Claude Sonnet 5

Sonnet 5 is the model most people should default to. It occupies the position that matters commercially: close enough to flagship quality on ordinary work that the difference rarely shows, at a price that makes high-volume use viable.

The practical test is simple. Run a task on Sonnet and on the tier above. If you cannot reliably tell the outputs apart, you were never buying capability — you were buying reassurance.

Sonnet's real advantage is the balance. It is fast enough for interactive use, capable enough for real engineering and analysis, and cheap enough that you stop rationing calls. That last point changes how people build: when a model call is cheap, you can afford to make several and pick the best, which often beats one call to a better model.

It follows instructions precisely, which matters more than raw capability for most production work. A model that reliably returns the JSON shape you asked for is worth more than a smarter one that improvises.

On genuinely hard reasoning — problems with several interacting constraints, or where the right answer requires noticing something not stated — it falls behind Opus and Fable. It is also more likely to accept a flawed premise rather than challenge it.

For long agentic runs, the per-step gap compounds. Twenty steps at slightly lower reliability is a meaningfully lower chance of finishing correctly.

Make Sonnet 5 the default and escalate deliberately. Most teams that audit their model spend find the same thing: the majority of calls did not need the tier they were sent to. Start here, measure where it actually fails, and route only those tasks upward.