The Biggest News from GPT-5.2 Isn’t the Benchmarks 🚨

Most GPT-5.2 commentary stops at “wow, smart”. I think it is “wow, cheap — and that changes everything”…

The collapse of the cost of reasoning…

Forget the benchmark charts OpenAI showed.

Forget the 100% AIME score, albeit it’s an awesome feat.

Forget SWE-Bench Pro.

Non-members can read here or subscribe to stay updated whenever I publish.

https://openai.com/index/introducing-gpt-5-2/

The real story of GPT-5.2 is buried in a single data point from the ARC Prize leaderboard:

🧠 90.5% accuracy on ARC-AGI-1 at $11.64 per task!****

[embed]

A year ago, hitting ~88% on the same benchmark cost an estimated $4,500 per task. Today, 90.5% costs $11.64. That’s a 390× cost reduction in 12 months. Even with conservative assumptions, we’re looking at two orders of magnitude cost collapse.

This one number matters more than every other slide in the launch believe me…

The Efficiency Frontier Just Moved — Again 📉➡️📈

Look at the ARC-AGI leaderboard. What you’re seeing is not “models getting better.” You’re seeing the efficiency frontier being redrawn every few weeks.

Claude Opus 4.5 -> Grok 4 -> Gemini 3 Deep Think -> GPT-5.2 Pro

All stacked diagonally from bottom-left to top-right — each one obliterating the economics of what came before it. Accuracy is no longer the competition but…

💰 Cost per solved task is the competition!

Why ARC-AGI Was Supposed to Be “The Wall”🧱

ARC-AGI wasn’t just another benchmark. François Chollet designed it in 2019 to resist brute-force scaling.

The thesis was brutal:

• LLMs pattern-match
• Intelligence requires abstraction
• Every puzzle is novel
• Generalization must happen from minimal examples

Humans solve ~95% of ARC puzzles effortlessly.

For years:

• Best AI systems struggled to reach 5%
• The 2020 Kaggle competition topped out at ~20%
• By 2023, still only ~33%
• GPT-3 scored 0% via direct prompting 😂

ARC-AGI became the benchmark that supposedly proved:

“Scaling alone won’t get us to general intelligence.”

Chollet himself said human-level performance would take many years.

December 2024 Changed Everything ⚡

Then OpenAI’s o3-preview happened.

• 🟡 87.5% accuracy
• First system to cross the human threshold (~85%)
• But at a brutal cost:
• ~1,024 attempts per task
• ~137 pages of reasoning per attempt
• 💸 $3,000–$30,000 per task like hiring a bunch up BCG, McKinsey, Bain consultants for days to create a solution report for each task… 🧈🧈

Impressive — but economically irrelevant. Until now.

GPT-5.2 Pro: Same Capability, Different World 🌍

Eleven months later: GPT-5.2 Pro hits 90.5% 🚀 at $11.64 per task. 🪙🪙 That changes everything.At $30,000 per task, AI only makes sense if a human costs $6,000/hour. At $11.64 per task, even Mechanical Turk at $5/task is more expensive.

⚠️ We quietly crossed human-cost parity for frontier reasoning sometime in the last few months.

Most people missed it…

What’s Actually in GPT-5.2 for Us? 👇

Not “better answers.”

Not “higher benchmarks.”

What GPT-5.2 really delivers is this:

🧠 Reasoning Is Now an Economic Primitive

Reasoning used to be:

• rare
• expensive
• experimental

GPT-5.2 turns reasoning into infrastructure. When abstract problem-solving costs ~$10, you don’t decide to use it, you assume it’s there.

🤖 Agent Architectures Just Got Simpler

Teams are collapsing:

• fragile multi-agent graphs
• brittle orchestration layers

into single mega-agents with 20+ tools. Why? Because GPT-5.2 can now carry long-horizon state, tool-calling, and execution without falling apart.

This isn’t a UX improvement. It’s an architectural one.

🧑‍💼 Professional Work Is the New Baseline

On GDPval (knowledge work across 44 occupations):

• GPT-5.2 wins or ties humans ~71% of the time
• at >11× speed
• at <1% the cost

Spreadsheets. Presentations. Plans. Models. The question is no longer: “Can AI help professionals?”

It’s:“Which parts of this workflow still justify a human?”

🧪 4. Science Becomes Cheaper to Try

GPT-5.2 Pro helped solve an open problem in statistical learning theory. Humans verified it. External experts reviewed it. The model explored. Humans judged. That’s the new pattern:

🧠 AI = exploration engine

🧑‍🔬 Humans = verification, interpretation, responsibility

This Was a Competitive Release — And It Shows 🥊

Three weeks ago:

• Google dropped Gemini 3
• Topped LMArena at 1501 Elo
• Set records on Humanity’s Last Exam

Inside OpenAI: code red 🚨. Projects shelved. GPT-5.2 fast-tracked. Requests for more polish overruled. This is the first OpenAI launch clearly driven by competition. And on ARC-AGI, they reclaimed the lead.

But again — the win isn’t accuracy. It’s cost.

ARC-AGI Is No Longer a Capability Barrier

It’s a Pricing Problem 💵. The benchmark designed to prove AI couldn’t generalize just became: a line item on a pricing page

390× efficiency improvement in one year.

ARC-AGI-2 already exists. Humans still hit ~100% but models are catching up. But the gap is no longer philosophical. It’s economic.

Final Words 🧠⚙️

If you’re building products in 2025 assuming:

• reasoning is expensive
• abstraction is rare
• generalization is slow

You’re building for a world that no longer exists. GPT-5.2 doesn’t replace humans yet. It forces us to be honest about:

• what humans are uniquely good at
• what is now economically indefensible to do manually
• and how fast agent-first systems will eat brittle workflows

Benchmarks look impressive but the real story is quieter and much harder to reverse.

All the best

Altan