GPT-5.2 Climbs to #5: Better, But Not Enough

From #29 to #5: What changed

GPT-5.2 has been on LMArena since its December 11, 2025 launch. The base model sat at #29 with 1438 Elo. GPT-5.2 High — the version with more compute — was marginally better at #28 with 1439 Elo. Neither was particularly impressive for a flagship model.

Then on February 10th, OpenAI quietly pushed a new checkpoint:gpt-5.2-chat-latest-20260210. This updated version scored 1478 Elo — a 40-point improvement over the base model. That's a massive jump. For reference, 40 Elo points is roughly the gap between Gemini 3 Pro and Claude Opus 4.6.

This is how modern AI development works. The December launch was the foundation. The February checkpoint represents months of post-training refinement — RLHF tuning, instruction following improvements, and response quality polish. OpenAI didn't release a new model. They made their existing model significantly better through iterative updates.

Better isn't best

Here's the uncomfortable truth for OpenAI. A 40-point improvement is genuinely impressive engineering. But it only gets them to #5. The gap to #1 is still 27 Elo points — roughly a 46% win rate heads-up against Claude Opus 4.6. And two Google models sit between them.

This matters because of who OpenAI used to be. GPT-4 launched at #1 on March 13, 2023 with no real competition. Across GPT-4, GPT-4o, o1, and GPT-4.5 Preview, OpenAI held the top position for 651 of the first 1,077 days of the modern AI race — 60% of the time. They weren't just in the race. They were the race.

Now their best model, after significant iteration, sits behind four others. The base GPT-5.2 was 67 points behind the leader. The new checkpoint cut that to 27. Progress, absolutely. But the days of OpenAI holding a comfortable lead are over.

The code leaderboard tells the same story

On the Code leaderboard, OpenAI's best showing is GPT-5.2 High at #5 with 1471 Elo — 90 points behind Claude Opus 4.6's 1561. The base GPT-5.2 is at #17 with 1395. Even GPT-5.2 Codex, the coding-specialized variant, sits at #27 with 1336.

Fifth on text. Fifth on code. Anthropic holds the top two spots on text and all four of the top positions on code. Google holds #3 and #4 on text. OpenAI is consistently good, but consistently behind.

The checkpoint strategy

What's worth noting is OpenAI's approach. Rather than waiting for GPT-5.3 to close the gap, they're pushing checkpoint updates to GPT-5.2 that deliver meaningful improvements. The February 10th checkpoint gained more Elo than some entire model generations.

This mirrors what Anthropic did with the Opus 4.5 → 4.6 transition and what Google did with Gemini 3 → 3.1 Pro. The industry is moving toward continuous improvement over discrete launches. The model you use today is measurably better than the one released under the same name three months ago.

For developers, this is actually good news. GPT-5.2 at 1478 Elo is a genuinely strong model — better than everything outside the top 4. If you're locked into OpenAI's ecosystem, you're getting competitive performance that's improving without any migration effort.

What comes next

Check the timeline on WhoLeads.AI and you'll see the pattern: the AI race has become a three-way relay where Anthropic, Google, and OpenAI trade momentum every few months. OpenAI dominated 2023-2024. Google had strong showings in early 2024 and mid-2025. Anthropic has led since late October.

GPT-5.2's 40-point improvement shows OpenAI isn't standing still. But the question isn't whether they're improving — it's whether they can improve faster than the two companies ahead of them. Right now, the answer is: not quite.

GPT-5.3 is presumably in the works. When it arrives, it'll tell us whether OpenAI can reclaim the top spot or whether #5 is the new normal for the company that started it all.