September 5, 2026

Bun 1.4.2 fixes regressions introduced in 1.4.1 that could break Elysia builds and retain AsyncLocalStorage context in memory. Maintainers also document a worker event-order fix affecting Discord clients and crash fixes for long-running or musl-based processes. If those paths match your stack, prioritize a staged upgrade with build, request-context, and memory checks.

Neon ran 100 synthetic support tickets through 42 models, grading policy compliance as well as cost and latency. Its cheapest listed models were not always cheapest per usable answer, so catalog token prices alone are a weak purchasing guide. Compare cost per accepted result on your own task set before changing routing; this is a vendor benchmark, not a general model ranking.

Vercel now exposes Astra through AI Gateway using the model identifier openai/gpt-6-astra, with setup instructions for coding agents. Existing Gateway users have another integration route without implementing a separate provider client. Check access and workload costs before adopting it; the announced capability improvements remain vendor claims.

Researchers published an investigation of roughly 18,000 public-wiki posts attributed to autonomous agents sharing web-research answers and discussing sandbox restrictions. The practical issue is that a nominally read-only browser can still cause writes on sites that mutate state through GET requests. Audit destination and write controls outside the model, and test those controls in an owned environment; the researchers’ attribution and internal-training explanations should remain explicitly qualified.

Simon Willison compared bicycle-riding pelican SVGs across Astra reasoning levels and the 5.6 models, reporting that fewer output tokens narrowed Astra’s cost premium. That is a useful reminder to measure whole-task cost, but an illustrative drawing task cannot justify switching a coding agent. Monitor broader repeated evaluations and keep production routing unchanged on this evidence alone.