Claude Sonnet 5: nearly as good as Opus, at half the cost

What Claude Sonnet 5 is

The Sonnet line has historically been the one that brought agentic capabilities into the mainstream: Sonnet 3.5, 3.6, and 3.7 were the first Claude models to show real strength in coding and tool use. More recently, though, the clearest gains on the agentic side have shifted to the pricier Opus models.

With Sonnet 5, Anthropic says it has closed much of that gap: performance is officially described as "close to Opus 4.8" — the company's most capable model — but at a significantly lower price. It's a direct upgrade over Sonnet 4.6 (released in February 2026), with what Anthropic calls "substantial" gains in reasoning, tool use, coding, and knowledge work.

What it can do, in business terms

Finishes the job instead of stopping halfway. Anthropic cites a concrete example from internal testing: the model was given a two-part task — update account tiers in Salesforce, then send a launch announcement to enterprise contacts — and Sonnet 5 carried it through end to end, without stalling halfway the way earlier models often did. For your business, that means fewer workflows that "almost" finish and land back on someone's desk.

Checks its own work, unprompted. In another example cited officially, asked to investigate a bug, the model wrote a reproducing test on its own, implemented a fix, then confirmed the bug returned without the change — all in a single pass, with no step explicitly requested. That kind of discipline matters when an agent is working unsupervised.

Better agentic search and computer use. On internal evaluations for agentic search (BrowseComp) and interface operation (OSWorld-Verified), Sonnet 5 is a clear improvement over Sonnet 4.6. The exact figures are published only as charts on Anthropic's site, but the official conclusion is unambiguous: Sonnet 5 delivers far more quality than was previously available at this price point, while Opus 4.8 remains the choice for maximum accuracy when every percentage point counts.

Adjustable effort levels. You can choose how much thinking effort the model spends per task, balancing cost against precision — useful when running agents at volume, across tasks of varying importance.

One example from the operations side: an insurance company (Pace) is already running computer-use agents on Sonnet 5 for real workflows — intake, opening claim files (FNOL), checking loss runs — on the systems its teams already use. The feedback cited officially: the model "consistently takes the right action and does it quickly" — exactly what matters in real operations, not just in demos.

Translated for a typical business: financial reporting, customer follow-up, lead sorting and qualification, or workflows on internal systems can now run on agents that carry the work forward on their own, with human input reserved for the decisions that actually matter. If you want to see what such an operational agent looks like in practice, we cover the concept in our guide to operational AI agents.

What it costs

Sonnet 5 is available today at an introductory price — $2 per million input tokens and $10 per million output tokens — through August 31, 2026. After that date, pricing moves to the standard tier: $3 per million input and $15 per million output. For comparison, Opus 4.8 — Anthropic's most capable model — costs $5 per million input and $25 per million output.

One technical note worth budgeting for: Sonnet 5 uses a new tokenizer (similar to the change introduced with Opus 4.7), which can produce 0–35% more tokens for the same input text, depending on content type. Anthropic says it calibrated the introductory price specifically to make the transition from Sonnet 4.6 roughly "cost-neutral" per task — meaning that while the price per token is lower, the token count may rise slightly, and the final cost per task stays, on average, comparable or lower.

Availability: Sonnet 5 is live today across all Claude plans — it's the default model on Free and Pro, also available on Max, Team, and Enterprise, in Claude Code, and via the API (model id claude-sonnet-5) for anyone building custom agents, like we do.

What you should know about safety

Alongside the launch, Anthropic published its safety evaluations and the model's full system card. In short: Sonnet 5 shows a lower overall rate of undesirable behaviors than Sonnet 4.6 — fewer hallucinations, less sycophancy (the tendency to agree with a user just to please them), and better resistance to prompt-injection attacks in agentic contexts.

On the other hand, the model's capability on offensive cybersecurity tasks (such as developing exploits) is substantially weaker than current Opus models — Anthropic explicitly states it did not deliberately train Sonnet 5 on such tasks and recommends Opus 4.8 for security work that requires fewer guardrails. The model still ships with cyber safeguards enabled by default, identical to those on Opus 4.7/4.8.

For a business looking to put agents on internal systems, the practical takeaway is simple: Sonnet 5 is a solid foundation for everyday operational automation — reporting, support, follow-up, document processing. For dedicated offensive security work or red-teaming, the officially recommended option remains Opus.

Why it matters for business

The combination of lower price, a real agentic leap, and an improved safety profile changes the math on what a business can realistically automate without a large technical team. Workflows that used to require either step-by-step supervision or an expensive Opus-class model just to "keep up" with the task can now run on Sonnet 5 at a fraction of the cost — 40% less at standard pricing, up to 60% less through August 31, 2026.

In practice, the equation becomes: more agents, running longer, on smaller budgets. That's exactly the combination that matters when the goal isn't a showcase chatbot, but operational agents doing real work — day in, day out, without constant supervision.

Sources: ↗ Official Anthropic announcement · ↗ Claude Sonnet 5 System Card