This AI Ended Software Engineering (again)?

Recent AI Model Releases and Competition
📌 Anthropic released Opus 4.6 around the same time OpenAI dropped its GPT Codeex model, leading to immediate competition on benchmarks.
📢 Anthropic ran a Super Bowl ad indirectly criticizing OpenAI for planning to introduce ads in ChatGPT, affirming their commitment not to run ads in Claude.
📊 Sam Altman highlighted that the usage of ChatGPT in Texas alone surpassed the total number of Claude users in the US, referencing the sustainability of an ad-supported model.

Anthropic Opus 4.6 Features and Benchmarks
🔍 Benchmarks suggest Opus 4.6 is roughly the same as Opus 4.5 in coding tasks, though tool use capability has improved.
🧠 The most significant upgrade in Opus 4.6 is the 1 million context window size, drastically improving its ability to handle complex tasks requiring extensive context retention.
💰 Pricing for Opus 4.6 remains the same as 4.5, but US-only inference is available at 1.x price for compliance-sensitive workloads.

OpenAI GPT Codeex and Model Strategy
💻 OpenAI is positioning the Codeex app (CLI, Web) as the future standard for software development, with GPT-5.3 Codeex being a model specifically tailored for this platform.
⚡ GPT-5.3 Codeex is reportedly 25% faster due to changes in inference tech, potentially linked to their partnership/acquisition of Cerebras.
🎮 GPT-5.3 Codeex demonstrated capability in autonomously iterating on games over millions of tokens, though some supervision was noted (e.g., "fix the bug").

Autonomous Software Development via Agent Teams (Anthropic)
🧑‍💻 Anthropic introduced Agent Teams in Claude Code, a feature where multiple agents share a task list, self-claim work, and communicate, distinct from the master-slave model of sub-agents.
⚙️ This orchestration layer was used by 16 agents to build a Rust-based C compiler from scratch, capable of compiling the Linux kernel (Linux 6.9).
📉 The resulting compiler, while impressive, is highly inefficient, outputting code worse than the lowest optimization setting of GCC, and it lacks essential components like its own assembler and linker.

Key Points & Insights
➡️ The 1 million context window in Opus 4.6 is a massive unblocker for complex tasks, although single-agent context limits remain a challenge for massive codebases.
➡️ The capability of AI to build a C compiler highlights that current performance gains (like Opus 4.6) are not 10x improvements over human experts in specialized fields like compiler design.
➡️ Developers must maintain optimistic attitude and continuous growth; relevance depends on improving at orchestration (like using Agent Teams) faster than AI models advance in raw intelligence.
➡️ The perceived exponential growth in LLM capabilities might be a sigmoid function's initial curve; stagnation or limitations in compute/physics could cap development speed.

📸 Video summarized with SummaryTube.com on Feb 07, 2026, 15:20 UTC