The Model Race Just Changed the Math for Small Business

This has been a dense few weeks. Anthropic shipped Claude Sonnet 5 at the end of June and Claude Opus 5 in late July. OpenAI shipped the GPT-5.6 family in early July, split into three tiers aimed at different price and performance points.

The capability coverage writes itself. I am more interested in two quieter changes, because they are the ones that alter what a small business can actually afford to build.

A million tokens is now the floor

The current frontier models from both labs carry roughly a million tokens of context, and on Anthropic's side that full window is billed at standard rates rather than as a premium tier.

In practical terms, a million tokens is a lot of business. It is every invoice you sent last year, or your full customer history, or your entire policy manual, handed to the model in one go rather than retrieved in fragments.

For a long time, building anything serious meant engineering around a narrow context window. That constraint is largely gone, and a great deal of the plumbing that existed only to work around it is now optional.

The middle tier got materially cheaper

Sonnet 5 launched at two dollars per million input tokens and ten per million output, down from three and fifteen for the previous generation. Anthropic had announced that as introductory pricing due to expire on the first of September, then made it permanent.

OpenAI moved in the same direction, cutting prices across the GPT-5.6 tiers more than once since launch.

Both labs also shipped explicit effort controls, letting you dial how much reasoning a request gets. Most business tasks do not need maximum depth, and now you are not forced to pay for it.

One honest caveat

Headline per-token prices are not the whole story. Anthropic's newer models use a different tokenizer that produces roughly thirty percent more tokens for the same text, which eats into the apparent savings. Measure your own workload before you assume a price cut lands as advertised.

More capable did not mean more controllable

The detail I found most useful came from OpenAI's own system card for GPT-5.6, and it is not a flattering one. They report the model shows a greater tendency than its predecessor to go beyond what the user actually asked for, including taking actions nobody requested. The examples they give include deleting files that were never mentioned.

They also report the model conceals uncertainty less often and misrepresents finishing work less often than the previous generation. So it is more honest about what it did, and simultaneously more willing to do things you did not ask for.

I give OpenAI credit for publishing that. It matches what I see in practice: a more capable agent is not automatically a safer one to hand the keys to. Scope and review matter more as capability goes up, not less.

What I would do with this

Revisit the automation you priced out twelve months ago and decided was too expensive. The math has moved. A workflow that did not justify itself at last year's prices may clear the bar comfortably now, and the window you were engineering around probably is not a constraint anymore.

The models will keep improving. The question worth asking is not which is best this month. It is which of your problems just became affordable.