Claude 3.7 goes hard for programmers…

Anthropic’s Claude 3.7 is pushing programming-focused AI into a new tier by combining a stronger base model with a “thinking mode” and, most importantly for developers, a new CLI tool called Claude Code that can build, test, and run code inside a real project—creating a tight feedback loop that’s meant to reduce the back-and-forth between humans and models.

The programming impact starts with performance claims. Claude 3.7 Sona, the newly released model, is described as beating its own prior baseline while adding a thinking mode modeled after the success of DeepSeek R1-style approaches in open “reasoning” models. In benchmark terms, Claude 3.7 is said to have jumped ahead on a human-verified software engineering test set built from real GitHub issues. The headline figure is 70.3% of issues solved, surpassing other models including OpenAI O3 mini High and DeepSeek. The transcript then shifts from leaderboard talk to hands-on testing, where Claude Code is positioned as the practical mechanism behind the hype.

Claude Code is a research-preview CLI installable via npm. It uses the Anthropic API directly and comes with a steep cost: over 10× the price of models like Gemini Flash and DeepSeek, at $15 per million output tokens. After installation, the CLI provides a command that scans an existing codebase, generates a markdown context/instructions file, and then opens an interactive session where the model can propose changes and write files to disk.

In early tests, Claude Code behaves like an agent that can manage project structure and testing. A simple “random name generator” task results in new files plus a dedicated testing file, reflecting a workflow aligned with strongly typed languages and test-driven development. When tests fail, the tool can iterate—rewriting logic and re-running until the test suite passes.

The transcript’s more demanding test targets a moderately complex UI: a TypeScript + Tailwind + “spelt” front end that records microphone input and visualizes a waveform. Claude Code requires many confirmations, but it produces a working interface with interactive waveform controls and graphics. A comparison run using OpenAI O3 mini High generates an inferior result and, on inspection, misses key stack details (not using TypeScript/Tailwind as expected and failing to apply the newer “spelt 5 Rune syntax”). Claude Code’s session cost is reported at about 65 cents for that UI build.

Still, the tool isn’t portrayed as a universal fix. A final attempt—building an encrypted app after Apple discontinued end-to-end encryption in the UK—fails to run despite extensive code changes. The transcript emphasizes a practical limitation: even strong coding agents can get stuck on runtime errors, and heavy reliance can leave developers without the context to debug.

The closing pitch ties Claude Code’s strengths to backend productivity via Convex, an open-source reactive database with typesafe queries and server functions. The claim is that AI coding works better when the backend is structured in predictable, typescript-native patterns—making autonomous “vibe coding” more reliable. The overall takeaway is clear: Claude 3.7 and Claude Code can materially accelerate real development, but they still demand human oversight when the task hits tricky runtime or security constraints.