Comparisons

A stated window is a capacity figure, not a usable-attention figure. The middle of a long input is measurably a worse place to be.

The API bill is visible on an invoice. The self-hosting bill is spread across salaries, on-call rotas and slipped projects.

A second datastore is a second thing to back up, monitor, secure and be woken by. That cost is invisible in a benchmark.

Serving one person and serving two hundred are different engineering problems. The tool that wins at one loses badly at the other.

Does not know something means retrieval. Does not behave right means fine-tuning. Was not told properly means prompting, and that is most cases.

The frontier model wins the comparison. The cheap model wins the invoice. Which matters depends entirely on your volume.

Fixed hardware cost against per-token billing, plus deprecation risk, data residency and the licence traps hiding inside models called open.

Reasoning tokens bill as output. A model thinking for 2,000 tokens before a 200-token answer charges you eleven times what you see.

Grok 4.6 and GPT-5.6 Sol Max tie at 61 on the AA index. One charges $6 per million output tokens, the other $30. The maths of that gap.