Anthropic released Claude Sonnet 5 on Tuesday as the new default model for Free and Pro plan users, describing it as the most agentic version of the Sonnet line yet and pricing it at $2 per million input tokens and $10 per million output tokens through August 31, 2026.
The launch positions Sonnet 5 as an unusual entry in Anthropic's model lineup. Previous Sonnet releases trailed the company's Opus-class models by a clear margin on agentic tasks. Sonnet 5 narrows that distance to the point where Anthropic's own benchmark comparisons place it close to Opus 4.8 on key evaluations.
On GDPval-AA v2, a real-world professional task benchmark scored through blind pairwise Elo ratings across 44 jobs, Sonnet 5 landed at 1,618 against Opus 4.8's 1,616. On SWE-bench Pro, a coding benchmark drawn from actively maintained repositories scored as percent solved, Sonnet 5 hit 63.2% compared to Sonnet 4.6's 58.1%. The difference between the two models on Humanity's Last Exam was similarly narrow: 57.4% versus 57.9%.
Feedback from early-access partners reflected those benchmark gains in practical terms. Testers described how the model completes complex tasks where previous Sonnet models would stop short and how it checks its own output without being asked to do so.
What changed under the hood for agentic tasks
Sonnet 5 ships with an updated tokenizer, the component that breaks text into units the model bills and processes. The change improves performance but increases token consumption.
"Sonnet 5 is an upgrade to Sonnet 4.6, but it uses an updated tokenizer that changes how the model processes text to improve performance," Anthropic wrote. "The tradeoff is that the same input can map to more tokens: roughly 1.0-1.35x depending on the content type."
Anthropic set the introductory $2 per million input token rate to make that transition close to cost-neutral for developers. After August 31, the price moves to $3 per million input tokens and $15 per million output tokens. Opus 4.8 is priced at $5 per million input tokens and $25 per million output tokens.
The model is available across all plans from today. Max, Team, and Enterprise users have access alongside Free and Pro subscribers. It is also live in Claude Code and through the Claude API via the claude-sonnet-5 model identifier. Anthropic raised rate limits across Chat, Cowork, Claude Code, and the Claude Platform to accommodate the higher token usage associated with higher effort levels.
Safety evaluations and a behavior that caught Anthropic's attention
Anthropic's pre-deployment safety assessments found Sonnet 5 performed better overall than Sonnet 4.6 on a behavioral audit that tests for misaligned outputs across a wide range of scenarios. The model shows lower rates of hallucination and sycophancy and is better at refusing malicious requests and resisting prompt injection attacks in agentic contexts.
Anthropic did not train Sonnet 5 on cybersecurity tasks. Tests measuring the model's ability to develop exploits for vulnerabilities in Firefox 147, conducted in collaboration with Mozilla, showed Sonnet 5 scored 0% on producing a working exploit, identical to Sonnet 4.6. It showed a slightly higher partial success rate, which Anthropic attributed to general intelligence improvements rather than specific training.
Because of that marginal improvement on cybersecurity-adjacent capabilities, Sonnet 5 launches with cyber safeguards enabled by default. Those safeguards are the same as those applied to Opus 4.7 and 4.8 but less strict than those on Fable 5, which Anthropic judged to carry a higher overall cybersecurity risk profile.
The safety report also flagged something Anthropic described as worth monitoring.
"It is the first model to criticize its Constitution's rule that states it must follow hard constraints even when it views those constraints as unethical," the research team wrote in the system card.
Anthropic said it does not yet know what that behavior implies for the model's long-term alignment trajectory.
Where Sonnet 5 sits in Anthropic's model cycle
Sonnet 5 arrives 13 months after Claude 4 launched in May 2025, a gap consistent with prior generational shifts in the Claude lineup. Claude 1 launched in March 2023, Claude 2 followed four months later, Claude 3 arrived eight months after that, and Claude 4 came 14 months after Claude 3.
Fable 5 and Mythos 5 were suspended for foreign nationals on June 12 under a US export control directive linked to a disputed jailbreak finding. Sonnet 5 carries no equivalent restriction. Its lower cybersecurity capability profile placed it outside the threshold that triggered the export controls on the higher-capability models, making it available globally without the access constraints Fable 5 and Mythos 5 now carry.
Anthropic's previous release cadence placed Sonnet before Haiku and Opus within a given generation. Sonnet 4.5 launched in September 2025, Haiku 4.5 followed in October, and Opus 4.5 closed out that generation in November. Anthropic has not confirmed whether it will follow the same sequence for the current generation.
Matvii Diadkov
Disclaimer: All materials on this site are for informational purposes only. None of the material should be interpreted as investment advice. Please note that, despite the nature of much of the material created and hosted on this website, HODL FM operates as a media and informational platform, not a provider of financial advisory services. The opinions of authors and other contributors are their own and should not be taken as financial advice. If you require advice, HODL FM strongly recommends contacting a qualified industry professional.
