This week in web: Next.js 16.3, Chrome's built-in AI, and the new AI model pricing tier

Three things shipped in the past two weeks that are worth tracking before you make tooling decisions.

Next.js 16.3: the SPA speed gap closes

Next.js shipped version 16.3 across three blog posts between June 25 and 29. The headline feature — Instant Navigations — addresses a real objection that comes up in almost every conversation about server-rendered React versus SPAs.

The tradeoff has always been: Next.js gives you SEO and fast first loads, but navigating between pages has a perceivable pause while the server renders. Client-rendered SPAs feel instant because the shell is already loaded. Teams picked a side and lived with the compromise.

Instant Navigations closes the gap using Partial Prefetching: a reusable shell per route is cached on the client, so navigation renders instantly while data streams in behind it. Routes get three explicit modes — Stream (shell-first, data follows), Cache (full route cached client-side), Block (old behavior for pages that need complete data before showing). The SPA "feel" is now opt-in, not opt-out.

The same release ships meaningful Turbopack improvements: persistent build cache (successive builds reuse prior work), a Rust-ported React Compiler (faster transforms), and `import.meta.glob` for Vite-style file imports. On large projects, build times drop noticeably.

There's also a set of AI tooling additions that signal where the development workflow is heading: AGENTS.md serves version-matched documentation to coding agents running in your project; a new Agent Browser mode lets agents drive a real browser and read your React component state tree directly. Vercel is clearly betting that development will increasingly involve agents navigating running apps, not just editing static files.

Chrome 148: on-device AI is now a stable platform API

Chrome 148 shipped the Prompt API as stable. Any website running in Chrome can now call Gemini Nano — an on-device language model — directly from JavaScript. No API key, no network request, no per-request cost.

The API accepts text, image, and audio inputs and returns structured JSON via schema constraints, making it composable with application data. Chrome also ships Gemma 197M for task-specific built-in APIs — summarization, translation, language detection — which abstract away prompt engineering and scale to a wider range of devices.

Chrome 149 brings WebMCP into origin trial: a protocol that turns websites into structured tool providers that AI agents can call by name, without scraping.

The practical constraints are real. Firefox and Safari have not committed to the Prompt API, and it shipped outside the normal web standards process — generating active discussion in the developer community. For a Chrome-only internal tool or a progressive enhancement with a clear fallback, this is shipping infrastructure today. For a public-facing feature you want working across all browsers, you need a server-side fallback path.

The AI model pricing shift

Claude Fable 5 — Anthropic's newest model, launched June 9 — moved to usage credits on July 7. Pricing: $10 per million input tokens, $50 per million output tokens. That's roughly double Opus 4.8.

Fable 5 is built for a specific use case: multi-day autonomous sessions, large-scale code migrations, long-horizon work where the model writes and tests its own code across many steps. For that work, the price can justify itself. For most web development tasks — generating components, code review, refactoring, content pipelines — it's the wrong tier.

The model worth benchmarking for production use is Claude Sonnet 5: $2 per million input tokens, $10 per million output tokens, introductory pricing through August 31, 2026. Anthropic positions it at near-Opus-4.8 performance on agentic tasks. If you're building AI features into a web app, Sonnet 5 is the practical default until the intro pricing ends and you have usage data to justify moving up.

The competitive context worth noting: Chinese models now serve roughly 45% of OpenRouter traffic (up from less than 2% a year ago). Tencent's Hy3 — a 295B-parameter mixture-of-experts model — is available free on OpenRouter through July 21. The pressure on AI pricing is structural.

The browser AI competition, briefly

All major browser vendors are now competing on AI execution rather than search integration. Perplexity (Comet), The Browser Company (Dia), Opera (Neon), and OpenAI (Atlas) are each launching browsers designed to act on your behalf inside web pages — filling forms, booking appointments, navigating multi-step workflows without manual input.

For web developers building production apps, this has a practical implication: browser agents will increasingly traverse your UIs the same way real users do. Navigation patterns, form structure, and ARIA labels start mattering for machine traversal, not just accessibility. Worth auditing with this in mind.

What this means together

The three main stories this week share a common thread: AI is being wired into the web stack at the framework level (Next.js 16.3), the browser level (Chrome 148 Prompt API), and the model infrastructure level (Fable 5 pricing, Sonnet 5 as default).

If you're building a React app, Instant Navigations removes a real objection to server-first architecture for performance-sensitive projects. If you're adding AI features, Chrome's Prompt API gives you an offline, zero-cost path for Chrome users with a clear degradation boundary. And if you've been defaulting to Fable 5, Sonnet 5 is the better starting point until you have a specific reason to spend more.