New models
Anthropic launches Claude Sonnet 5: approaching Opus performance for less
New mid-tier model performs nearly as well as Opus 4.8 on many agentic tasks, but for less than half the price. Available from today via API and all Claude subscriptions.
Anthropic has launched Claude Sonnet 5, a model designed to autonomously execute complex tasks. The new model approaches the performance of the much more expensive Opus 4.8, but is more than twice as cheap. According to the developer, it represents a significant improvement over predecessor Sonnet 4.6 in areas such as reasoning, tool use, coding and knowledge work [1].
What has been announced
Claude Sonnet 5 is intended as the most 'agentic' Sonnet model to date. It can make plans, use tools such as browsers and terminals, and operate autonomously at a level that only larger and more expensive models achieved a few months ago [1]. Anthropic's Sonnet line (3.5, 3.6, 3.7) previously demonstrated impressive coding and tool capabilities. Recently, however, the biggest advances in agentic capability were reserved for the Opus class. Sonnet 5 narrows that gap [1].
The model scores comparably to Opus 4.8 on evaluations such as BrowseComp (agentic search) and OSWorld-Verified (computer use), especially at higher 'effort levels'. At medium effort, Sonnet 5 offers substantially better cost efficiency than Opus 4.8 [1]. The model outperforms Sonnet 4.6 on metrics including Humanity's Last Exam, SWE-Bench Verified and various code evaluations [1].
In terms of safety, Sonnet 5 refuses malicious requests more frequently and is better protected against prompt injection attacks than Sonnet 4.6. The model exhibits fewer hallucinations and less sycophantic behaviour. In an automated behavioural audit, Sonnet 5 scored lower — therefore safer — than Sonnet 4.6, though slightly higher than Opus 4.8 and Claude Mythos Preview [1].
Anthropic's evaluation of cybersecurity capabilities shows that Sonnet 5 was never able to develop a working exploit for vulnerabilities in Firefox 147. The model does have a slightly higher 'partial success' rate than Sonnet 4.6, presumably due to general intelligence improvements. Cybersecurity capabilities remain clearly behind Opus 4.8 and Mythos 5 [1]. Nevertheless, cyber-safeguards are enabled by default.
Availability and pricing
Claude Sonnet 5 is available from 30 June 2026 via all subscriptions: it is the default model for Free and Pro, and available for Max, Team and Enterprise users. Developers can use it via the Claude API with the model name `claude-sonnet-5` [1].
Pricing is $2 per million input tokens and $10 per million output tokens [1]. By comparison: Opus 4.8 costs $5 input and $25 output per million tokens. Sonnet 5's introductory price has now been made permanent [1]. Note: the model uses a new tokenizer that can convert the same input into roughly 1.0 to 1.35 times as many tokens, depending on content type [1].
What you can do with it
According to early-access partners, Sonnet 5 completes complex tasks where previous Sonnet models stopped, checks its own output without explicit instruction, and does so at an attractive price [1].
Use cases from the announcement:
- Software engineering and coding: the model continues with coding, tool use and debugging in complex technical contexts, including in 'brownfield code' with race conditions and hidden tests [1].
- Multi-step automation: one user describes a two-part task — updating Salesforce account tiers and sending a launch message — which Sonnet 5 completed from start to finish, where earlier versions got stuck halfway [1].
- Computer use and workflows: for insurance tasks (intake, FNOL, loss runs), Sonnet 5 executes the correct action on existing systems [1].
- Data analysis and research: the model investigates live data, produces insights and conducts legal research [1].
- Bug investigation: one engineer reported that Sonnet 5, unprompted, wrote a reproducing test, implemented the fix, and then 'stashed' it to confirm the bug returned without the change — all in one go [1].
Analysis
For Dutch organisations looking to deploy agentic AI, Sonnet 5 is relevant because it offers an attractive balance between performance and cost. Until now, many agentic tasks were reserved for Opus models, which are two to three times more expensive. With Sonnet 5, these capabilities come within reach of smaller budgets.
The strong performance in coding and multi-step automation makes the model suitable for software teams wanting to experiment with AI agents, for example for pull request handling, refactoring or automated debugging. For analysis and research workflows too, where switching between tools and systems is often currently manual, Sonnet 5 can bring continuity.
The safety improvements and built-in cyber-safeguards are important for organisations wanting to run models in production without extensive proprietary oversight. The fact that Sonnet 5 is stronger at refusing malicious requests and better protected against prompt injection makes it more accessible for applications where user input is unpredictable.
Finally, Sonnet 5 offers a middle ground for Dutch professionals who want the flexibility of agentic models but don't yet have a use case for the more expensive Opus. Availability via all Claude subscriptions and API enables rapid pilots.
Sources
This article was produced with AI assistance from the original sources cited and checked against our editorial policy. Editorial policy