Anthropic launches Sonnet 5: rivals Opus 4.8 at lower cost, discounted through August

Anthropic unveiled Sonnet 5 this week, positioning it as the company's most capable mid-tier model to date. The release aims to narrow the performance divide between the Sonnet and Opus families, while offering significant cost savings for developers.

Performance gains across reasoning and coding#

Sonnet 5 delivers measurable improvements over its predecessor, Sonnet 4.6, particularly in areas like complex reasoning, tool use, and software development. On several benchmarks—including OSWorld-Verified and agentic search tasks—the model achieves performance levels comparable to Opus 4.8 when running at the "Extra High" reasoning setting. However, this higher reasoning level also comes with proportionally higher costs, meaning Opus 4.8 remains preferable for tasks where accuracy demands justify the premium.

Anthropic's internal testing revealed that Sonnet 5 often succeeds where earlier Sonnet versions would terminate prematurely on intricate multi-step tasks, suggesting real-world usability gains beyond what benchmark scores alone capture.

Introductory pricing and availability#

To encourage adoption, Anthropic is running a limited-time promotion through August 31. API users can access Sonnet 5 at $2 per million input tokens and $10 per million output tokens—well below the standard rates of $3/$15 per million that will take effect afterward. A company spokesperson framed this as a "migration window," allowing teams to evaluate Sonnet 5 against their production workloads at minimal cost. Anthropic has also temporarily raised rate limits for Chat, Cowork, and Claude Code users to support heavier token consumption from higher reasoning levels.

Safety considerations and cybersecurity stance#

Anthropic deliberately kept Sonnet 5's cybersecurity safeguards intact while avoiding specialized training on exploit discovery. In tests involving software like Firefox 147, the model showed a marginally higher rate of partial security breaches compared to Sonnet 4.6, but remained unable to develop fully functional exploits. This measured approach—maintaining guardrails without aggressive defensive overtraining—reflects lessons from the regulatory pressures that led to the swift withdrawal of Fable 5. The risk of government intervention against Sonnet 5 appears minimal given its limited offensive capability in this domain.