Sonnet 5 launches at an introductory $2 per million input tokens, and the week in Claude

PLUS: Fable 5 is back as of Jul 1, Claude Science launches in beta, and Claude Desktop comes to Linux

By SamJul 5, 2026

Morning, everyone. Here's what changed in Claude from Jun 28 to Jul 4.

This week, Anthropic launched Sonnet 5. Fable 5 came back after US export controls were lifted. And scientists got their own Claude app.

Plus 12 quick hits. Every claim has a link.

Sonnet 5 is out, and Anthropic says it's close to Opus 4.8

Anthropic has a new Sonnet model. It's now the default on the Free and Pro plans.

On Jun 30, Anthropic launched Claude Sonnet 5. It says the model's performance "is close to that of Opus 4.8, but at lower prices." Per Anthropic, its higher-effort performance can match Opus 4.8 on some tasks.

Max, Team and Enterprise users can pick it too. Per the release notes, it launched at introductory pricing of $2 per million input tokens and $10 per million output. It has a 1M token context window.

The release notes say Sonnet 5 uses a new tokenizer. It makes about 30% more tokens for the same text. The exact increase depends on the content and workload.

On Anthropic's automated behavioral audit, Sonnet 5 shows an overall lower rate of misaligned behavior than Sonnet 4.6. Its rate is somewhat higher than Opus 4.8 and Mythos Preview. Sonnet 5 launched with cyber safeguards on by default.

Fable 5 is back. The US dropped its export controls.

Anthropic brought Fable 5 back on Jul 1, with a new safeguard.

On Jun 12, the US government applied export controls to Fable 5 and Mythos 5. That meant restricting access for foreign nationals, inside or outside the US.

The order took effect immediately. Anthropic says it had no reliable way to verify nationality in real time. So it turned both models off for all users.

On Jun 30, Anthropic said the controls were lifted. Fable 5 came back on Jul 1 for users globally.

That covers the Claude Platform, Claude.ai, Claude Code and Claude Cowork. Anthropic said it would re-enable AWS, Google Cloud and Microsoft Foundry "as quickly as possible."

On Pro, Max, Team and select Enterprise plans, Fable 5 is included for up to 50% of weekly usage limits through Jul 7. After that, it runs on usage credits.

Per Anthropic, the controls came after a report by Amazon researchers. They had found a way past Fable 5's safeguards. Anthropic says its testing found many less capable models could find the same vulnerabilities.

Anthropic trained an improved classifier, working closely with the government. It says the reported technique is now blocked in over 99% of cases.

In a very small fraction of cases, the model may give information that isn't detailed enough to help an attacker. Blocked requests go to Opus 4.8, and users are told.

Anthropic says the new classifier also flags harmless requests more often during routine coding and debugging.

Anthropic wants one scale to rate jailbreaks

Anthropic wants a shared way to rate how serious a jailbreak is.

On Jul 2, Anthropic shared a draft framework. It calls it the Cyber Jailbreak Severity (CJS) scale.

The levels run from None (CJS-0) to Critical (CJS-4). Anthropic says the levels are meant to grow fast: each step up is several times more serious than the last.

A jailbreak is scored on four things:

How far it takes an attacker beyond existing tools.
How many attack tasks it works on.
How much effort it takes to turn into an attack.
How easy it is to get.

Anthropic says it's working on the framework with its Glasswing partners. It calls it an early draft. It also launched a HackerOne program where researchers can submit cyber jailbreaks they find in Fable 5.

The same post lists the cyber tasks Fable 5's classifiers aim to block and those they don't. Secure coding and log analysis are on the list they aren't meant to block.

Scientists get their own Claude app

Anthropic launched an app for scientists. It's in beta.

On Jun 30, Anthropic released Claude Science. It's for Pro, Max, Team and Enterprise users, on macOS and Linux. Team and Enterprise admins have to turn it on.

It runs on your laptop or on a remote machine over SSH. A main agent has over 60 curated skills and connectors for fields like genomics and proteomics. A reviewer agent checks citations and calculations.

Each figure comes with the exact code and environment that made it. Anthropic says large or sensitive datasets don't have to leave your systems. Only the context needed for each step goes to Claude.

