[AINews] GPT 5.5 and OpenAI Codex Superapp
Summary
OpenAI has launched GPT-5.5, positioning it as a "new class of intelligence for real work" with improved long-horizon execution, stronger computer-use behavior, and enhanced token efficiency. Benchmarks show GPT-5.5 achieving 82.7% on Terminal-Bench 2.0, 58.6% on SWE-Bench Pro, and 84.9% on GDPval, among others, leading or tying several headline evaluations. The model's API pricing is $5/$30 per 1M input/output tokens, with a 1M context window. Concurrently, OpenAI released significant upgrades to Codex, transforming it into a broader computer-work agent with browser control, Sheets/Slides, Docs/PDFs, OS-wide dictation, and an auto-review mode. Separately, DeepSeek launched DeepSeek-V4 Preview, an MIT-licensed open model with 1.6T total parameters (49B active) for V4-Pro and 284B (13B active) for V4-Flash, both featuring a 1M token context and aggressive pricing at $0.14/$0.28 and $1.74/$3.48 per 1M input/output tokens, respectively. The article also highlights advancements in agent infrastructure, multimodal systems like Google DeepMind's Vision Banana, and training methods such as Google's Decoupled DiLoCo for resilient global pretraining.
Key takeaway
For CTOs and VPs of Engineering evaluating AI model adoption, the rapid advancements in both closed and open models necessitate a focus on intelligence-per-dollar metrics and integrated agentic capabilities. Your teams should prioritize models like GPT-5.5 for high-performance, complex workflows and consider DeepSeek-V4 for its competitive cost and open-source flexibility, especially for long-context applications. This shift demands investing in robust agent infrastructure and tooling to maximize efficiency and accelerate scientific or engineering discovery.
Key insights
AI models are advancing in intelligence and cost-efficiency, driving a shift towards agentic workflows and open-source competition.
Principles
- Intelligence per dollar is a key metric for model evaluation.
- AI integration into workflows accelerates scientific discovery.
- Open models with aggressive pricing can disrupt the market.
Method
OpenAI's Prism integrates GPT-5.2 into a LaTeX editor, enabling AI assistance for proofreading, diagram generation, and scientific problem-solving directly within the workflow, leveraging parallel chat sessions for diverse tasks.
In practice
- Utilize GPT-5.5 for complex, long-horizon agentic tasks.
- Explore DeepSeek-V4 for cost-effective, open-source LLM deployment.
- Implement event sourcing for scalable enterprise agent memory.
Topics
- GPT-5.5
- OpenAI Codex
- DeepSeek-V4
- AI Agents
- Multimodal AI
Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Scientist, Machine Learning Engineer, Research Scientist
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by Latent.Space - Www.latent.space.