Anthropic introduces Claude Sonnet 5, a more affordable agent model.

      Anthropic has introduced Claude Sonnet 5, its most capable mid-tier model to date. It performs comparably to the flagship Opus 4.8 on various tasks, yet is priced at less than half its cost. On June 30, 2026, Anthropic announced that Sonnet 5 is available immediately across all plans. The model is designed not only to respond but to take action; it can devise plans, operate browsers and terminals, and function autonomously for extended periods. Such capabilities previously required larger, more expensive models.

      The offering is straightforward. Sonnet 5 delivers near-flagship performance at a mid-tier price, closely rivaling Anthropic’s most powerful model, Opus 4.8, in reasoning, tool utilization, coding, and knowledge-based tasks. It significantly surpasses its predecessor, Sonnet 4.6, and is much less expensive to operate.

      Focused on affordability

      Pricing is central to this launch. Sonnet 5 is priced at $2 per million input tokens and $10 per million output tokens, with this introductory rate effective until August 31, 2026. After that date, the prices will increase to $3 and $15. In contrast, Opus 4.8 is priced at $5 and $25. TechCrunch has labeled the model as a more economical option for running agents, which is precisely the intention.

      The timing is crucial. Companies hurried to implement AI agents but later hesitated due to the costs involved. Agents operate in loops, use tools, and consume tokens rapidly. Thus, a model that closely aligns with Opus's quality at a much lower price resonates with this concern. It also addresses a market seeking cost reductions following soaring enterprise AI expenses.

      However, there is a caveat in the details. Sonnet 5 employs a new tokenizer, which means that the same text can convert into up to 1.35 times more tokens than previously. Anthropic has set the initial pricing to keep the transition relatively cost-neutral. While the advertised rates seem low, the token count may increase.

      How well does it perform?

      According to Anthropic’s benchmarks, Sonnet 5 is a noticeable upgrade from 4.6, though it hasn't fully matched Opus. In an agentic coding test, it achieved a score of 63.2%, compared to Opus 4.8’s 69.2% and Sonnet 4.6’s 58.1%, based on preliminary reports. In one knowledge work benchmark, it slightly surpassed Opus. Anthropic also features an "effort" dial, allowing developers to balance cost and accuracy between the two models.

      Early users have reported that the model completes complex tasks where earlier Sonnets failed and that it verifies its own outputs independently. However, these assertions stem from the company's promotional materials and should be taken with caution. Independent evaluations will provide the true insights.

      Improved safety with a caveat in cybersecurity

      Anthropic claims that Sonnet 5 demonstrates better safety than 4.6. It more frequently declines malicious requests and shows resistance to prompt-injection attacks, where concealed instructions try to take over the agent. Additionally, it has decreased instances of hallucination and flattery. In an automated audit of misaligned behavior, it recorded safer results than 4.6 but performed worse than Opus 4.8 and the Mythos preview.

      Cybersecurity is a critical concern. Anthropic did not train Sonnet 5 for cybersecurity functions, and it struggled with creating software exploits. In a test conducted with Mozilla on the Firefox browser, the model failed to produce a working exploit. Nevertheless, Anthropic has equipped it with real-time cybersecurity safeguards by default, similar to those used with Opus 4.7 and 4.8. These protective measures are lighter than those surrounding Fable 5, its more restricted public model.

      A well-strategized discount

      The low price is a strategic move, not an act of generosity. Anthropic is in a competitive race to attract developers, and a competent, economical agent model is key to winning their interest. The company also develops a significant portion of its own code using Claude, so an improved, cost-effective Sonnet benefits its engineers as well. Moreover, it is progressing towards a planned public offering, where both revenue growth and developer reach are important.

      The broader context pertains to cost. Operating agents continuously can lead to substantial expenses, and Anthropic has set ambitious revenue goals to support its model developments. Sonnet 5 aims to address both of these challenges: lowering capability costs, keeping developers within the ecosystem, and allowing the effort dial to manage the remaining factors.

      Claude Sonnet 5 is currently active within Claude’s applications, Claude Code, and the API, featuring higher rate limits overall. For most developers, the primary question is no longer whether the model is intelligent enough but whether it is affordable enough for extensive use. Anth