GPT-5.2 at 10: Progress, Pricing, and Public Perception

James Turner

Welcome back to 48-Hour AI—I'm James Turner, and today we’ve got a huge milestone to talk about: GPT-5.2, dropping on OpenAI’s 10-year anniversary. Yep, it’s been a decade. Oddly, feels like both forever and—I dunno—a blink. Anyway, OpenAI’s not pulling their punches with this one. I mean, get this: GPT-5.2 just posted a 74% win rate on GDPVal, which for folks who missed the past episodes, that’s the main benchmark for “economically valuable” knowledge work. Last time, GPT-5.1 managed, what, like 39%? So this is basically doubling up, and more. And that’s not just cherry-picked—this thing is beating or at least tying human experts across 44 different professions. Like, real humans, not just the ones who live in Excel spreadsheets all day.

James Turner

But benchmarks—let’s get a little nerdy—are where things get wild. It scored 100% on AIME 2025, which is a top-tier math competition, and over 92% on GPQA-Diamond, the science and knowledge grand tour. For the code and SWE folks, SWE-Bench Pro accuracy took a big jump, especially the “xhigh” parameter, so if you remember last episode, 5.1 Codex Max really struggled there. Now, 5.2’s back in the game. Even more, on the Verified harness, it hit 80%. That’s real productivity.

James Turner

Another thing people might overlook? Long-context handling. This allows the model to “read” and actually use far more information in one shot. If you look at the MRCRv2 results, it’s pretty clear: 5.2 keeps up its memory way better as the input tokens increase. That’s big for tasks like summarizing or keeping a complex discussion straight rather than dropping context after the first few paragraphs.

James Turner

Now, I know—benchmarks aren’t everything. We covered that when DeepSeek v3.2 launched. Labs love numbers, but real users care if all that actually translates to getting stuff done. Still, nobody can look at these numbers and not see a step-change. So, of course, OpenAI’s hyping the “professional work” angle pretty hard. And, honestly, it seems justified—at least on paper. There’s nuance, though, but I’ll get to that in a sec.

James Turner

Can we talk about the elephant in the room? Price. Seriously, a 40% jump over GPT-5.1. The API clocks in at $1.75 per million input tokens, and a mind-boggling $14 per million output tokens, though you can whittle that number down with a hefty cache discount—if you’re clever. And Twitter, Discord, even some pretty salty subreddits are losing it. Developers are out here doing back-of-the-napkin math and freaking out about high-effort tasks burning up $168 just for a single million output tokens. For people with agent stacks or building RAG pipelines—ouch.

James Turner

A lot of the chatter is “How much is benchmark performance worth if you’re actually limiting people’s access?” Perplexity was the first one to get GPT-5.2 for their Pro and Max users—even before ChatGPT Plus I think—which led to a whole new round of benchmarking races, just to see if the premium cost justified it. And the frustration was real: rate limits kept slamming the brakes even for people who shelled out for max plans. Cursor, Copilot, VS Code, pretty much every big ecosystem tool rolled out GPT-5.2 integration almost overnight, but it didn't stop the sticker shock from, well, stinging.

James Turner

It reminds me of a theme we've hit on in other episodes—if the tool's powerful, but you have to ration usage or rethink workflows to avoid bill shock, a lot of devs just...won't. You end up with this weird situation where all the “superpower” upgrades exist, but teams have to benchmark like crazy before they even consider switching over fully. And, of course, that creates a kind of arms race—people hustling to test out real-world performance before the invoices pile up.

James Turner

And Twitter—X, or whatever—lit up with wild takes, memes, and, like, full-on existential debates about value. One thread I saved had two posts right next to each other: “Is GPT-5.2 suddenly worse for everyone?” and “GPT-5.2 is so good it’s AGI!” Just...classic AI launch day stuff. Alright, let’s follow that thread into the reactions and actual user experiences. That’s where things get spicy.

James Turner

So—public perception. It’s a mess, and also kind of hilarious. A lot of praise for the scientific and code benchmarks, like people absolutely floored by the math and science numbers. But the flipside? Regular users pile into Reddit with “my job as a developer is done, time to hit the beach,” and then you have a bunch more ranting, “Wait, is GPT-5.2 actually worse lately?” Sometimes both on the same thread. The memes come fast. There’s this one—“GPT-5.2 is AGI,” right under a screenshot of it tripping over, like, counting the letter ‘R’ in ‘strawberry.’ Just brutal.

James Turner

Don’t get me wrong, some of these critiques hit real issues. The notorious “miscounting the letters” bug? Still around. Also, it’ll spit out a pretty spreadsheet, but, yeah, the numbers don’t always add up if you sanity check them. The vision upgrades are real but, even in OpenAI’s own notes, Gemini 3 still edges ahead in some key areas, especially in multimodal tasks. And for verified SWE-bench code, Opus 4.5 sometimes pulls ahead in certain harnesses. So we’re far from crowned champion territory.

James Turner

The bigger industry picture—remember when we talked about agent workflows and local context in recent episodes? Now, that all ties back in. You’ve got folks who want the bleeding-edge power, but grapple with API caps, pricing, or workflow quirks, especially on new integrations like Cursor or Copilot. Rate limiting, context window weirdness, even stuff like Cursor’s “rewind” feature confusing people about what actually gets restored—real world’s always messier than the launch party.

James Turner

And then there’s all the noise. Memes about OpenAI’s finances, hot takes on Discord about benchmarking before committing, or the running joke that “every release is AGI until it flunks kindergarten spelling.” I think—like, as we saw back in the Anthropic episode too—adoption and trust aren’t just about raw numbers. They’re about reliability, transparency, and whether these things actually make your day easier, not just your benchmarks fancier.

James Turner

So, that’s GPT-5.2—a massive leap on paper, a complicated story in practice, and a true 48-hour AI roller coaster. We’ll keep tracking how this one lands in the real world as the integrations shake out and, honestly, I wanna hear your horror stories—or your wins. That’s it for this round—catch you in 48 hours, and keep those receipts handy. There’s always more on the horizon.