Claude Sonnet 5 Is Here. The Opus vs. Sonnet Trade-off Just Collapsed.

The Trade-off That Forced Your Hand Is Gone

Anthropic launched Claude Sonnet 5 on 30 June 2026, and it does something the previous Sonnet tier never quite managed: it performs close to Opus 4.8 in agentic tasks, at a substantially lower cost. That single fact rewrites the model-selection logic most teams have been running for the past year.

Until now, the calculus was blunt. You used Sonnet for speed and cost efficiency. You used Opus when the task genuinely needed frontier-level reasoning, sustained tool use, or complex coding. Sonnet 5 narrows that gap far enough that the default answer for many production workloads just changed.

What Happened and When

Anthropic announced and released Sonnet 5 on 30 June 2026. It is available immediately across all Claude plans, including Free and Pro, where it is now the default model. Max, Team, Enterprise, Claude Code, and API access are all live from the same date, as confirmed on the official Claude Sonnet 5 announcement.

Pricing at launch is introductory: $2 per million input tokens and $10 per million output tokens through 31 August 2026. From 1 September 2026, it moves to $3 per million input and $15 per million output. For context, Opus 4.8 sits at $5 per million input and $25 per million output.

How the Pricing Stacks Up

The table below shows the cost comparison directly from Anthropic's verified pricing. Use it to sense-check whether your current model routing still makes financial sense.

Source: Anthropic's Claude Sonnet 5 launch page, 30 June 2026. Introductory pricing applies through 31 August 2026.

How It Actually Works: Effort Levels, Not a Fixed Mode

The mechanism worth understanding is the effort dial. Sonnet 5 is not simply faster or slower than Opus. It operates across a range of configurable effort levels on the Claude API, meaning you can tune cost and capability per task rather than committing to a fixed model tier.

At medium effort, Sonnet 5 is already substantially more cost-efficient than Opus 4.8. At extra-high effort, it matches Opus 4.8 on some agentic tasks, including BrowseComp (agentic web search) and OSWorld-Verified (computer use). That flexibility is the real architectural change here.

What Changes for Agentic Workflows and Coding Teams

Anthropic's framing is deliberate: Sonnet 5 is positioned as the most agentic Sonnet model yet. Early access partners reported that it completes multi-step tasks where Sonnet 4.6 would stall, self-checks its own output without being prompted, and handles brownfield code, including race conditions and hidden test failures, without reverting to surface-level patches.

For growth and engineering teams running AI agents at scale, this matters in two concrete ways. First, if you are currently routing complex tasks to Opus purely for reliability, that routing logic deserves a re-test. Second, if you have been holding back on agentic deployments because Opus costs made them hard to justify at volume, the economics have shifted enough to revisit the business case. If you want broader context on how agentic AI is already operating inside businesses, the analysis of how OpenAI's own data shows agentic AI running business operations is worth reading alongside this.

The Safety Point That Often Gets Skipped

Anthropic's safety evaluations found Sonnet 5 shows a lower rate of undesirable behaviours than Sonnet 4.6 in agentic contexts. It also has a substantially reduced ability to assist with cybersecurity tasks compared to current Opus models. That is a deliberate capability constraint, not an oversight, and it matters if you are assessing model risk for enterprise deployments.

The Action to Take Before 31 August

The introductory window closes on 31 August 2026. If you are evaluating Sonnet 5 for production agent workloads, do it now while the cost floor is at its lowest. Run your most demanding agentic tasks, the ones you currently route to Opus, through Sonnet 5 at high effort and compare outputs directly. The pricing gap at launch makes this the lowest-risk window to benchmark. And if your team is also thinking about the broader infrastructure shift underneath models like this, the piece on what custom inference chips mean for AI cost and speed adds useful context on where the economics are heading across the industry.