Bijan Bowen begins with Anthropic's positioning for Claude Sonnet 5, including its introductory token pricing, updated tokenizerTokenization turns input into discrete units that a model can represent and process. and benchmark gains over Sonnet 4.6. The model approaches older Opus results in some coding tests and adds another selectable effort level, suggesting a capable middle-tier option for agentic development.
His practical testsEvaluation is the systematic assessment of a system's behaviour or results against explicit criteria. are much less consistent. Sonnet 5 builds a feature-rich browser operating systemCode generation uses AI or another automated system to create source code from instructions, examples, schemas, or higher-level specifications., a drum-kit simulation and several 3D games, but long generation times, awkward interfaces and incomplete physics recur. A laptop replica gets some controls and details right while reversing the keyboard, and a Formula 1 game includes two camera modes but suffers from inverted steering and poor viewpoints.
The strongest individual outputs are still offset by slow high-effort runs and heavy token useCost per completed task measures the total AI, tool, infrastructure, retry, and repair expense for each verified useful outcome.. Bowen concludes that Sonnet 5 is competent rather than exceptional, and questions whether its post-introductory price can compete with faster, cheaper coding models. The new Claude desktop beta for Linux interests him more than the model upgrade itself.
Watch on YouTube
