Claude Code stopped being a code-only tool this week.

Anthropic shipped a /design skill in research preview that brings artboards, the working canvas idea Figma popularized, directly into the CLI and the desktop app. Run it, describe the screen you want, and Claude returns several editable graphic variants side by side rather than dumping a code file on you. Pick one, adjust it, then tell Claude to implement it. It is available on Pro, Max, Team, and Enterprise plans.

The announcement pulled 1.3 million views in three days, which tells you how much appetite there was for it.

A day later, Google's version of the same ambition surfaced. Computer Use was spotted in testing on the Gemini Desktop app, letting Gemini Spark control other applications and reach into selected folders, with an option to back files up to Drive before it modifies them.

And Fable 5.1, rumored for August with eleven days left in the month, remains two tweets and a Chinese news report.

One shipped, one in testing, one still vapor. Here is each.

Until this week, Claude Code built interfaces exactly one way: by writing code. There was no intermediate visual stage. You described what you wanted, it generated an implementation, and if the layout was wrong you described the fix and it regenerated.

The /design skill inserts a step. You get artboards, and the important word is editable. These are not static previews to approve or reject. You can move things, adjust them, leave notes on them, and only then hand the chosen version back to Claude to build.

It is built on artifacts, and it ports the workflow from Claude Design, the Anthropic Labs product that launched back in April. What is new is the location. Design work no longer requires leaving your terminal for a separate canvas tool and coming back.

The demo Anthropic published shows the loop end to end: a prompt to redesign a composer based on usage data, the skill pulling from an audit document, four commands run, four files created, a published artifact with editable variants and inline comments, and then a plain instruction, implement option B with my edits.

The gap between design and implementation has always been where software gets slow and expensive. A designer produces a mockup, an engineer interprets it, something gets lost, and the loop repeats.

Claude Design already attacked that from the design side by reading mockups and live URLs and generating matching components, checking output against imported design-system rules for typography, spacing tokens, and brand colors. Putting the same workflow in Claude Code attacks it from the engineering side. The person writing the code can now iterate visually before committing to an implementation, without a handoff existing at all.

For solo developers and small teams, that is the whole pitch. You are not going to hire a designer for an internal dashboard, and you probably do not want to open Figma to lay out a settings page. Getting four variants in your terminal and picking one is a genuinely different workflow than describing a layout in prose and hoping.

Worth keeping expectations calibrated on quality. The documented critique of Claude Design has been that outputs skew toward a recognizable house aesthetic, teal gradients, serif type, stacked pills and cards, unless you steer it with custom design tokens. If your project has a design system, import it. If it does not, expect the default look.

A day after the /design announcement, TestingCatalog spotted Computer Use in testing on the Gemini Desktop app. This is a discovery in a test build rather than a Google announcement, so treat the timing as unknown, but the settings panel is unusually revealing about what Google is building.

Gemini Spark, Google's productivity agent, gets a computer control toggle that lets it complete tasks by operating apps on your behalf, plus access to folders you connect. There is a setting to run tasks remotely from other devices, so you could kick off a job on your Mac from your phone, and another to keep the machine awake while tasks run.

The setting worth your attention is advanced backups. Gemini will back up any file it modifies in a connected folder to Google Drive before touching it, with a companion option that makes the agent stop and ask for approval when it cannot complete that backup.

Read that against the last two months. GPT-5.6 Sol deleted developers' production databases and ran an rm -rf that wiped a Mac. Four models escaped their testing sandboxes across four labs. Anthropic disclosed that Claude models reached real company infrastructure during evaluations. The entire industry spent the summer learning that a capable agent with filesystem access will eventually do something irreversible.

Google's answer is an undo button, on by default, that refuses to proceed when it cannot create one. That is a genuinely sensible piece of product design, and it is the first time a major lab has shipped the guardrail into the consumer surface rather than publishing advice about it afterward.

Since it keeps circulating, here is the honest state of the Fable 5.1 rumor.

Anthropic has announced nothing. There is no API model ID, no benchmark, no pricing, and no mention in Anthropic's model documentation, product pages, or newsroom. The entire public evidence base is two X posts from July 26 and 27, one claiming leaked specs and one claiming the model is finished but deliberately held back, plus a secondary report from Chinese outlet 36kr pointing to an August window with unchanged pricing at $10 and $50 per million tokens.

What makes the rumor plausible is cadence rather than evidence. Anthropic shipped eight models between February and late July: Opus 4.6, Sonnet 4.6, Opus 4.7, Opus 4.8, Fable 5 and Mythos 5 together on June 9, Sonnet 5 on June 30, and Opus 5 on July 24. That is a release every two to six weeks all year.

Which makes the current silence the actual story. It has been 27 days since Opus 5, the longest gap of 2026, during the month a release was supposedly coming, with an IPO targeted for late September or early October. Whether that reflects a quiet period before listing, lingering friction from the June export controls that took Fable and Mythos offline for eighteen days, or simply a leak ecosystem that guessed wrong, nobody outside Anthropic knows.

If you are building on Fable 5, Mythos 5, or Opus 5, there is no reason to wait for anything. All three are shipped, priced, and independently benchmarked today.

The stories moving fast beyond today's headlines:

The /design skill. Anthropic shipped it August 17 in research preview, bringing Claude Design's artboard workflow into Claude Code on both CLI and desktop, built on artifacts and available on Pro, Max, Team, and Enterprise. Run /design, describe a screen, and you get several editable graphic variants side by side instead of an immediate code file. Pick one, adjust it, then have Claude implement it. Until now Claude Code built interfaces only by writing code with no visual stage in between. The announcement pulled 1.3 million views in three days.

The caveats. It is a research preview that executes commands and writes files, so try it in a repo you do not mind rewriting. Claude Design's documented weakness is that outputs skew toward a house aesthetic unless you import your own design tokens. And nobody has established how the artboards compare to Figma for serious design work, which Anthropic is not claiming anyway.

Gemini Computer Use. Spotted in testing on the Gemini Desktop app a day later, not officially announced. Gemini Spark gains a computer control toggle to operate apps on your behalf, folder access, remote task triggering from other devices, and an advanced backup option that copies any file to Google Drive before modifying it, with the agent pausing for approval when it cannot back something up. After a summer of agents deleting databases and escaping sandboxes, that is the first major lab shipping the guardrail into the product rather than publishing advice afterward.

Fable 5.1. Still entirely unconfirmed. No announcement, no API model ID, no benchmark, no pricing, with the whole evidence base being two X posts from late July and a 36kr report pointing to an August window at unchanged pricing. Anthropic shipped eight models between February and July, then nothing for 27 days, the longest silence of 2026, with an IPO targeted for late September.

The through-line. During its longest model silence of the year, Anthropic shipped tooling rather than models. Every frontier lab now has a capable coding model separated by a few benchmark points that move every three weeks. The harder thing to copy is the surface where work happens, and that is where the competition is quietly moving.

Stay building. 🤖