The first week of July opened with a change that quietly reset the default model on Free and Pro accounts, so let’s start there.
Sonnet 5 becomes the default model
On June 30, Anthropic shipped Sonnet 5 and set it as the default model on Free and Pro plans right away (Max, Team, and Enterprise all get access too, you just have to pick it yourself). It’s the most agentic Sonnet yet, which in practice means it plans multi-step work better, handles tools like a browser or terminal more reliably, and loses the thread less often during longer, autonomous runs – the kind of thing that not long ago needed a bigger, pricier model.
Launch pricing is $2 per million input tokens and $10 per million output tokens, running through August 31, after which it reverts to the standard $3/$15. TechCrunch called it “a cheaper way to run agents,” and that’s fair, since this level of autonomy used to cost noticeably more. If you’re on the Free or Pro plan, you’re probably already talking to Sonnet 5, even if nobody announced it to you with fanfare inside the app. And if you don’t have an account yet, we walk through setting one up in a separate lesson: A Claude account from scratch.
Which model is actually answering in your session is something /model will tell you – worth a look especially if you’ve gotten into the habit of forcing Opus for every single task. For routine work, refactoring, writing tests, small fixes, Sonnet 5 is more than enough, and at the current promotional pricing the difference on your bill can be significant. Keep Opus for the moments that genuinely need the strongest reasoning available, rather than leaving it on as the default for everything.
Claude Science walks into the lab
The same week, Anthropic announced Claude Science, a version tuned for scientific labs, with a particular eye on pharma research. It’s more than a chat interface bolted onto a lab bench: Claude Science is meant to help design experiments, work through sequencing data, and comb through scientific literature at a pace a single research team simply couldn’t match on its own.
Most readers of this blog won’t touch it directly. It is, though, an early sign that Anthropic might be moving toward industry-specific variants of Claude instead of sticking to one general-purpose product for everyone. If you work with a client in healthcare or pharma, this is a good moment to ask whether they’ve already heard about it, before they hear it from someone else first.
What’s new in Claude Code
This week’s Claude Code changelog brought one setting that saves tokens and one warning that saves your nerves.
A warning before your login expires
Claude Code 2.1.202 added a “Dynamic workflow size” setting to control the suggested agent count in dynamic workflows, so you get more say over whether Claude Code splits a task across two agents or ten. If you’ve noticed dynamic workflows spinning up more parallel agents than a task actually needed, burning tokens without a clear payoff, this setting finally gives you a lever, instead of leaving it entirely to the model’s judgment.
A day later, in 2.1.203, Claude Code started warning you before your login expires. That sounds minor until you’ve lost a session halfway through a long task because your token just quietly invalidated and you had no idea it was coming. Now you get a heads-up before it happens, so you can log back in without losing your working context.
/doctor trims an overgrown CLAUDE.md
A more interesting change landed in 2.1.206, on July 9: /doctor learned to propose trimming an overgrown, checked-in CLAUDE.md when you have one in your repo. That’s a real problem anyone who’s worked with Claude Code for a while will recognize – the instructions file grows with every session, every new rule, every exception to a rule, until it turns into a wall of text the model has to read on every single run instead of actually using. We run a similar audit on our own instruction files regularly, so we’d recommend running /doctor this week rather than waiting until the file’s grown enough that trimming it becomes a project of its own.
/commit-push-pr now also auto-approves git push to your repo’s configured remote, one less click on every commit. Small changes, but this is exactly the kind of detail that separates a tool you have to babysit from one that just gets on with it.
To end on: an AI that listens to the rainforest
On July 7, WILDLABS announced its 2026 awards: $360,000 spread across 16 conservation-technology projects. The one that stuck with us most was Shuar women in the Amazon working with researchers to build open acoustic classifiers grounded in their own local knowledge of the forest, supporting both the research itself and the rainforest restoration their community is already leading.
There’s also a project monitoring endangered Humboldt penguins in Peru entirely by sound, with no need to catch or disturb the animals at all. Bioacoustics turns up often enough across this year’s list to be hard to miss: cheap acoustic sensors, AI, and local knowledge together do what would have needed expensive equipment and months of fieldwork just a few years ago. It’s worth keeping this kind of use case in mind next time someone claims AI is mostly about generating content and replacing jobs: sometimes it’s just a cheaper way to hear something nobody could hear before.