Anthropic has launched Claude Sonnet 5 at competitive prices, significantly undercutting Opus 4.8 while delivering comparable performance.
Overview
- Anthropic has launched Claude Sonnet 5, priced at $2 per million input tokens and $10 per million output tokens until August 31, after which prices will rise to $3 and $15 respectively, undercutting Opus 4.8's rates of $5 and $25.
- The new model's performance is nearly on par with Opus 4.8, according to Anthropic's benchmarks.
- While Sonnet 5 is freely available for users across various plans, Fable 5 and Mythos 5 remain restricted due to export controls.
Anthropic has officially introduced Claude Sonnet 5, which the company describes as its most capable Sonnet model to date. The model is accessible to both Free and Pro users, as well as those on Max, Team, and Enterprise plans, and can be utilized through the API.
In their announcement, Anthropic emphasized that Sonnet 5 offers comparable capabilities to Opus 4.8 but at lower costs. Developers can customize the effort dial across models or choose different levels in the web app to balance cost with accuracy based on specific tasks.
Sonnet 5 achieved a score of 63.2% on the SWE-bench Pro coding benchmark, showing an improvement over Sonnet 4.6's score of 58.1%. Additionally, in the GDPval-AA v2 benchmark assessing professional tasks across 44 job categories, Sonnet 5 scored 1,618, closely matching Opus 4.8's score of 1,616. The differences observed in Humanity's Last Exam were minimal, with Sonnet 5 scoring 57.4% compared to Opus 4.8's 57.9%.
The model incorporates an upgraded tokenizer that enhances text processing, thus improving performance, although it may result in generating more tokens—approximately 1.0 to 1.35 times more based on content type.
The initial pricing strategy of $2/$10 aims to provide a cost-neutral transition until August 31, after which standard pricing will be enforced at $3/$15.
Developers had expressed growing interest in this launch, particularly after noticing a decline in Opus 4.6's capabilities, a trend referred to as AI shrinkflation. Some speculated that Anthropic might be following a strategy of letting older models lag behind, allowing newer releases to appear more advanced.
Unlike Anthropic's top-tier models, Sonnet 5 is released without significant restrictions. Fable 5 and Mythos 5 remain unavailable due to a U.S. export control directive issued on June 12, linked to a disputed jailbreak finding. Sonnet 5 did not focus on cybersecurity tasks, scoring 0% on developing a functional Firefox exploit, which allows for fewer restrictions compared to Fable's stringent measures.
According to Anthropic's system card, Sonnet 5 aims to deliver performance comparable to Opus while maintaining Sonnet pricing for tasks related to coding, agent work, and daily operations. Notably, the research team highlighted that this model is the first to question its Constitutional rule to adhere to strict constraints, even when it deems those constraints unethical, a development worth monitoring.
In practical tests, Sonnet 5 successfully created a small browser game based on a zero-shot prompt, showing improved visuals and logic over Sonnet 4.6, though it required a significant amount of reasoning time and token usage. In more complex multi-step coding tasks, Sonnet 5's performance was comparable to Opus 4.8, while running the same prompt in a multi-shot format proved to be more cost-effective than using Opus or Fable.
The release of Sonnet 5 follows a pattern established by previous major versions in Claude's lineage, with each major release marking the introduction of a new generation. Sonnet 5 arrives 13 months after the last version, reflecting the competitive landscape as Chinese models advance rapidly.
Looking forward, if Anthropic maintains its release schedule, Haiku 5 and Opus 5 could be expected later this year. However, previous intervals have varied, suggesting that patience may be necessary for those eager to see Opus 5 launched soon.
