In this issue · 4 chapters
OpenAI dropped GPT 5.5 last week. Opus 4.7 landed a couple of weeks ago. Are they good? Yeah, for the most part. Does it really matter? Not as much as it used to.
The model wars stopped mattering around Christmas last year, we just didn't notice.
What genuinely matters now is the harness - the tools, memory, and workflow wrapped around the model.
That’s the new battleground.
And the story on that front this week is wild: Anthropic just admitted they quietly made Claude Code dumber for the whole of March. Sam Altman fired direct shots at them in the 5.5 launch.
GPT 5.5 Is Out. Don't Focus On The Model.
The new model dropped last week. Their subtitle: "A new class of intelligence for real work." Note they didn't say coding. They said work. That’s a very specific and important choice.
The press release spells it out: researching online, analysing data, creating documents and spreadsheets, operating software, moving across tools until a task is finished. That's knowledge work. That's what most of you reading this actually do for a living…
I put out a, ahem, somewhat challenging video about this specifically that is upsetting a fair few people on social media:
@iamkylebalmer ChatGPT 5.5 is specifically coming for white collar knowledge workers. They straight up are telling us this and people are ignoring what’s... See more
How do you actually find GPT5.5? Because it’s not immediately obvious if you are using it or not!
Inside ChatGPT: click the dropdown, you get Instant / Thinking / Pro, but the actual model numbers are hidden. Thinking and Pro are 5.5, Instant is still 5.3. Go to Configure to see them. Or grab it inside Codex, their agent harness (free on the $0 plan, which Claude Code isn't). Or via the API if you're a developer.
OpenAI are tired of people obsessing over model numbers. They hid them. The direction is clear: stop thinking about "which version am I on" and start thinking about "what job do I give it." The model is the engine. The engine used to be the bottleneck. It isn't anymore. I agree with them here. Let’s talk harnesses.
The Model Is Not The Thing. The Harness Is.
Every week we get a new model. 4.5, 4.6, 4.7 Opus. 5.3, 5.4, 5.5 GPT. Blah blah blah.
The benchmarks go up a few percent. Twitter argues about whether the new one is actually better. A week later another one drops. It's exhausting and it's mostly noise.
Every benchmark result
What matters increasingly is the harness. The harness is everything wrapped around the model: the ability to use tools (Google Drive, Notion, your CRM), memory across tasks, the ability to iterate and check its own work, permissions and authorisations, multi-agent coordination. The model sits in the middle of this. The harness is everything else.
Good model + weak harness = useless. Google Gemini Pro on the API is brilliant on benchmarks. BUT nobody uses it for coding because the harness is weak.
Good model + strong harness = ChatGPT Codex
Amazing model + amazing harness = Claude Code.
The next twelve months of AI are not about which model is smartest. They're about which harness lets you do the most work. Claude Code, Codex, and Open Claw are the three that matter right now. Everything else is chat, and chat by itself is a solved problem.
Sam Altman Just Fired Direct Shots
Read paragraph two of Sam's 5.5 launch tweet. It’s the important one here:
"We believe in democratisation. We want people to be able to use lots of AI. We aim to have the most efficient models, the most efficient inference stacks, and the most compute. We want our users to have access to the best technology and for everyone to have equal opportunity."
That is not generic fluff. That is Sam firing directly at Anthropic, the makers of Claude. The same week Anthropic tested removing Claude Code from the $20 plan. The same week they got caught gaslighting about Claude quality - and finally admitted it. The same month they gated their next model - called Mythos - behind US government partnerships and a select list of enterprise customers.
OpenAI are positioning as the AI of the people.
And saying that Anthropic are only for the elites.
The comparison lands because it's true. Anthropic make 80% of their revenue from enterprise and API. They don't want $20 prosumer customers. They certainly don’t want free customers.
OpenAI have nearly a billion users, most of them free, and they're keeping it that way.
So…I’m with Sam here. His words here are backed by their actions - mass access to AI.
I loved the third paragraph less. "We love you" from a multi-billion-dollar CEO is giving weird energy. Unfortunately that’s also the one that everyone focused on instead of the genuinely fascinating second paragraph. Why you gotta be weird Sam? 😅
What This Means If You Don't Code
Lots of you will read this and think “well, I don’t code. I don’t use Claude Code, Codex or any of that sort of stuff”. This isn’t important for me. My job and livelihood are safe.
Anything that gets good at coding will be good at other multi-step knowledge work. The same skills - planning, using tools, checking its own output, iterating - transfer directly from "build this feature" to "restructure this sales page" or "turn this 64-page PDF of feedback into a completed revision by morning."
That last one isn't hypothetical. I gave Claude Code a 64-page feedback PDF from my business partner last night, told it to plan, implement, and verify the changes, then check its own work until done. It worked for three hours while I slept. Delivered a completed revision and a status report before I had coffee.
Previously that would have been my time. Or I would have hired someone. Now AI carries the task.
You can do this too. You don't need to be technical. Codex installs in ten minutes. It uses your existing ChatGPT subscription. You give it tasks in plain English. It goes and does them.
The people who've clocked this are quietly pulling ahead. EVEN IF you are not a coder this applies!
Kyle