Anthropic will back up to 50 Claude Science projects with up to $30,000 in credits. Applications are open through Jul 15.

Claude Desktop now runs on Linux, in beta

Linux users get the Claude desktop app, in beta.

On Jun 30, Anthropic said on X that Claude Desktop is now available on Linux in beta. It works on Ubuntu and Debian and includes Claude Code, Claude Cowork and chat. It's on all paid plans.

Per the setup docs, you need Ubuntu 22.04 or later, or Debian 12 or later, on x86_64 or arm64. Other Debian-based systems may work but aren't officially tested.

Computer Use and dictation aren't in the Linux beta yet.

You install it from Anthropic's apt repository. Updates then arrive through your regular package updates.

California gets Claude at half price

California's state agencies can now get Claude at 50% off per seat.

On Jun 29, Governor Gavin Newsom announced a partnership with Anthropic. State agencies can access Claude "at a 50% discounted price per seat." It comes with free workforce training and technical help.

The same offer covers cities and counties. Per the governor's office, Claude is the first AI productivity tool that will be available to all state agencies.

"AI should not replace the human work of government," Newsom said.

Claude Code now runs subagents in the background by default

Claude keeps working while most subagents run.

Claude Code 2.1.198 shipped on Jul 1. Subagents now run in the background by default.

Claude gets a notice when they finish. This was a gradual rollout before.

Per the week 27 notes, Claude still runs a subagent in the foreground when it needs the result before continuing.

Background agents started from claude agents now commit and push when they finish code work in a worktree. They also open a draft PR. Before, they stopped to ask.

The built-in Explore agent now uses the main session's model, capped at Opus. It used to run on Haiku.

Per Claude Code's week 27 notes, its Chrome integration is now out of preview for everyone on a direct Anthropic plan.

Quick hits

Sonnet 5 is in GitHub Copilot for Pro, Pro+, Max, Business and Enterprise users (Jun 30). Rollout will be gradual. Business and Enterprise admins turn it on in model policy settings.
Some old API settings return errors on Sonnet 5 (Jun 30). Per the release notes, non-default temperature, top_p or top_k values return a 400 error. Priority Tier isn't available on it.
Claude in Microsoft Foundry is generally available, hosted on Azure (Jun 29). To start, Opus 4.8 and Haiku 4.5 are available. You can pick where inference runs, including a US data zone.
The Claude apps gateway lets companies run Claude Code on Amazon Bedrock and Google Cloud with single sign-on and per-user cost tracking (Jun 29). It's self-hosted. It can set spend limits per organization, group or user.
Claude Enterprise admins get spend alerts at 75% and 90% of an org spend limit (Jul 2). Users get in-app notices at 75% and 95%.
Enterprise admins can now limit which models and effort settings users get (beta, Jul 1), per the support release notes.
Fast mode for Opus 4.6 is gone (Jun 29). Fast requests now run at standard speed and standard rates. They don't return an error.
Claude Managed Agents sessions can override the agent's setup (Jun 30). You can swap the model, system prompt, tools, MCP servers or skills for one session. The agent itself is unchanged.
Admins can set an org default model for Claude Code in the org console (2.1.196, Jun 29). It shows as "Org default" in /model when you haven't picked one.
claude mcp list and get no longer start MCP servers a repo approved for itself in a committed settings file (2.1.196, Jun 29). Untrusted workspaces show "Pending approval" instead.
Claude Code now loads up to 5 stacked slash skills, not only the first (2.1.199, Jul 2). Server rate-limit errors unrelated to your usage limit now retry for subscribers.
Claude Code's "default" permission mode is now called "Manual" (2.1.200, Jul 3). AskUserQuestion dialogs no longer move on by themselves by default.

Worth your time

Loop engineering: getting started with loops (Jun 30): the Claude Code team on turn-based, goal-based, time-based and proactive loops.
Claude Code week 27 (Jun 29 to Jul 3): a one-page roundup of versions 2.1.195 to 2.1.201.

That's the week. The next issue lands Sunday, Jul 12.

Reply and tell me what you're building. I read every reply.

— Sam