The fine print behind GPT-5.6 and Claude Sonnet 5

The last ten days delivered the biggest launch stretch of the year. On June 26, @OpenAI announced a limited preview of GPT-5.6: three models called Sol, Terra, and Luna, spanning a frontier flagship, a balanced everyday workhorse, and a fast budget tier. Four days later, Anthropic shipped Claude Sonnet 5, its most agentic mid-tier model yet, with a 1 million token context window and an introductory price of $2 per million input tokens and $10 per million output through August 31.

Both headlines sound like straightforward good news: more capability, lower prices. The real story in each case is in a follow-up post, and both follow-ups point the same direction. The number in the announcement is not the number you will actually live with.

OpenAI walked its launch past Washington first

The detail that jumped out was in OpenAI's second post. The company said it plans to make all three models generally available in the coming weeks, but for now, "at the request of the U.S. government," it is starting with a limited preview among a small group of trusted partners in Codex and the API, per @OpenAI.

Read that against last month's news. In June, an export-control order forced Anthropic to switch off its brand-new Fable 5 model for every customer worldwide, an episode we covered here. That pause lasted 19 days before Fable 5 came back on July 1 with additional safeguards. OpenAI clearly took notes. Rather than launch first and get a national-security letter later, it previewed GPT-5.6 to the government ahead of release and is rolling out slowly, on purpose.

My read: this is now the template. A frontier model release in 2026 has a government checkpoint in it, whether it happens before launch by choice or after launch by order. If you are building on these models, availability is no longer just a question of vendor uptime. It is a policy variable.

There is still plenty to like in the launch itself. Terra is pitched at roughly GPT-5.5-level performance at about half the cost, and Sol is the only model that unlocks a new maximum reasoning effort and an "ultra" mode that runs teams of subagents. But you cannot use any of it yet unless you are on the partner list.

Sonnet 5's "same price" costs about 40 percent more

Anthropic's Sonnet 5 launch came with a friendlier headline: same standard pricing as its predecessor, plus a two-month introductory discount. Then Simon Willison ran the numbers. As @simonw posted, "the new tokenizer makes it ~1.4x more expensive for English."

Here is the mechanism, from his full write-up. Sonnet 5 splits text into tokens differently than Sonnet 4.6. The same document now produces roughly 30 percent more tokens, and since you pay per token, an unchanged sticker price still means a bigger invoice. By his measurements English text costs about 1.4 times more, Spanish about 1.33 times, Python code about 1.27 times, while Simplified Mandarin comes out roughly flat.

To be fair to Anthropic, tokenizers change for legitimate engineering reasons, and the intro discount offsets much of the increase for now. But the discount ends August 31, and nothing in the launch material made the tokenizer math obvious. It took an outside developer counting tokens on real documents to surface what migrating actually costs.

What this means if you run a business on AI

Two launches, one lesson: the announced number and your number are different numbers. Availability gets shaped by policy you do not control. Cost gets shaped by measurement units the vendor can redefine. It is the same as a contractor quoting you the old rate per square foot and quietly changing what counts as a square foot.

None of this is a reason to sit out. Sonnet 5 and GPT-5.6 are genuinely strong tools. It is a reason to check three things before you commit a workflow to any model: what it costs on your actual documents, not the pricing page; what happens to that price when the promotional window closes; and what your fallback is if access changes for reasons that have nothing to do with you.

Most small businesses have never priced their AI usage against their own real workload. That is exactly the kind of thing New Face Design's free process audit is built to catch, before the intro pricing expires and the invoice does the talking.