Anthropic makes Claude Sonnet 5 default across plans

Anthropic has launched Claude Sonnet 5, making it the default Sonnet model across several user plans.

The release puts Sonnet 5 at the centre of Anthropic's mainstream product line. It is available to Free, Pro, Max, Team and Enterprise users, as well as developers using the Claude API, Claude Code and the Claude Platform.

The model is designed to handle more autonomous tasks than earlier Sonnet versions, including planning, using tools such as browsers and terminals, and carrying out multi-step work with less user intervention. Anthropic positions it as closer in performance to its larger Opus-class models while keeping pricing below that range.

That marks a shift in how Anthropic is pitching the Sonnet tier. Earlier Sonnet releases were framed as useful for coding and tool use, while the strongest recent gains in autonomous behaviour were associated with Opus models. Sonnet 5 is intended to narrow that gap.

Pricing and access

For developers, Anthropic is offering introductory pricing of USD $2 per million input tokens and USD $10 per million output tokens, before moving to USD $3 per million input tokens and USD $15 per million output tokens. Sonnet 5 is also available through all existing Claude plans and is now the default model for Free and Pro users.

Anthropic has raised rate limits across Chat, Cowork, Claude Code and the Claude Platform to account for the higher token usage tied to higher effort settings. Users can adjust those settings depending on the task and the balance they want between cost and output quality.

Sonnet 5 also uses an updated tokenizer, changing how text is processed and, in some cases, increasing token counts for the same input. Anthropic says the introductory pricing is intended to make the transition roughly cost-neutral for users moving from Sonnet 4.6.

Performance claims

Internal evaluations showed Sonnet 5 outperforming Sonnet 4.6 in reasoning, coding, tool use and general knowledge work, according to Anthropic. The company also says the model can match Opus 4.8 on some tasks at higher effort levels, while covering a broader range of price and performance options than the larger model.

Anthropic cited tests including BrowseComp, which measures agentic search, and OSWorld-Verified, which assesses computer use. It says Sonnet 5 improves clearly on Sonnet 4.6 across these benchmarks and offers stronger cost efficiency at medium effort levels.

Anthropic also pointed to feedback from early access partners, who reported that Sonnet 5 was more likely to complete complex tasks, check its own output and continue working through problems that earlier Sonnet models would leave unfinished.

Safety measures

Alongside the performance push, pre-deployment safety assessments found Sonnet 5 showed a lower overall rate of undesirable behaviour than Sonnet 4.6, according to Anthropic. The company says the model was better at refusing malicious requests and resisting prompt injection attacks in agentic settings.

Anthropic also reported lower rates of hallucination and sycophancy than the previous Sonnet release. On its automated behavioural audit, which tests for conduct such as deception or cooperation with misuse, Sonnet 5 received a safer overall score than Sonnet 4.6, it said.

Still, Sonnet 5 did not outperform all of Anthropic's larger models on those measures. The company reported somewhat higher rates of misaligned behaviour than Opus 4.8 and Claude Mythos Preview on the same assessment.

Cybersecurity was a separate focus in the safety review. Anthropic said it had not deliberately trained Sonnet 5 on cyber tasks and that the model performed substantially worse than Opus 4.8 and Mythos 5 on tests involving potentially dangerous cyber skills, such as developing software exploits.

On one evaluation involving vulnerabilities in the Firefox browser, Sonnet 5 was never able to produce a full working exploit, according to Anthropic. However, the model showed a slightly higher rate of partial success than Sonnet 4.6, which the company attributed to broader intelligence gains rather than specific cyber training.

Because of that change, Anthropic has enabled cyber safeguards by default for Sonnet 5. The safeguards are designed to detect and block dangerous cyber usage in real time and are the same as those used with Claude Opus 4.7 and 4.8.

The launch highlights how competition among AI model providers is moving beyond raw benchmark scores towards a mix of price, autonomy and operational safeguards. By making Sonnet 5 the default model for its mass-market plans, Anthropic is putting that trade-off in front of a wider group of users than before.