Anthropic was officially released on June 30, 2026Claude Sonnet 5, the most powerful agent model in its Sonnet series to date. Compared with the previous generation Sonnet 4.6, Sonnet 5 has achieved leaps and bounds in the core capabilities of agents such as reasoning, tool use, programming and knowledge work, while the pricing maintains the consistent cost-effectiveness advantage of the Sonnet series. More importantly, it brings Opus-level agent capabilities to the mid-range price range for the first time - Anthropic says its performance is "close to Opus 4.8, but at a much lower price."
core competencies
The biggest selling point of Sonnet 5 isA huge jump in the capabilities of intelligent agents. It's able to make plans, use tools like browsers and terminals, and run autonomously for long periods of time - tasks that just a few months ago would have required larger, more expensive models, the Sonnet 5 can now do.
- Agentic Coding: It scored 63.2% in the SWE-Bench agent programming benchmark, far exceeding Sonnet 4.6’s 58.1% and approaching Opus 4.8’s 69.2%. Ability to independently handle multi-step software engineering tasks, including code writing, tool invocation, and debugging.
- knowledge work: even in some knowledge work benchmarksSlightly more than Opus 4.8, outstanding performance in in-depth research, subtle judgment, etc.
- Complete tasks autonomously: Early testers reported that Sonnet 5 can complete complex tasks that would “stop midway” in previous versions, andProactively check your own output, no explicit instructions are required.
- contextual understanding: Supports 1M token context window, suitable for processing large-scale code bases and complex documents.
Pricing and availability
Sonnet 5’s pricing strategy is extremely competitive. From now until August 31, 2026, enjoyInitial discount price: USD 2 per million input tokens, USD 10 per million output tokens. Starting from September 1st, standard pricing will resume: input $3/output $15. Compared to Opus 4.8’s $5/$25,Claude Sonnet 5 offers approximately 60% lower cost while providing similar performance.
Sonnet 5 is now available as a Free and Pro planDefault modelOpen to all users and supports Max, Team and Enterprise plans. Developers can useClaude API usage model identifierclaude-sonnet-5, and can also be called on AWS, Microsoft Foundry and other platforms.
safety assessment
Anthropic's security assessment shows Sonnet 5's overall bad behavior rateLower than Sonnet 4.6. In terms of agent security, the model is better at rejecting malicious requests and resisting hint injection attacks. Rates of hallucinations and pandering behavior are also lower than in previous generations.
However, Sonnet 5 is still far inferior to Opus 4.8 and Mythos 5 in terms of network security capabilities. Complete success rate of Sonnet 5 in Firefox exploit testsis 0%. Anthropic is enabled by default for Sonnet 5Network security real-time protection(Cyber Safeguards), detect and block dangerous network operations.
User experience/limitations
Many early partner companies have given positive feedback.ZapierEngineers said "it completed end-to-end tasks that would have failed halfway before";LovableThe co-founders praised its "clean and consistent rejection of unsafe requests"; Sourcegraph commented on its "continuous coding, tool usage and debugging capabilities in complex technical environments."
limitations: The tokenizer of Sonnet 5 has been updated. The same input may generate more tokens (about 1.0-1.35 times), but the initial discount price has taken this into account. Furthermore, Opus 4.8 is still the better choice for tasks that require the highest precision, such as complex network security work. Anthropic recommends users switch between Sonnet 5 (price/performance) and Opus 4.8 (high accuracy) as needed.
Overall Score
| 维度 | Score | evaluate |
|---|---|---|
| functional completeness | 9.0 / 10 | Agent capabilities, programming, and knowledge work are all close to the Opus 4.8 level, 1M context, but some high-precision tasks are still weaker than the flagship model |
| 易用性 | 9.2 / 10 | All plans are available by default, and the API is simple and easy to access.Claude Code integration is smooth |
| Cost-effectiveness | 9.5 / 10 | The price starting from $2/$10 reaches 90% of the performance of Opus 4.8, and it is extremely cost-effective during the first launch discount period. |
| 中文支持 | 8.5 / 10 | The Chinese ability is stable, but the understanding and expression of Chinese in professional fields are slightly inferior to OpenAI and domestic models. |
| 输出质量 | 9.0 / 10 | Code quality and knowledge work output are reliable, independent inspection output reduces errors, and the hallucination rate is lower than the previous generation |
Overall rating: 9.0/10
Claude The release of Sonnet 5 marks an important turning point in the AI industry:Agent capabilities are no longer the exclusive privilege of flagship models. Anthropic packs close to Opus-level agent capabilities into a mid-range price, which means that the threshold and cost for developers to build AI Agents will drop significantly. For enterprises, the combination of Sonnet 5 and Opus 4.8 - which can adjust between performance and cost on demand - offers an attractive proposition. Anthropic’s move also brought direct price pressure to OpenAI’s GPT-5.6 and Google’s Gemini 3.5 series.
Want to know more AI tool reviews? access AI Dash Discover the best AI tools.