Laravel 11 Slims Down + Claude 3 Arrives

Archive note: this is a backdated post, written years later while rebuilding this site. It’s dated to the moment it covers, but the hindsight is real.

Two releases this March, one for each half of how I work now.

Laravel 11: subtraction as a feature

Laravel 11 (March 12) is the rare major release whose headline is deletion. The new application skeleton is radically slimmer: no app/Http/Kernel.php (middleware registration moves into the framework, customized via bootstrap/app.php), no nine config files you never touched (they live in the framework now, published only when you need to override), no Console/Kernel.php (scheduling moves to routes/console.php). A fresh app is a handful of files that all mean something.

Two years of this blog’s framework coverage has tracked the same arc — 9 settling the cadence, 10 adding types, and now 11 removing ceremony. The skeleton you git init today is closer to the framework’s idea than it’s ever been: PHP 8.2 floor, casts() as a method, per-second rate limiting, a /up health route — and Laravel Reverb, a first-party WebSocket server, ending the “Pusher or self-hosted soketi?” deliberation with an official answer.

Migration note for the worried: existing apps keep their old skeleton and run fine on 11. The slimming applies to new apps; nobody’s Kernel.php explodes. Upgrade cost: the usual mild annual afternoon.

Claude 3: the second opinion gets good

On March 4, Anthropic released Claude 3 in three tiers (Opus, Sonnet, Haiku), and the top model does something genuinely new in my testing: it makes the GPT-4 pairing workflow contested. Opus matches or beats GPT-4 on the benchmarks that matter for working devs — and exceeds it, in my experience, on the squishier dimensions: long-context comprehension (whole-module pastes stay coherent), following complex instructions without drift, and a writing register that doesn’t need de-chatbot-ifying.

The strategic note matters more than the scorecard. December’s observation — that a real multi-vendor race changes builder economics — is now operational reality: three frontier labs, leapfrogging quarterly, with my MODEL_PROVIDER abstraction layer built during the OpenAI boardroom weekend finally earning its keep on merit rather than fear. Sonnet and Haiku also reset the price-performance floor for API work — the “good enough, fast, cheap” tier is where most production LLM calls actually live, and competition there compounds directly into shipped features.

The connecting thread

It struck me writing this that both releases are the same maturity signal in different domains. Laravel at 13 years old is confident enough to remove — the framework knows what it is, so the scaffolding can stop performing complexity. The model ecosystem at 16 months old is competitive enough that no single vendor is load-bearing — the capability is becoming a commodity layer, which is what infrastructure looks like when it grows up.

Boring framework, contested models, everything faster and slimmer than last year. The stack is in the best shape of its life, and — checking the December ledger early — the useful story is indeed compounding at the floor. As scheduled.