What Changed: A New Cost-Performance Frontier
Anthropic released Claude Sonnet 5 on June 30, 2026, positioning it as the most agentic Sonnet model yet. For many developers, the agentic AI era began with Sonnet-class models: Claude Sonnet 3.5, 3.6, and 3.7 were the first to show impressive coding and tool-use skills. However, recent gains in agentic capabilities have been concentrated in Opus-class models. Sonnet 5 narrows that gap: its performance is close to Opus 4.8, but at lower prices.
The model is priced at $2 per million input tokens and $10 per million output tokens, a price that became permanent on August 10, 2026 (an earlier plan to raise it to $3/$15 in September was scrapped). This is a substantial improvement over its predecessor, Sonnet 4.6, on key agentic dimensions: reasoning, tool use, coding, and knowledge work.
How It Works: Effort Levels and Benchmarks
Anthropic compared Sonnet 5, Sonnet 4.6, and Opus 4.8 on the agentic search evaluation BrowseComp and the computer use evaluation OSWorld-Verified, at different “effort” levels. The effort parameter lets you control how much compute the model uses, trading cost for performance.
On these benchmarks, Sonnet 5 (orange line) is a strict improvement over Sonnet 4.6 (gray line) and covers a much wider range of cost-performance options than Opus 4.8 (yellow line). It provides substantially improved cost efficiency at medium effort; at higher effort, its performance can match Opus 4.8 on some tasks. This means you can dial in the right balance of cost and performance for your specific use case.
Early access partners reported that Sonnet 5 is much more agentic than its predecessors. Testers described how it finishes complex tasks where previous Sonnet models would stop short, how it checks its own output without being asked, and how it does all this at an attractive price point. For example, one tester said: “I asked Claude Sonnet 5 to investigate a bug. Unprompted, it wrote a reproducing test, implemented the fix, then stashed it to confirm the bug came back without the change. All in a single pass.”
Practical Use Cases: From Software Engineering to Insurance
Partner feedback highlights Sonnet 5’s follow-through and cost efficiency across domains:
- Software engineering: One partner noted it “handles sustained coding, tool use, and debugging well across messy technical contexts,” especially for workflows where follow-through and technical grounding matter. Another said it “carried each one through to a tested, verified result on its own,” freeing engineers to focus on judgment and final sign-off.
- Legal research: Eve, a legal AI platform, said Sonnet 5 “sits on the Pareto frontier” for plaintiff-law tasks, with clearest gains in legal research and analysis at a price-to-performance ratio that made migration easy.
- Insurance automation: Pace, which runs computer-use agents for insurance workflows, said Sonnet 5 “consistently takes the right action and does it quickly,” which is what real insurance work demands.
- Data exploration: ClickHouse agents use Sonnet 5 to explore live data and produce insights on the fly, noting it “reasons in tighter steps and gets our users to answers noticeably faster.”
These examples illustrate how Sonnet 5 can handle multi-step, end-to-end tasks that previously required more expensive models, making agentic automation more accessible.
Limitations and Trade-offs
Anthropic’s pre-deployment safety evaluations found Sonnet 5 overall safer than Sonnet 4.6. It is better at refusing malicious requests and resisting prompt injection attacks, and shows lower rates of hallucination and sycophancy. On their automated behavioral audit, Sonnet 5 scored lower (safer) overall than Sonnet 4.6, but somewhat higher rates of misaligned behavior compared to Opus 4.8 and Claude Mythos Preview.
Notably, Anthropic did not deliberately train Sonnet 5 on cybersecurity tasks. On evaluations testing dangerous cyber skills, such as developing software exploits, it performed substantially worse than Opus 4.8 and Mythos 5. In a Firefox exploit development evaluation, Sonnet 5 never successfully developed a full working exploit (0.0% success), though it had a slightly higher partial success rate than Sonnet 4.6. Because of this, Anthropic launched Sonnet 5 with cyber safeguards enabled by default, the same as those in Opus 4.7 and 4.8, but less strict than those for Fable 5.
For product builders, this means Sonnet 5 is a strong choice for general agentic tasks, but if you need advanced cybersecurity capabilities (e.g., penetration testing), you may still need Opus-class models. Also, while misaligned behavior rates are lower than Sonnet 4.6, they are higher than Opus 4.8, so high-risk applications warrant careful evaluation.
Concrete Takeaway: Test Effort Levels on Your Workload
Claude Sonnet 5 is available across all plans: it is the default for Free and Pro, and available to Max, Team, and Enterprise users. Developers can access it via the Claude API using the model name claude-sonnet-5. Anthropic has also increased rate limits across Chat, Cowork, Claude Code, and the Claude Platform to accommodate the higher token usage of higher effort levels.
If you’re evaluating Sonnet 5, the key is to test different effort levels on your typical tasks. The cost-performance curve is wider than Opus 4.8, so you can find a sweet spot that balances cost and quality. For many workloads, Sonnet 5 may deliver near-Opus performance at a fraction of the cost, making it a practical choice for scaling agentic AI in production.
For full details, refer to the Claude Sonnet 5 System Card, which reports a broader set of evaluations and safety results.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.