Claude Opus 5 closes 2024 SDK gaps but still struggles with 2025 API changes

·1 views

A developer re-ran their open-source tool SDKProof against Anthropic's newly released Claude Opus 5 to measure how accurately it generates code for recently updated SDKs. The tool evaluates model output by type-checking generated code against real installed packages, with a compiler — not another AI — determining pass or fail. Opus 5 scored 100 on Vercel AI SDK 7 and Zod 4, up from 90 on the previous Opus 4.8, correctly using renamed parameters and updated syntax introduced in those libraries. However, Prisma 7 and Next.js 16 — both of which shipped breaking changes in late 2025 — scored 87 and 92 respectively, as the model still generates outdated API patterns it likely learned during training. The findings suggest that a model's SDK accuracy is tied less to overall capability and more to how recently a library's last major version was released relative to the model's training cutoff.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Parent Shares Experience Teaching Children the Forth Programming Language

A writer named Anna Liberty has published an article on her website Grace Full Liberty about teaching children the Forth programming language. The piece explores her personal approach to introducing kids to this stack-based language, which dates back to the 1970s. Forth is considered unconventional compared to modern beginner languages like Python or Scratch. The article was shared on Hacker News, where it attracted some attention from the programming community.

0 comments Read more at Hacker News

ProgrammingDEV Community ·

How Oracle Silently Handles Duplicate Keys During Unique Index Updates

Oracle Database allows multi-row updates on unique columns without raising constraint errors, a behavior not shared by databases like PostgreSQL, MySQL, SQLite, and MongoDB. A developer investigated Oracle's internal handling by creating a table with a unique index and a row-level trigger to pause execution between row updates. The experiment revealed that Oracle does not process unique index updates in a straightforward row-by-row manner that would trigger duplicate key violations mid-statement. By dumping index leaf blocks during execution, the researcher aimed to observe how Oracle manages temporary duplicate states within a single update statement. The findings shed light on how Oracle's B-tree index mechanics differ from other databases when enforcing uniqueness constraints during set-based operations.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Browser game Dario Dash uses client-side Groth16 proofs to verify scores on-chain

Dario Dash is a browser-based endless runner built on the Dusk blockchain that uses zero-knowledge proofs to ensure submitted scores are legitimate. After completing a ranked run, the player's browser generates a Groth16 proof locally using a Circom circuit of roughly 421,000 constraints, covering movement, collisions, item pickups, and scoring rules. The smart contract independently reconstructs public inputs, recomputes the obstacle schedule from the run seed, and verifies the proof before accepting any score. Each seed can only be used once, preventing replay attacks, and the proof is bound to the submitting account to stop players from copying others' valid proofs. Proof generation takes 10 to 60 seconds depending on hardware, and requires downloading an approximately 200 MB proving key that is cached by the browser after the first use.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

AI Coding Tools Will Hollow Out Mid-Tier Developer Roles, Not the Profession

A veteran engineering leader argues that AI-assisted coding will not eliminate software developers but will compress the middle tier of the market, much like cheap digital cameras gutted mid-range photography work. Routine tasks such as CRUD apps, simple integrations, and internal tools built to a known spec are now cheaply handled by AI models, threatening the junior-to-mid segment most. While some developers may move upward by developing architectural judgment and systems thinking, the author warns this represents leaving the middle tier rather than saving it. Even if many mid-level developers upskill, the high end of the market — architects, security specialists, and trusted technical advisors — is narrow by definition and cannot absorb everyone. The article draws on the author's two decades of leading dev teams and personal experience rebuilding a side project in weeks with AI that previously took months of evenings.

0 comments Read more at DEV Community