Model: Claude Sonnet 5, compared with Sonnet 4.6 and Opus 4.8.
Evaluation: The official presentation shows BrowseComp (Agentic Search) and OSWorld-Verified (computer use), and reports additional capability and safety evaluations in the System Card.
Inference control: effort tiers; official charts compare cost and performance at different effort levels.
Pricing: $2 per million input tokens and $10 per million output tokens; Anthropic says this introductory price became the permanent price on August 10.
API model name: claude-sonnet-5.
Availability: The default model on Claude Free/Pro; also available on Max, Team, and Enterprise, and through the Claude API.
The agent evaluations use tool calls, long-running tasks, and different effort tiers; complete figures should be taken from the official charts and System Card.
The official conclusion is that Sonnet 5 is clearly better than Sonnet 4.6 at agentic reasoning, tool use, coding, and knowledge work, with some high-effort tasks approaching Opus 4.8.
In the official cost-performance charts for BrowseComp and OSWorld-Verified, Sonnet 5 offers a broader range of cost options than Opus 4.8; Anthropic particularly emphasizes the cost efficiency of medium effort.
The official safety evaluation says that Sonnet 5's rates of harmful behavior, hallucination, sycophancy, and prompt-injection hijacking are generally lower than Sonnet 4.6's.
Cybersecurity boundary: In the Firefox exploit evaluation, neither Sonnet 5 nor Sonnet 4.6 completed a working exploit (both scored 0%); Sonnet 5's partial success rate was slightly higher. Network-security safeguards were therefore enabled at launch.
The official positioning is a “more agentic Sonnet”: it is suited to multi-step coding, tool calls, and browser/terminal workflows, while effort tiers provide room to choose among cost and completion-rate tradeoffs. Anthropic also explicitly states that its dangerous cyber capabilities are below those of Opus-level models, but that this does not justify omitting safety safeguards.
The official materials consist of vendor testing and partner feedback, not independently reproduced experiments; specific configurations and complete scores for BrowseComp, OSWorld, and other evaluations should be taken from the System Card.
“Approaching Opus 4.8” depends on the task, effort, and budget, and cannot be generalized to equivalent performance across all text or coding tasks.
Pricing and default behavior may still change with the API product; at the time of collection, the page's 2026-08-10 update was used as the reference.
Use the same task set and call Sonnet 5 separately with medium, high, and xhigh effort.
Fix the tool definitions, timeout, maximum total tokens, and retry rules; record the success rate, tool calls, input/output tokens, and wall-clock time.
Rerun Sonnet 4.6 and Opus 4.8 in the same harness, without mixing different system prompts or different tool permissions.
For coding and cybersecurity tasks, separately retain records of safeguards, authorization, and human review; do not infer that a system is risk-free from its failure to complete an exploit.
Official API pricing: $2/M for input and $10/M for output; the official page says this price changed from an introductory price to a permanent price on August 10.
Anthropic describes Sonnet 5 as better able to continue executing, self-checking, and completing deliverables in complex multi-step tasks.
The official safety evaluation reports both a lower overall rate of harmful behavior and the boundary that network-security safeguards still need to be enabled.
Official positioning phrase: “Our most agentic Sonnet yet”.
The official release page treats effort as a cost/performance control, rather than as a single fixed capability level.
Claude Sonnet 5
