What AI builders are saying about Claude Opus 4.5

Get every new Claude Opus 4.5 video summarised in your inbox.

Building real software in under an hour

Two creators independently found that Claude Opus 4.5 could produce fully functional, commercially viable applications within a single working session. Wes Roth demonstrated three distinct projects — a browser-based RPG with voiced dialogue and combat mechanics, a sailing simulator with dynamic weather physics, and a spreadsheet application with live stock data — each completed in roughly 25 to 35 minutes and costing between $17 and $26 in API credits. Jack Roberts similarly used Claude Opus 4.5 as the core engine for an end-to-end YouTube video production pipeline, assembling competitor research, scriptwriting, audio processing, and video assembly into a workflow that produces a finished faceless video for approximately $2 in generation costs.

Both creators framed Claude Opus 4.5 as a credible replacement for commercial software or specialist labour, rather than merely a productivity aid. Roth's framing was that AI coding agents had reached a threshold where building a personal tool is cheaper and faster than subscribing to an existing product. Roberts positioned the pipeline as a replacement for traditional video editors, provided creators contribute their own voice and rewrite AI-generated scripts to avoid what he called the 'AI slop apocalypse'.

Benchmark performance as evidence of a capability leap

Coverage of Claude Opus 4.5 paid close attention to benchmark results as a way of contextualising the real-world demonstrations. Wes Roth cited a CoreBench score of 95.5% — measuring the model's ability to replicate machine learning research papers — reached in December 2025, contrasting it sharply with a score of 21% recorded just three months earlier in September 2024. This trajectory was presented not merely as a product milestone but as supporting evidence for Anthropic co-founder Jack Clark's argument that fully automated AI research and development could arrive by the end of 2028.

Roth's treatment of these figures was notably sober rather than celebratory. The benchmarks were offered as data points within a broader discussion of recursive self-improvement, economic disruption, and the limits of current alignment techniques — a framing that sits alongside, but does not contradict, the more enthusiastic applied demonstrations found in his and Roberts' build videos.

Practical access limits and cost considerations

Creators who tested Claude Opus 4.5 directly through Anthropic's paid plans noted meaningful constraints on availability. Wes Roth explained that the model consumes up to 50% of a user's weekly usage allowance, meaning heavy use of other Anthropic models can crowd out access to Opus 4.5 entirely. He also noted that after 7 July 2026, the model would no longer be available through standard paid plans and would require API access instead.

On the question of content safety, Roth reported that classifiers which had been flagged as a potential concern before launch proved non-intrusive in his testing. He recommended using the feedback mechanisms within Claude Code or Claude.ai to report any false positives. These practical notes complement Roberts' cost breakdown — roughly $2 for a complete video project — suggesting that for builders willing to work via the API, the per-task cost can remain very low even as access through consumer plans becomes more restricted.