On June 30, 2026, Anthropic launched Claude Sonnet 5, which the company calls the best Sonnet yet at agentic work. It plans on its own, uses a browser and a terminal, and runs autonomously at a level that, just a few months ago, required bigger and costlier models. According to Anthropic, Sonnet 5 comes close to Opus 4.8 at a much lower price. For developers and anyone who regularly hands work off to AI, this one is worth trying. Hung Phat breaks down what you need to know.
What Claude Sonnet 5 Is and How It Improves on Sonnet 4.6
Anthropic positions Sonnet 5 as the most clearly agentic release in the Sonnet line: it can plan, call tools such as a browser and a terminal, and carry out many steps from start to finish without hand-holding. The Sonnet line kicked off the agentic era with 3.5, 3.6, and 3.7. More recently, the biggest leaps have come from the Opus line, so a Sonnet release that narrows the gap with Opus 4.8 is something developers are watching closely.
Compared with its predecessor, Sonnet 4.6, Sonnet 5 is ahead on nearly every important agentic front: reasoning, tool use, coding, and knowledge work. The scores Anthropic published show the specific gains, with Opus 4.8 alongside for easy comparison.
A few numbers stand out. On Terminal-Bench 2.1 (command-line coding), Sonnet 5 scores 80.4% versus 67.0% for Sonnet 4.6, nearly touching Opus 4.8 at 82.7%. On SWE-bench Pro it is 63.2% versus 58.1%. On the OSWorld-Verified computer use test it reaches 81.2% versus 78.5%, right behind Opus 4.8 at 83.4%. Humanity's Last Exam with tools comes in at 57.4% versus 46.8%, almost level with Opus 4.8 (57.9%). On knowledge work, GDPval-AA v2, which uses a different scoring scale, Sonnet 5 earns 1618 points, above Opus 4.8's 1615.
The overall idea is that you get near-Opus quality at Sonnet pricing. You can read the full announcement in Anthropic's Claude Sonnet 5 launch post, along with the more detailed System Card.
Working With Sonnet 5: Planning, Tool Use, and Autonomous Runs
The biggest selling point of this release is its ability to work independently. Anthropic measured it on two familiar tests: BrowseComp for agentic search tasks, and OSWorld-Verified for computer operation. Sonnet 5 improves solidly over Sonnet 4.6, while Opus 4.8 still leads when you need the highest accuracy.
The model also lets you adjust its effort level, including a very high setting called xhigh. For the same task, you can lower the effort to save money, or raise it in exchange for accuracy. That lets you balance cost and quality project by project instead of being stuck at one fixed level.
Feedback from early-access teams is fairly consistent: Sonnet 5 follows through to the end on tasks that earlier Sonnet versions often left unfinished, and it checks its own results without being prompted. One partner gave it a two-part job, updating account tiers in Salesforce and then sending a launch announcement to enterprise customers, and Sonnet 5 completed it from start to finish. Another engineer asked it to investigate a bug, and the model wrote a reproduction test, fixed the bug, then removed the fix to confirm the bug came back, all in a single pass.
Other teams mentioned debugging messy legacy code (race conditions, hidden tests), legal research and analysis, and running insurance workflows through computer use. The common thread is finishing multi-step work at a comfortable cost, which is exactly what a real agent needs to do.
Safety Evaluation and Cybersecurity Capabilities of Claude Sonnet 5
On safety, Anthropic says Sonnet 5 is generally better than Sonnet 4.6. The model refuses malicious requests more reliably, resists injected instructions in prompt injection attacks better, and makes things up and flatters users less than the earlier version.
On the automated behavioral audit, which scans for many kinds of misaligned behavior, Sonnet 5 scores lower than 4.6, meaning it is safer. Even so, the score is still higher than for the two more powerful models, Opus 4.8 and Mythos Preview, and Anthropic is upfront about that limit.
The cybersecurity side needs a clear explanation. Anthropic did not train Sonnet 5 specifically for this area. On a test of building exploits for Firefox 147 browser vulnerabilities (run with Mozilla, with the flaws patched in Firefox 148), both Sonnet 5 and 4.6 scored 0.0% at producing a complete working exploit. Sonnet 5 edges ahead only slightly on partial success, most likely because of its higher general intelligence rather than any dedicated training. Its cyber capabilities are well below those of Opus 4.8 and Mythos 5.
Because it is slightly ahead of 4.6, Sonnet 5 ships with real-time cyber safeguards turned on, using the same mechanism as Opus 4.7 and 4.8, which blocks dangerous usage the moment it is detected. For security work that requires loosening the guardrails, Anthropic recommends choosing Opus 4.8 for that task.
Pricing, Plans, and What Laptop Users Should Care About
On pricing, Anthropic is offering Sonnet 5 at an introductory rate through August 31, 2026: 2 USD per 1 million input tokens and 10 USD per 1 million output tokens. After that date, pricing returns to the standard 3 USD and 15 USD. For comparison, Opus 4.8 is currently priced at 5 USD and 25 USD.
The chart makes it clear that Sonnet 5 is considerably cheaper than Opus 4.8 on both token types. One technical note: Sonnet 5 switches to a new tokenizer, so the same content can be counted as more tokens, roughly 1.0 to 1.35 times depending on the type of content. Anthropic set the introductory price so the transition adds almost no extra cost.
Sonnet 5 is available on every plan from launch day. It is the default model for Free and Pro, is available to Max, Team, and Enterprise users, runs in Claude Code and on the Claude Platform under the name claude-sonnet-5, and can also be used in Cowork. Anthropic has also raised usage limits to accommodate the higher token consumption when you push the effort level up.
What about the person sitting at the keyboard? Sonnet 5 runs in the cloud, so you do not need a powerful machine just to call it. But when you let it work agentically through Claude Code, opening a terminal, running tests, building projects, and handling multi-step operations, your own laptop determines whether everything runs smoothly or stalls. A machine with 16GB of RAM or more, a multi-core CPU, and a fast NVMe drive makes the write-and-run loop noticeably smoother. If you are thinking about buying a machine for working with AI and Claude Code, Hung Phat can recommend a configuration that fits your budget and your programming needs. Take a look at the high-spec laptops currently in stock.
Where is Claude Sonnet 5 stronger than Sonnet 4.6?
Sonnet 5 beats 4.6 in reasoning, tool use, coding, and knowledge work. For example, Terminal-Bench 2.1 is 80.4% versus 67.0%, OSWorld-Verified is 81.2% versus 78.5%, and Humanity's Last Exam with tools is 57.4% versus 46.8%.
How much does Claude Sonnet 5 cost?
The introductory price through August 31, 2026 is 2 USD for input and 10 USD for output per 1 million tokens, then it returns to the standard 3 USD and 15 USD. Opus 4.8 is more expensive, at 5 USD and 25 USD.
Do I need a high-spec laptop to use Claude Sonnet 5?
The model itself runs in the cloud, so you do not need a powerful machine to call it. But when running agentic work through Claude Code (terminal, tests, builds), a laptop with 16GB of RAM or more, a multi-core CPU, and a fast NVMe drive will make the work smoother.
Is Sonnet 5 safe for agentic work?
Anthropic rates Sonnet 5 as safer than 4.6: it refuses bad requests better, resists prompt injection better, and makes things up and flatters users less. Its cybersecurity capability is low, and the company has turned on real-time cyber safeguards by default.
