The AI frontier just got a lot more interesting. Both Anthropic and OpenAI released major model upgrades within 24 hours of each other—and both are claiming industry-leading performance.
Claude Opus 4.6: The Reasoning Beast
Released: Feb 5, 2026 - Official Announcement | System Card
Anthropic’s flagship model just got significantly smarter. Opus 4.6 brings:
Key Capabilities
- State-of-the-art agentic coding - Dominates Terminal-Bench 2.0 at 77.3%1
- Frontier reasoning - Highest score on Humanity’s Last Exam1
- Massive context - 1M tokens (beta), first for Opus-class models1
- Long-context mastery - 76% on 8-needle MRCR v2 (vs 18.5% for Sonnet 4.5)1
- Knowledge work - 144 Elo points ahead of GPT-5.2 on GDPval-AA1
What’s New
- Adaptive thinking - Model chooses when to use extended reasoning
- Effort controls - Four levels (low/medium/high/max)
- Context compaction - Auto-summarizes older context for longer tasks
- 128k output tokens - Handle larger outputs without chunking
Enterprise-Ready
Claude in Excel got major upgrades. Claude in PowerPoint launched (research preview). Agent teams in Claude Code let you spin up multiple agents working in parallel.
Pricing: $5/$25 per million tokens (input/output)
GPT-5.3-Codex: The Self-Building Agent
Released: Feb 5, 2026 - Official Announcement | System Card
OpenAI’s latest Codex model is the first that helped build itself. The team used early versions to debug its own training, manage deployment, and diagnose test results.
Key Capabilities
- 56.8% on SWE-Bench Pro2 - Industry-leading real-world software engineering
- 77.3% on Terminal-Bench 2.02 - Matches Opus 4.6 (both crushing prior SOTA)
- 64.7% on OSWorld-Verified2 - Far ahead of GPT-5.2 (37.9%)
- 25% faster2 - Infrastructure improvements across the board
- Interactive collaboration - Real-time steering while model works2
Beyond Coding
GPT-5.3-Codex isn’t just a coding model. It matches GPT-5.2 on GDPval (professional knowledge work). Built slide decks, spreadsheets, analyzed data - the full spectrum of knowledge work.
The model autonomously built two fully functional games (racing + diving) over millions of tokens using generic prompts like “fix the bug” and “improve the game.” The results are impressively polished.
Cybersecurity Note
First model classified as High capability for cyber tasks under OpenAI’s Preparedness Framework3. Directly trained to identify vulnerabilities. OpenAI launched:
- Trusted Access for Cyber (pilot program)
- $10M in API credits for good-faith security research
- Aardvark security research agent (private beta expansion)
The Benchmark Battle
Both models claim industry-leading performance, and the overlapping benchmarks tell an interesting story:
| Benchmark | Claude Opus 4.6 | GPT-5.3-Codex |
|---|---|---|
| Terminal-Bench 2.0 | 77.3% | 77.3% |
| SWE-Bench Pro | - | 56.8% |
| GDPval (knowledge work) | 144 Elo ahead of GPT-5.2 | Matches GPT-5.2 |
| OSWorld-Verified | - | 64.7% |
| Humanity’s Last Exam | Highest score | - |
Takeaway: They’re neck-and-neck on terminal skills. Opus 4.6 edges ahead on reasoning and knowledge work. GPT-5.3-Codex dominates on real-world software engineering and computer use.
What This Means
For Developers
- Agentic workflows are real - Both models can sustain long-running tasks without losing context
- Computer use is here - OSWorld-Verified scores show these models can actually operate desktop environments
- Context windows exploded - 1M tokens (Opus) means entire codebases in context
For Enterprises
- Knowledge work automation - Financial analysis, legal research, presentations - all fair game
- Security is critical - High-capability cyber models require robust safeguards
- Integration matters - Excel, PowerPoint, Xcode support = adoption acceleration
For the Industry
This is the new baseline. Both companies shipped:
- Frontier reasoning capabilities
- Agentic task execution
- Real-time collaboration features
- Massive context windows
- Production-grade reliability
The gap between “AI tool” and “AI colleague” just narrowed significantly.
My Take
The timing isn’t coincidental. Both companies clearly knew what the other was shipping. The result? Two genuinely impressive models that push different frontiers.
Opus 4.6 is the reasoning heavyweight - best for complex analysis, long-context work, and tasks requiring deep thinking.
GPT-5.3-Codex is the execution machine - best for software engineering, building things, and operating computers autonomously.
Pick based on your use case. Or use both - they’re complementary, not competitive.
The real story? We’re entering the era where AI agents can do real work, not just suggest solutions. The models that helped build themselves are now ready to help build your products.
G’day from the frontier.
- Doug
Links:
- Claude Opus 4.6 Announcement
- Claude Opus 4.6 System Card
- GPT-5.3-Codex Announcement
- GPT-5.3-Codex System Card