Anthropic is facing a major problem. While Fable 5—by far their most powerful AI model—is supposed to be accessible only to U.S. users, the company behind Claude chose to withdraw the LLM entirely. Weeks later, Claude Sonnet 5 has now surfaced, a release that seems somewhat anticlimactic at first glance.
Benchmarks from an AI model developer often set the tone. Independent tests later reveal just how good a new LLM really is. This also applies to Claude Sonnet 5, which has been available for a few hours as of this writing. Sonnet 5 and Opus 4.8—a larger and more expensive model—are virtually neck-and-neck on the benchmarks. Both perform roughly equally well in coding, reasoning, computing tasks, and knowledge work. Anthropic emphasizes the agentic nature of the latest model, which is better than ever.
Don’t worry about security
For years, Anthropic has cultivated a habit of generating an ungodly amount of fear, uncertainty, and doubt surrounding new model releases. Nothing could be further from the truth this time: Sonnet 5 is actually meant to come across as harmless and mundane. The U.S. government banned the use of Fable 5 and Mythos 5 by non-U.S. citizens due to an alleged threat to national security. A “cyber-superweapon”—that’s how Anthropic CEO Dario Amodei described the characterization he heard from early users of Mythos Preview. After countless models hit the market without much scrutiny, that unchecked pace has been interrupted.
With Sonnet 5, Anthropic apparently has little to fear. Even though there are additional security guardrails—more so than with its predecessor, Sonnet 4.6—these measures are merely on par with the protections Anthropic provided for the use of Opus 4.7 and 4.8.
Returning to the restrictions: OpenAI is also feeling the consequences, with a limited rollout of GPT-5.6 in its three variants. GPT-5.6 Sol Ultra can be viewed as a kind of Fable (and scores higher than Mythos/Fable on OpenAI’s own benchmarks); Sol itself is a kind of Opus; Terra is a potential competitor to the Sonnet series; and Luna is a compact model similar to Gemini 3.5 Flash or Claude Haiku.
Not quite Opus and not quite attractive enough?
According to Anthropic, Sonnet 5 bridges the gap with the more powerful Opus 4.8—but at a lower price. According to Anthropic, it’s a significant improvement over its predecessor, Sonnet 4.6, in areas such as complex reasoning, tool usage, coding, and knowledge work. Starting today, it’s the default model for Free and Pro users and available to Max, Team, and Enterprise customers, as well as in Claude Code and on the Claude Platform. With a pop-up focused on everyday AI use—such as planning a vacation—Anthropic also appears to be pursuing a Gemini- or ChatGPT-like approach alongside this release.
The pricing—which is more relevant to professional users who typically focus on coding—is noteworthy. Through August 31, Anthropic is charging an introductory rate of $2 per million input tokens and $10 per million output tokens. After that, the rates will increase to $3 and $15, respectively. By comparison, Opus 4.8 costs $5 and $25 per million tokens. Between the two models, users can adjust the effort level to balance cost and performance.
Of course, that doesn’t tell the whole story. Consider GLM-5.2 from the Chinese company Z.ai, an open-source model that, based on figures from the renowned Artificial Analysis, performs on par with some recent Claude Opus models. However, it is five times cheaper per token… but consumes far more tokens per output than Opus typically does. In that regard, Sonnet 5 could prove to be an interesting competitor via the API.
Safer in an agentic context
Anthropic states that Sonnet 5 exhibits less undesirable behavior than Sonnet 4.6 and is more resistant to prompt injection attacks. The model also hallucinates less. Nevertheless, in the internal audit, it shows a higher percentage of “misaligned” behavior than the more capable Opus 4.8 and Mythos Preview.
When it comes to cybersecurity tasks, Sonnet 5 actually performs deliberately poorly. In a way, it’s quite comical to see Anthropic emphasizing this topic: don’t expect much from this brand-new model in this specific area. In a test developed in collaboration with Mozilla to write exploits for Firefox vulnerabilities, the model never succeeded in building a working exploit. Furthermore, Anthropic has enabled cybersecurity safeguards by default—the same ones used in Opus 4.7 and 4.8.
Of course, this caution didn’t come out of nowhere. Fable 5 and Mythos 5 have been blocked since mid-June and have shown no signs of returning. Meanwhile, the policy has shifted to controlled access for Mythos 5 in particular, with the U.S. government determining which organizations are allowed to use the most powerful models.
Modified tokenizer
Sonnet 5’s seemingly weaker performance on paper compared to Fable 5 and Opus 4.8 is to be expected given this smaller-scale LLM. Nevertheless, Sonnet models have built a solid reputation for helping users more easily stay within their five-hour and weekly token budgets under their subscription plans. Now, the step forward appears to be primarily architectural.
Sonnet 5 uses a new tokenizer that processes text differently. As a result, the same input can yield more tokens—approximately 1.0 to 1.35 times as many, depending on the type of content. Anthropic states that the introductory price has been set so that the transition is roughly cost-neutral. Rate limits have also been increased in Chat, Cowork, Claude Code, and the Claude Platform. Developers can access the model via the Claude API as claude-sonnet-5.