox exploits, it never once produced a fully working exploit.

Pricing and Availability

Sonnet 5 launches with an introductory price of $2 per million input tokens and $10 per million output tokens (through August 31, 2026). After that, pricing increases to $3 and $15 respectively. For comparison, Opus 4.8 costs $5 for input and $25 for output. The model is immediately available across all plans — from Free through Pro, Max, Team, to Enterprise. European users, including those in the Czech Republic, have full access through Claude.ai and the API.

Fable 5 and Mythos 5: Back After Three Weeks

Alongside the Sonnet 5 launch came a major relief for Anthropic. The US government lifted export controls on June 30 on the Fable 5 and Mythos 5 models, which had been completely unavailable since June 12.

Here's what happened: On June 9, Anthropic released Fable 5 and Mythos 5 — two models built on the same foundation but with different safety guardrails. Within three days, Washington intervened. The trigger was a report from Amazon researchers who found a technique to bypass Fable 5's safety filters and get the model to identify software vulnerabilities, including demonstrating one exploit.

Anthropic subsequently found that many older and less capable models exhibited the same behavior — including Claude Haiku 4.5, Sonnet 4.6, Opus 4.8, GPT-5.4, GPT-5.5, and Kimi K2.7. No unique Mythos-level cyber capabilities were exposed. Nevertheless, the company worked with the government to train an improved safety classifier that now blocks the technique in over 99% of cases.

As of July 1, Fable 5 is once again available globally — including in the European Union and Czech Republic — on Pro, Max, Team, and select Enterprise plans. Through July 7, it's included in up to 50% of weekly usage limits, after which it will be billed via usage credits. Mythos 5 has been restored for select US organizations under the Glasswing program.

"Long Cat" from China: LongCat 2 Sidesteps US Restrictions

While Washington focuses on regulating domestic AI companies, China is producing models that could render the entire censorship effort moot. According to Hospodářské noviny, on the same day as Sonnet 5's release, LongCat 2 was unveiled — an open-weight model from China developed without Nvidia chips.

LongCat 2 comes with an open license, meaning anyone can download its weights, run it on their own hardware, and further modify it. This is a fundamental difference from Anthropic or OpenAI models, which operate only as closed cloud services. No US export control can prevent the spread of an open-source model.

And this isn't an isolated case. Even before LongCat, GLM-5.2 from the Chinese company Z.ai had emerged — an open model with a million-token context window that, according to independent testing, can find security vulnerabilities in software at roughly one-sixth the cost of Anthropic's models. On coding and long-horizon benchmarks, GLM-5.2 beats GPT-5.5 and in some tests approaches Claude Fable 5.

The Paradox of US Regulation

The situation creates a peculiar paradox: the more strictly the US regulates its own AI companies, the more space opens up for Chinese open-source alternatives. While Fable 5 was completely unavailable for three weeks and OpenAI had to immediately restrict its new Sol, Terra, and Luna models to selected partners, Chinese models are freely downloadable from GitHub and Hugging Face.

For European companies and developers, this poses a practical question: rely on regulated American models whose availability can be cut off overnight by political decisions, or turn to open Chinese alternatives that no one can shut down? The European Commission has previously warned about over-reliance on American models — and current developments are proving that warning correct.

Meanwhile, Anthropic is working on a systemic response. Together with Amazon, Microsoft, and Google, it's developing a unified framework for assessing jailbreak severity that should help the entire industry respond faster and more consistently to security findings. The company is also deepening its collaboration with the US government — offering pre-release model access for independent testing, threat intelligence sharing, and joint research capacity.

What This Means for Czech Users

For the average Czech Claude.ai user, the main headline is that Fable 5 is once again available in the Czech Republic — after three weeks of downtime. Sonnet 5 as the new default model for Free and Pro plans delivers significantly better performance without a price increase.

Businesses and developers, however, should consider the broader implications. If you're building a product on a closed American API, you're vulnerable to Washington's political decisions. Open-source models like GLM-5.2 or LongCat 2 eliminate this dependency — you can run them on your own infrastructure, and no export ban can shut you down. The trade-off, of course, is the need for your own computing power and technical expertise.