Claude Sonnet 5 Is Now the Default - and the 'Just Pay for Opus' Advice Is Dead

For most of the past year, the standard advice for anyone serious about AI-assisted coding went like this: the mid-tier models are fine for autocomplete, but if you're running agents, pay for the big one. Anthropic just undercut its own version of that advice.

Claude Sonnet 5 shipped on June 30 and immediately became the default model for Free and Pro plans and in Claude Code (Anthropic). Anthropic's own framing is unusually direct about what it's for: "the most agentic Sonnet model yet," able to "make plans, use tools like browsers and terminals, and run autonomously at a level that, just a few months ago, required larger and more expensive models."

Translated out of launch-speak: the work you were paying Opus prices for in early 2026 now mostly runs on the cheaper tier.

The numbers

Anthropic says Sonnet 5's performance "is close to that of Opus 4.8, but at lower prices," and that on cost-performance curves it "in some cases matches Opus 4.8's capability levels" (Anthropic). TechCrunch's launch coverage framed it more bluntly: a cheaper way to run agents (TechCrunch).

Note the asymmetry in that pricing table. At intro rates, Sonnet 5 costs 60% less than Opus on output tokens. Even at the September rates it's 40% less. Agent workloads are output-heavy - every plan, every tool call, every retry is output - so the practical savings sit near the top of that range for exactly the workloads where model spend hurts most.

Who should actually switch

Solo developers and small teams on Claude Pro: you already switched. Sonnet 5 is the default now; you got the upgrade without opting in. The only decision left is whether anything you do still needs manual escalation to Opus.

Teams running Claude through the API for agent pipelines: this is the real money decision. If your pipeline standardized on Opus because Sonnet 4.6 couldn't hold long agentic sessions together, that assumption is stale. Run your eval suite against Sonnet 5 before your next budget cycle. A team pushing 500M output tokens a month through Opus pays $12,500 for output alone; the same volume on Sonnet 5 costs $5,000 now and $7,500 after August. If quality holds on your tasks, that's not a rounding error.

People choosing between Claude and GPT-5.6: the launch timing here is transparently competitive. Nine days after Sonnet 5 landed at $2/$10, OpenAI shipped GPT-5.6's Terra tier at $2.50/$15. Through August, Sonnet 5 is the cheaper of the two mid-tier options; from September they're nearly identical in price and the decision is about fit - Claude's agentic tooling and Claude Code integration versus OpenAI's Codex ecosystem. Our Claude Code vs Cursor vs Copilot comparison walks through that seat-level choice.

Where Opus still earns its keep

A default change is not a replacement, and it would be dishonest to write this as "Opus is obsolete." Anthropic's own language - "close to," "in some cases matches" - concedes a gap at the top. Work that genuinely benefits from the strongest available reasoning still exists: architectural reviews across a large unfamiliar codebase, subtle concurrency debugging, high-stakes judgment calls where a wrong answer costs more than the tokens ever will.

The practical rule most teams should adopt: Sonnet 5 as the workhorse default, Opus reserved for escalation. That routing pattern used to require discipline to save money. Now it's how Anthropic ships the product.

The August 31 cliff

Don't let the fine print bury this: the $2/$10 pricing is promotional and expires August 31, 2026. After that, input rises 50% and output rises 50% to $3/$15.

If you're doing procurement math right now, or writing a cost model into a proposal, use the September numbers. Intro pricing is real money saved this summer, but any decision with a horizon past Labor Day should be justified at $3/$15 - if the switch only makes sense at promo rates, it doesn't make sense.

What to do now

• On Claude Pro or Claude Code: nothing. You're already on it. Escalate to Opus manually when a task actually warrants it.
• On the API with Opus as your agent default: run a two-week A/B on your real workload before August 31, while the trial itself is cheapest.
• Comparing against OpenAI: decide on September pricing ($3/$15 vs Terra's $2.50/$15), not the promo.
• Track your escalation rate. If under 10% of your tasks genuinely need Opus, your default was wrong even before Sonnet 5 - now it's expensive and wrong.

Current plan details live on our Claude pricing page, which we've re-verified as of this post's date.