Artificial intelligence · DevelopmentJune 30, 20263 min · updated August 12, 2026
Anthropic launches Claude Sonnet 5, its most agentic Sonnet model
Rephrased by Daillac
Source: Anthropic ↗
In brief
- Sonnet 5 plans, uses tools and carries out longer tasks with greater autonomy.
- It is available in Claude, Claude Code and the Anthropic API.
- Anthropic reports stronger safety behaviour than Sonnet 4.6 while retaining cyber safeguards.
A Sonnet model designed for execution
Anthropic positions Claude Sonnet 5 near its Opus models on several agentic tasks at a lower cost. Improvements focus on reasoning, tool use, coding and sustained multi-step work.
The model is the default for Free and Pro plans and remains available on Max, Team and Enterprise. Developers can use it in Claude Code and through the API as
claude-sonnet-5, simplifying trials in existing workflows.US$2 / M
input tokens at the announced introductory price through August 31, 2026; output is listed at US$10 per million.
Introductory pricing · Source: Anthropic
Measure total cost, not token price alone
A cheaper model can cost more if it adds steps or supervision. Conversely, completing and checking a task more reliably may reduce operating cost. Before migrating, compare quality, duration, token use and human intervention on real scenarios, and plan a resilient multi-model AI architecture.
Safety results and migration details
Anthropic says Sonnet 5 rejects some harmful requests more effectively and better resists prompt injection than Sonnet 4.6. It also reports fewer hallucinations and less sycophancy in internal evaluations, while acknowledging that unwanted behaviours remain more common than in its strongest models.
A safer model does not replace narrow permissions, tool isolation or code review. Teams should also account for the updated tokenizer: Anthropic estimates that the same input may use roughly 1 to 1.35 times as many tokens depending on content.
- Replay a representative sample of real tasks.
- Compare first-pass acceptance and retry rates.
- Measure input, output and tool calls separately.
- Test rate limits and provider outages.
Move from demonstration to production
A technical capability becomes reliable only after permissions, evaluations, cost limits, and failure behavior are defined. Test actions, exposed data, and human recovery—not just answer quality.
InfographicProduction guardrails · shareable block
01
Authorize
Least privilege, approved tools, and a clear boundary between reading, proposing, and executing.
02
Evaluate
Versioned test sets, adversarial cases, latency budgets, and quality thresholds.
03
Recover
Action logs, human interruption, failure handling, and a practiced rollback.
The decisive criterion is operational
The right option is not necessarily the strongest model in a demo. It is the combination that meets data, cost, latency, availability, and control requirements while preserving the ability to switch or disable a function.
Editorial follow-up questions
Should an agent receive normal user privileges?+
Not by default. Rights should be limited by task and duration, with a distinct identity and log.
What should be tested beyond answers?+
Tool calls, refusals, sensitive data, cost overruns, latency, and partial failures.
When is human approval required?+
Before irreversible actions, external communication, financial commitments, or critical data changes.
How can vendor dependence be reduced?+
Isolate model access, version evaluations, and test a fallback path.
Compare models on real business processes
Daillac can build an evaluation set around your data, tools and quality targets.
Sources & method
Summary of Anthropic’s announcement and safety documentation, interpreted from a product perspective.
Read the original source: Anthropic ↗News feed
August 3, 2026
Web development · Performance
Next.js 16.3 promises lower memory use and instant navigation
July 23, 2026
Governance · Artificial intelligence
Ottawa consults Canadians on transparency for AI systems and content
July 14, 2026
Artificial intelligence · Canada
Anthropic commits C$10 million to Canadian AI research