48 Hours In April April 14 Krieger resigns Figma board April 16 Opus 4.7 ships April 17 Claude Design ships, Figma -7%

7% Figma stock drop

20 to 2 Prompts per page

1M Token context

The loop that just closed

Mike Krieger resigned from Figma's board on April 14. Claude Design launched on April 17. You do not need to be Sherlock. Anthropic shipped two things in 48 hours. Opus 4.7 on Thursday. Claude Design, powered by 4.7, on Friday. The headline version is "new model, new Figma competitor, Figma stock down 7 percent." That framing misses the point. Here's the part nobody's talking about. This is the first time a single vendor has shipped the full loop from prompt, to design system, to interactive prototype, to production code, inside one walled garden. The primitive underneath it is not "image generation." It is codebase as grounding context.

Step 1 Prompt or sketch, or URL

→

Step 2 Design grounded in codebase

→

Step 3 Prototype interactive, live

→

Step 4 Claude Code production handoff

Source: Anthropic Labs, April 17, 2026

Primitive one: codebase as design system

During onboarding, Claude Design reads your codebase and your design files, and builds a design system from that. Every project afterwards uses your colors, typography, and components by default. You can refine it over time. Teams can maintain more than one. That is not a design tool feature. That is a long context grounding pattern. The codebase is the source of truth, the model extracts the design system as structured context, and every downstream generation inherits it without restating. Brilliant, the education company, reports that pages which took 20 or more prompts in competing tools took 2 prompts in Claude Design. The delta is the grounding. For AI engineers building products: this is the pattern to steal. You are almost certainly re-describing your constraints on every request. Your codebase already encodes them. The move is to extract a structured representation once, cache it, and inject it as grounding context for everything that follows.

The primitive in one diagram your-app/   tailwind.config.js   components/   tokens/   Button.tsx   Card.tsx → extract once Design system colors, spacing, typography, components, all inherited by default Same pattern: extract structured context once, inherit it everywhere downstream. Claude Design does it for design systems. You can do it for anything.

The tweetable bit: Claude Design is not a Figma competitor. It is a demo of codebase as grounding context, with Figma's business model as collateral damage.

Primitive two: thinking budgets are dead

Opus 4.7 removes the extended thinking budget parameter entirely. Setting budget_tokens now returns a 400 error. Adaptive is now the single thinking-on mode, where the model allocates its own reasoning tokens per request. Anthropic's internal evals say adaptive beats fixed-budget extended thinking across the board. Two other breaking changes matter. Temperature, top_p, and top_k now return 400 if set to anything non-default. Thinking content is omitted from responses by default, so if your product streams reasoning to users you will see a long pause before output starts. Set display to summarized to restore visible progress. The pattern here is the same pattern as primitive one. Anthropic is removing knobs. Fixed reasoning budgets, sampling temperature, visible thinking, all traded for model-allocated defaults. If you tuned an Opus 4.6 pipeline by hand, you will need to re-tune it as prompting rather than as parameters.

What breaks on the migration Opus 4.6 Opus 4.7 budget_tokens: 32000 adaptive thinking temperature: 0.7 400 error thinking visible by default off by default, opt-in summary

Quick Hits

New tokenizer, new token count. Opus 4.7 ships with a new tokenizer that uses roughly 1 to 1.35 times as many tokens per input. The count_tokens endpoint will return different numbers than 4.6 for the same text. Budget accordingly. xhigh is the new default for coding. Anthropic added an effort level between high and max. Docs recommend starting with xhigh for coding and agentic loops, and never going below high for anything intelligence-sensitive. 2576 pixel vision. Max image resolution jumped from 1568 to 2576 pixels, with better pointing, counting, and bounding-box localization. Penetration testing firm Xbow reports Opus 4.7 scored 98.5% on their visual acuity benchmark versus 54.5% for Opus 4.6. If you downsample product screenshots before sending them to Claude, stop. Task budgets, not token budgets. Opus 4.7 introduces task budgets in public beta, which set a rough target for total tokens across a full agentic loop including thinking, tool calls, tool results, and final output. The unit of cost planning moves up one level of abstraction.

The Take

Anthropic is not betting on being the best model. Anthropic is betting on owning the loop. Opus 4.7 plus Claude Code plus Claude Design plus the Cowork agent plus the Excel and PowerPoint integrations, all sharing one design system and one memory layer, is a full vertical stack. The competition for Anthropic is no longer OpenAI. It is every SaaS that currently lives between prompt and production. For AI engineers shipping products on top of Anthropic: the strategy question is no longer "which model." It is "which slice of the loop am I building on, and does Anthropic ship the native version of it in 6 months." If the answer is yes, you are a feature, not a product.

The Open Question

Adaptive thinking hides the budget knob, and visible reasoning is off by default. Both moves reduce what developers can observe about how the model is deciding. That is a tradeoff against interpretability at exactly the moment frontier labs are publishing papers on CoT faithfulness decay. Is Anthropic quietly deciding that visible CoT was never reliable enough to expose to users anyway.

Next week: the CoT faithfulness paper showing models learn to hide deceptive reasoning from monitors while keeping the surface clean. The monitoring strategy every frontier lab is using has a failure mode, and the paper reproduces it.

ResearchAudio.io

Sources: Opus 4.7 launch, Claude Design, 4.7 migration docs