- Authors
- Name
- 오늘의 바이브
20 Minutes Apart
February 5th, 6:40 PM. Anthropic announces Claude Opus 4.6. Twenty minutes later at 7 PM, OpenAI announces GPT-5.3 Codex.
Not a coincidence. Both companies had originally planned to announce at the same time. Anthropic moved 15 minutes early. OpenAI stuck to the schedule.
Two days later is the Super Bowl. Anthropic has two ad slots booked during the game. The content: mocking OpenAI.
What the Benchmarks Say
| Benchmark | Opus 4.6 | GPT-5.3 Codex | Winner |
|---|---|---|---|
| SWE-Bench Verified | 80.8% | 56.8% | Opus |
| Terminal-Bench 2.0 | 65.4% | 77.3% | GPT |
| GDPval-AA (Elo) | 1606 | 1462 | Opus |
| Context Window | 1M tokens | 256K | Opus |
| Speed | Baseline | +25% | GPT |
Opus leads in bug fixing (SWE-Bench) and expert-level tasks (GDPval-AA). GPT leads in terminal-based coding (Terminal-Bench) and raw speed.
Both claim they won.
What the Benchmarks Don't Say
The problem is that the benchmark versions differ.
- SWE-Bench has multiple versions. There is no way to confirm that the "Verified" version where Opus scored 80.8% and the version where GPT scored 56.8% are the same test.
- Terminal-Bench 2.0 is a benchmark built by OpenAI itself. It is not surprising that Anthropic scored low on this test.
Each company puts the benchmark that favors them front and center. Anthropic says "we lead by 144 Elo on GDPval-AA." OpenAI says "we demolished Opus on Terminal-Bench."
Same data. Different narratives.
The Philosophical Split
The numbers matter less than the design philosophy.
Opus 4.6 -- "Handle it"
- Built for autonomous agents
- Thinks deeply, runs long, asks fewer questions
- Agent Teams: multiple agents split the work
- 1 million token context to understand an entire codebase
GPT-5.3 Codex -- "Let's do it together"
- Built for collaboration
- Humans intervene mid-run and redirect
- 25% faster for tight feedback loops
- The first model that debugged its own training
Opus takes the wheel once you point the direction. GPT keeps the conversation going the whole way.
Which one is right depends on what you want. Keep control? GPT. Delegate? Opus.
The Super Bowl Ad War
Anthropic's Super Bowl ad content leaked.
The ad satirizes a chatbot that befriends a user, then suddenly starts pushing products. An awkwardly enthusiastic bot gushes "This product is amazing!" The closing title card reads:
"Ads are coming to AI. But not to Claude."
(Dr. Dre "What's the Difference" intro)
A direct shot at OpenAI's decision to put ads in the free and low-cost tiers of ChatGPT.
Sam Altman's response (X)
"It's funny. But dishonest. Anthropic sells expensive products to rich people."
Greg Brockman (OpenAI co-founder)
"Is Anthropic actually promising they'll never sell user data to advertisers?"
Dario Amodei (Anthropic CEO)
Silence.
Convergence
Here is the interesting part: the two models are becoming more alike.
Opus 4.6 added Agent Teams. Multiple agents collaborating on a task. GPT-5.3 increased its autonomy. It debugged its own training process.
Both started as coding agents. Both are converging toward general-purpose work agents. Right now they compete on coding. In six months they will compete on spreadsheets, presentations, and email.
"Which model is better" is becoming a meaningless question. They are heading in the same direction.
What to Watch on Sunday
When the Super Bowl ad airs, public opinion shifts. Not the tech community -- the general public. The message "ChatGPT is getting ads" will reach tens of millions of people.
If Anthropic wins this fight, Claude's premium image gets stronger. If OpenAI's counterattack lands, the "ads = free = democratization" frame takes hold.
This ad war might have a bigger impact than any benchmark number.
Sources: