Claude Code's Default Model Just Changed: What Sonnet 5 Actually Gets You

I opened Claude Code Monday morning and the model in the corner of my terminal said something I didn't recognize: Sonnet 5. I hadn't touched a config file. Anthropic just swapped the default out from under me, mid-week, no announcement banner, no opt-in.

That's worth pausing on before we get to the specs. Claude Sonnet 5 shipped June 30, 2026, and it's now the default model in Claude Code and the default for Free and Pro plans on claude.ai. Max, Team, and Enterprise users can still pick Opus manually, but the out-of-the-box experience for almost everyone building with Claude changed overnight. If you run cohorts, coach a team, or just have a business partner who "uses Claude sometimes," this is the version of Claude they're on right now whether they noticed or not.

What actually shipped

Three things, and only one of them is the headline model swap.

It's the new default, everywhere. Free and Pro plans on claude.ai default to it automatically. Claude Code defaults to it. If you've been telling clients "we build on Claude," this is now the "Claude" you mean.

The context window is native at 1M tokens, no toggle required. This is the part I actually care about. Sonnet 4.6 could reach a 1M-token window too, but on most plans you had to switch on usage-based credits to get there. It wasn't part of the default upgrade path. With Sonnet 5, 1M tokens is simply the standard configuration: no beta header, no separate toggle, billed at the same standard rate as everything else.

Pricing is aggressive. $2 per million input tokens, $10 per million output, through August 31, then $3/$15 after that. Compare that to Opus 4.8's standard $5/$25. Even post-promo, Sonnet 5 runs at roughly 60% of Opus's per-token cost. During the promo window it's 40%.

Why "native default" beats "available in beta"

I've watched enough cohort attendees use Claude Code to know that almost nobody flips a beta flag. Not because they can't. Because a feature gated behind a header or a credits toggle might as well not exist for someone who just wants to point Claude at a repo and get to work. Defaults are what actually get used. Options are what get read about in a changelog once and forgotten.

That's the real story here, more than the token count. The thing that used to require you to know it existed, find the setting, and turn it on is now just what happens when you open Claude Code. If you've been running into "please continue by reading the file, it's long" style context walls on a bigger codebase or a long research thread, that friction mostly went away this week, for free, without you doing anything.

The benchmarks, and the one that surprised me

Sonnet 5 doesn't beat Opus 4.8 across the board, and Anthropic isn't claiming it does. On SWE-bench Verified, a benchmark of real GitHub bug fixes, Sonnet 5 scores 72.7% against Opus 4.8's 79.4%. Opus is still the harder-problems model.

But on Terminal-Bench 2.1, which measures exactly the kind of long, multi-step, tool-using work Claude Code does all day (running commands, chaining scripts, recovering from a failed step), Sonnet 5 scores 80.4%. Opus 4.8 scores 74.6%. Sonnet 4.6 scored 67.0%.

That's the number I keep coming back to. A cheaper, faster model just out-scored the flagship on the specific skill Claude Code sessions live or die by: running a long sequence of terminal steps and staying correct across all of them. It doesn't mean Opus is obsolete. It means the everyday default got meaningfully better at the everyday job, not just cheaper at it.

What I'm actually doing differently this week

Two changes, both small, both real:

I stopped reflexively switching to Opus for normal build work. In the cohorts I've been telling people "switch to Opus for anything that matters." That advice is now wrong for a big chunk of Claude Code work. I'm reaching for Opus specifically when a task is architecturally gnarly (the SWE-bench-style hard bug, the ambiguous refactor), and letting Sonnet 5 run the actual building, which is most of the day.

I stopped manually managing context so carefully. With a real 1M-token window as the standing default, I've been less precious about /compact-ing mid-session on longer builds. That's a small workflow change, but it's the kind of small friction that used to eat five minutes a dozen times a day.

None of this is hype. It's a pricing and defaults change that happens to line up with a real capability jump on the exact task most of you are using Claude Code for. If you haven't opened your terminal since June 30, go look at what model it says you're on. You might already be running it.

Sources: Introducing Claude Sonnet 5 · What's new in Claude Sonnet 5 · Introducing Claude Opus 4.8 · Sonnet 5 vs Sonnet 4.6 vs Opus 4.8 benchmarks