Claude Sonnet 5: Near-Opus Agentic Performance at $2/$10, and How to Choose Between Sonnet and Opus

Anthropic released Claude Sonnet 5 on June 30, 2026, and positioned it plainly: not a new frontier flagship, but a mid-tier model that brings much of Opus-class agentic capability to a lower price. This article covers what Anthropic says the model does, how pricing works, how it compares with Sonnet 4.6 and Opus 4.8, and how teams can decide which model to use.

What Anthropic is claiming

In its launch announcement, Anthropic calls Sonnet 5 “the most agentic Sonnet model yet”: it can make plans, use tools such as browsers and terminals, and run autonomously at a level that previously required larger, more expensive models. The company’s framing of the gap is specific: “Sonnet 5 narrows the gap: its performance is close to that of Opus 4.8, but at lower prices.” It describes the model as a substantial improvement over Sonnet 4.6 in reasoning, tool use, coding and knowledge work.

The announcement also includes cost-performance curves at different effort levels. According to Anthropic, Sonnet 5 is a strict improvement over Sonnet 4.6 across those curves, offers substantially better cost efficiency at medium effort, and at higher effort can match Opus 4.8 on some tasks. That last qualifier matters: “close to” and “on some tasks” are not the same as parity.

Why the Sonnet tier matters

Anthropic notes that many developers first used agentic AI through Sonnet-class models, from Claude Sonnet 3.5 through 3.7, but that recent gains in agentic capability had been concentrated in the more expensive Opus line. Sonnet 5 is the attempt to bring those gains back to the tier most production workloads actually run on.

Pricing

At launch, Sonnet 5 was priced at $2 per million input tokens and $10 per million output tokens, described as introductory pricing. Anthropic has since stated that this pricing was made permanent, rather than reverting to the standard Sonnet price of $3/$15 (announcement; pricing docs). For comparison, Opus 4.8 is priced at $5 per million input tokens and $25 per million output tokens.

On a simple per-token basis, that makes Sonnet 5 60% cheaper than Opus 4.8 on both input and output. Real costs depend on how many tokens each model uses for a task, which is why Anthropic presents cost-performance curves rather than price alone.

Availability

From launch, Sonnet 5 became the default model on Claude’s Free and Pro plans, and is available on Max, Team and Enterprise plans. Developers access it via the API as claude-sonnet-5; model details are on the Sonnet 5 overview page.

Safety evaluation

Anthropic reports that its pre-deployment evaluations found an overall lower rate of undesirable behaviours than Sonnet 4.6, better refusal of malicious requests, better resistance to prompt-injection hijacking, and lower rates of hallucination and sycophancy. It also says Sonnet 5 has much lower ability to perform cybersecurity tasks than current Opus models. These are the vendor’s own findings; the system card gives the details.

Choosing between Sonnet 5 and Opus 4.8

A practical way to decide:

• Start with Sonnet 5 at medium effort for high-volume work such as coding assistance, customer-support agents, data analysis and document processing. This is where Anthropic claims the biggest cost-efficiency gain.
• Raise effort before switching models. Because Sonnet 5 can approach Opus 4.8 at higher effort on some tasks, adjusting effort is often cheaper than moving to the larger model.
• Use Opus 4.8 for the hardest tasks, such as long, ambiguous research or complex multi-system engineering where reliability matters more than cost.
• Measure on your own tasks. Run a sample of real workloads through both models and compare success rate and total tokens used, not just per-token price.

What “effort” means in practice

Anthropic’s comparisons are presented by effort level, so it helps to know what that setting does. In the Claude API, the effort parameter lets developers trade thoroughness for speed and token use: lower effort produces faster, cheaper responses with less extended reasoning, while higher effort lets the model spend more tokens thinking and checking its work. The same model can therefore sit at several points on a cost-performance curve. That is why Anthropic argues that, for many workloads, the real choice is not simply “Sonnet or Opus” but “which model at which effort level”.

Migration checklist

Teams moving production traffic from Sonnet 4.6 or Opus to Sonnet 5 can reduce risk by:

1. Building an evaluation set of 50 to 200 real tasks with known good outcomes before switching.
2. Comparing total cost per completed task, including retries, rather than list price per token.
3. Re-checking prompts and tool definitions. A more agentic model may take more actions on its own, so confirm that tool permissions and confirmation steps still match your risk tolerance.
4. Rolling out gradually, for example routing a small share of traffic first and monitoring error rates and user feedback.
5. Keeping a fallback to the previous model until results are stable.

How it fits the market

Mid-tier models are where most enterprise volume sits, so competition there is mainly on price-performance. Sonnet 5’s pitch is that teams can get near-flagship agentic behaviour without flagship prices. Whether that holds up will depend on independent evaluations and on how it performs on each organisation’s own workloads.

Sources

• Anthropic: Introducing Claude Sonnet 5
• Claude API pricing
• Claude Sonnet 5 model overview