Anthropic has announced Claude Sonnet 5, a new language model built specifically for agentic use[1]. Its ability to make plans, use tools such as browsers and terminals, and run tasks autonomously has improved substantially, delivering work that only larger, more expensive models could handle a short time ago, now at a more affordable price[1]. This article walks through the key points on performance, pricing, and safety evaluations.
Closing the gap with Opus 4.8 while surpassing Sonnet 4.6
According to Anthropic, the era of agentic AI (AI that judges and carries out work autonomously) began in earnest with Sonnet-class models[1]. Claude Sonnet 3.5, 3.6, and 3.7 were the first models to show strong skills in coding and tool use[1]. More recently, however, the clearest gains in capability had come from the higher-end Opus class[1].
Sonnet 5 is positioned to narrow that gap[1]. Its performance approaches that of the higher-end Claude Opus 4.8 while keeping prices down[1]. Compared with its predecessor Sonnet 4.6, it shows steady improvements in the core aspects of agentic performance: reasoning, tool use, coding, and knowledge work[1].
Anthropic presents comparisons on two evaluations: BrowseComp, which measures agentic search, and OSWorld-Verified, which measures computer use[1]. On these benchmarks Sonnet 5 clearly outperforms Sonnet 4.6, and while Opus 4.8 remains the better choice when the highest accuracy is required, Sonnet 5 offers higher-quality options than before at a lower price[1]. Between Sonnet 5 and Opus 4.8, users can adjust the "effort" (how much processing effort to apply) to find the right balance of cost and performance[1].
Feedback from early access partners
Anthropic also shares impressions from partner companies that tested the model ahead of its release[1]. A common theme was its autonomy: it "finishes complex tasks where previous Sonnet models would have stopped short" and "checks its own output without being asked"[1].
The development service Lovable, for example, praised how cleanly and consistently it refuses unsafe requests[1]. In the legal field, Eve said its price-to-performance ratio for plaintiff-side legal tasks made migration an easy choice, while the data analytics platform ClickHouse reported that it reaches answers faster in tighter steps[1]. Pace, which runs computer-use agents for insurance workflows, also noted that it consistently takes the correct action quickly[1].
Results of the safety evaluations
Anthropic has published the results of its pre-deployment safety evaluations[1]. Overall, Sonnet 5 is an improvement on Sonnet 4.6, with better refusal of malicious requests and greater resistance to prompt injection (attacks that try to hijack the model's instructions)[1]. Rates of hallucination (stating things that are not true) and sycophancy (excessively agreeing with the user) also fell[1].
On the other hand, compared with the more capable Opus 4.8 and others, it showed somewhat higher rates of certain undesirable behaviors[1]. On cybersecurity, Anthropic explains that Sonnet 5 was not deliberately trained for such work and is substantially weaker than higher-end models at dangerous tasks such as exploiting software vulnerabilities[1]. Because its capability is slightly higher than the previous generation, however, it ships with cyber safeguards—which detect and block dangerous use—enabled by default[1]. The full evaluation results are compiled in the "Claude Sonnet 5 System Card"[1].
Availability and pricing
Claude Sonnet 5 is available across all plans from the same day as the announcement[1]. It is the default model on the Free and Pro plans and is also available to Max, Team, and Enterprise users[1]. It can be used in Claude Code and on the Claude Platform, and called via the API as "claude-sonnet-5"[1].
As an introductory price, it is set at 2 USD (about 320 yen) per million input tokens and 10 USD (about 1,620 yen) per million output tokens through August 31, 2026[1]. After that period it moves to standard pricing of 3 USD (about 490 yen) for input and 15 USD (about 2,430 yen) for output[1]. ※1 USD = 162 JPY (as of July 1, 2026)
Note that Sonnet 5 uses a new tokenizer (the mechanism that splits text into tokens), so the same input may map to 1.0–1.35 times more tokens than before[1]. The introductory pricing is set so that, even accounting for this change, the cost of switching is roughly unchanged[1]. Rate limits (the cap on usage per fixed period) were also raised across Chat, Cowork, Claude Code, and the Claude Platform[1].
Summary
Claude Sonnet 5 is an agentic-focused model designed to deliver performance approaching the higher-end Opus 4.8 at a lower price. Alongside improvements in reasoning, coding, and tool use, it also advanced over its predecessor in safety evaluations—though the fact that the price rises after the introductory period, and the practical cost implications of the tokenizer change, are worth checking before use.
