This was one of those weeks where your AI news feed scrolled faster than you could read it. Three major model releases in three days, each one with a launch post, a pile of charts and a chorus of people declaring that everything has changed.

Some things did change. Most of your job didn't. Here's what actually happened, what I think matters for developers, and how I'm deciding whether to switch anything.

What was released

In order:

• 5 August: gpt-oss. OpenAI released two open-weight models, gpt-oss-120b and gpt-oss-20b, under the Apache 2.0 licence. It's OpenAI's first open-weight language model release since GPT-2.
• 5 August: Claude Opus 4.1. Anthropic released an update to Opus 4, its top model and one many developers use for coding and agent work.
• 7 August: GPT-5. OpenAI's new flagship, available to all ChatGPT users, including people on the free plan.

Any one of these would have been the headline of a normal week.

GPT-5: the model that decides when to think

The most interesting thing about GPT-5 isn't a benchmark. It's the design. OpenAI describes it as a unified system: instead of you choosing between a fast model and a "thinking" model, GPT-5 decides for itself when a question deserves a quick answer and when it needs to reason for longer.

If you've used ChatGPT over the past year, you'll know why that matters. The model picker had grown into a small puzzle: which one is fast, which one reasons, which one is good at code? Most people just used the default and never touched the rest. Folding that decision into the system removes a choice most users were never equipped to make.

For developers, there are two practical consequences:

• Response times will vary more. The same chat can answer one message instantly and think for a while on the next. That's by design.
• Your colleagues just got an upgrade. Making it available to free users means the people around you, including non-developers who paste code into ChatGPT, are now using a newer model by default. Expect the quality of "I asked ChatGPT and it said..." to change, for better and occasionally for worse.

gpt-oss: why the licence is the headline

Open-weight models aren't new. Llama, Mistral, Qwen and DeepSeek have been around for a while. What's notable here is OpenAI joining in, and the Apache 2.0 licence, which is about as permissive as licences get: use it commercially, modify it, ship it.

The smaller gpt-oss-20b is the one I find most interesting for everyday developers, because a model of that size is within reach of running on your own hardware. That opens up the use cases where sending data to an external API is a problem: internal documents, sensitive records, offline tools, or just experimenting without watching a usage bill.

The trade-off doesn't change: running a model yourself means you own the hosting, the updates and the guardrails. "Open" means you're allowed to. It doesn't mean it's free to operate.

Opus 4.1: the quiet one

Opus 4.1 got less attention because it landed two days before GPT-5, but if you already use Claude for coding, it's the release that affects your daily work most directly. Point releases like this are often the most useful kind: same behaviour you're used to, with improvements, and nothing to relearn.

What I'd actually do this week

Every big launch comes with the temptation to switch everything immediately. Here's the routine I'd suggest instead.

1. Test on your own tasks, not the launch charts

Launch benchmarks measure what the vendor chose to measure. Your work is specific: your framework, your codebase, your conventions. Keep a small set of real tasks you've done before, where you know what a good answer looks like:

• a bug you fixed recently, with the error message you started with
• a refactor of a messy class in your codebase
• a feature description that needs tests
• a code review of a real diff with a known problem in it

Run the new model on those, side by side with what you use now. An hour of this tells you more than a week of reading takes.

2. Change one thing at a time

If you have an LLM in production, don't swap the model and the prompt in the same release. Pin the model version, change it deliberately, and compare outputs on saved examples. A model upgrade is a dependency upgrade. Treat it like one.

3. Don't chase every release

There will be another big launch next month, and the month after. Switching tools every time has a cost: new quirks to learn, prompts to retune, habits to rebuild. If what you use works well, "I'll evaluate it properly when I have a reason" is a perfectly good answer.

The bigger picture

The striking thing about this week isn't any single model. It's that the best AI tools are now coming from several companies at once, in both closed and open forms, and the gaps between them keep shifting. That's good news for developers. It means the useful skill isn't loyalty to one tool, it's knowing how to evaluate them quickly and plug them into a workflow that still has tests, reviews and a human deciding what ships.

Have you tried GPT-5, gpt-oss or Opus 4.1 on real work yet? What's the first task you'd use to compare them?