GPT-5.6 goes public, Tencent open-sources a monster, and coding models get cheap

Every Tuesday I round up what actually happened in AI the week before, filtered through one question: does it change what I do Monday morning? Not what sounds big in a headline.

Three weeks ago one of last week’s headline models was still locked behind a government review. Now anyone can call it. Here is what mattered.

GPT-5.6 finally shipped to everyone

On July 9 OpenAI moved the GPT-5.6 family, the Sol, Terra, and Luna models, from a 20-org government preview to general availability. ChatGPT, Codex, and the API, rolled out worldwide in about a day. The safety review tied to its cybersecurity abilities cleared faster than expected. The same week OpenAI shipped GPT-Live, a voice model that listens and talks at the same time instead of waiting for you to stop. The bigger GPT-Live-1 runs on paid plans, a mini version on the free tier.

What it means for builders: last time I wrote about this it was a rumor with three code names and a waitlist you were not on. This week you can call it. That is the pattern now. Frontier models preview to a handful of vetted orgs, clear a review, then land on everyone at once. If you build on OpenAI, the gap between announced and available is the thing to plan around. And if you build anything with voice, full duplex is the line between a phone tree and a conversation. A model that can be interrupted changes what you can build.

Tencent open-sourced a serious model and dropped the fine print

Tencent released Hy3 on July 6. It has 295 billion parameters but only 21 billion fire on any given task, so it runs far lighter than the headline number. A 256K context window. Up on Hugging Face today. The story is not the benchmarks, though they are strong. It is the license. Tencent’s April preview of this model shipped under a house license that banned use in the EU, the UK, and South Korea. Hy3 is full Apache 2.0. No territory carve-outs. Use it, ship it, sell what you build on it.

What it means for builders: an open model you can actually build a business on needs two things. Good weights, and a license that does not trap you. For most of this year the good open weights came wrapped in terms a lawyer would flag. That is changing. When a top-five tech company puts 295 billion parameters under Apache 2.0 with no strings, the floor for free and unrestricted just moved up. If your costs are eaten by per-token API bills, this is the week to price out running your own.

Cheap coding models are a category now

Three code-focused models landed at prices that would have been the discount tier six months ago. xAI shipped Grok 4.5 on July 8, aimed at code, around two dollars per million tokens in. GPT-5.6 ships with a Codex mode built for it. Anthropic’s Sonnet 5, out the week before, sits at two dollars in and ten out through the end of August. The pattern is not subtle. The model makers decided coding is the killer app, and they are racing each other down on price to own it.

What it means for builders: the thing I do all day, hand work to a model that writes code and check what comes back, got cheaper again. A year ago the good coding model was the expensive one. Now there are three of them fighting over your bill. If you priced out an AI coding workflow in the spring and decided it cost too much, the math is different now. Re-run it before you write the idea off.

What I did about any of it

Not much, on purpose. My calculator site shipped two new tools last week and I did not touch either build. But I did ship a bug. One calculator went live with a Calculate button that did nothing. Over 500 tests passed. The math was perfect. The button reached for a script that was never built and threw a 404 at real users.

None of last week’s news would have caught that. A faster model does not check your work. A cheaper one does not either. The lesson I keep relearning is the boring one. Better tools raise the ceiling on what you can build. They do not remove the job of looking at the thing you shipped and pressing the button yourself.

So I added the check I was missing. A robot loads the real page and presses the button before anything goes live now. Cheap models, expensive habit.

More next Tuesday. If someone forwarded you this, subscribe at blog.jaredhebb.com/newsletter.

Jared