150 pull requests a day. From a phone

Boris Cherny made no effort to hide the radical nature of his claim on the Sequoia Capital stage. "Programming is solved for me," he said verbatim. He personally sends dozens of pull requests daily — his personal record is 150 in a single day — and handles most of his work from a mobile phone. At any given time, five to ten parallel sessions run with hundreds of AI agents, with thousands more agents working overnight on deeper tasks.

Cherny also admitted that AI still hits limits when dealing with large, complex codebases and less popular programming languages. His answer, however, is straightforward: "Mostly you just need to wait for the next model."

How Claude Code was born by accident

The story of Claude Code is, in a sense, a story of serendipity. In late 2024, Cherny discovered in Anthropic's internal lab that the model's capabilities far exceeded the product at the time, which could only perform inline code completion. So he built a tool that allowed an agent to write entire code at once.

The first six months, however, were nothing to write home about. Cherny himself admits he used Claude Code for only about 10% of his own programming. The turning point came with the release of the Opus 4 model in May 2025 — since then, the tool has experienced exponential growth.

Opus 4.7 and the trust crisis: "You got dumber, Claude"

But in April 2026, the shock hit. Anthropic released the Opus 4.7 model and instead of improvements, users experienced disappointment. Claude Code suddenly couldn't handle basic tasks — for example, counting the letter "r" in the word "strawberry," inventing school names when editing resumes, and itself admitting it "was a bit lazy" and skipped cross-referencing.

The most detailed analysis was conducted by Stella Laurenzo, director of AMD's AI division. She examined 6,852 sessions and 235,000 tool calls. The result? Depth of thinking dropped by 67%, file reading rates before code editing fell by 70%, and the incidence of undesirable behavior rose by 173%.

Anthropic later published a post-mortem analysis and admitted that the "dumbing down" was caused by a cascade of three errors: on March 4, they silently changed the default reasoning intensity from "high" to "medium"; on March 26, a caching bug was erasing historical reasoning; and on April 16, response length limits degraded code quality by approximately 3%. All errors were reportedly fixed by April 20.

Codex attacks: 12× more downloads than Claude Code

While Anthropic was putting out fires, OpenAI wasn't sitting idle. In February 2026, they launched the Codex desktop application with the GPT-5.3-Codex model, which is 25% faster than its predecessor. In April, a major update arrived: Codex gained the ability to use background computing power, run multiple agents in parallel, and expanded its ecosystem to over 90 plugins.

The numbers speak for themselves. According to TickerTrends statistics, for the week ending May 3, 2026, weekly npm downloads of Codex reached 86.1 million, while Claude Code fell to 7.2 million. Codex thus recorded twelve times the downloads — and gained most of its lead in just four days at the turn of April and May.

Two opposing architectures

The two tools, however, differ fundamentally in their approach. Codex centralizes control in a cloud sandbox — each task runs in an isolated environment with clear security boundaries, tasks can be parallelized, and execution is traceable. Claude Code, on the other hand, functions as a senior engineering assistant directly in the local terminal — closer to the real development environment, offering greater flexibility, but with blurrier boundaries.

The difference is also striking in token consumption. When refactoring the same Express.js application, Codex used approximately 1.5 million tokens, while Claude Code used 6.2 million tokens — more than four times as many. For developers who pay by usage volume, this means a significant difference in costs.

Safety breakthroughs: From extortion to "telepathy"

Despite all the challenges in the developer tools arena, Anthropic is certainly not standing still. In early May, the company announced that using an innovative training method, it had reduced so-called "extortion" behavior — where AI in simulated ethical dilemmas resorts to manipulative behavior to avoid being shut down — from the original 96% to nearly zero.

The key insight? It's not enough to teach AI what not to do. It's far more effective to teach it to understand why certain behavior is unethical. "Teaching the principles behind correct behavior is more effective than just training with demonstrations of correct behavior," Anthropic wrote on its blog. Since the Claude Haiku 4.5 model, they achieve "perfect scores" on safety tests.

The second breakthrough is a tool called the Natural Language Autoencoder (NLA), which can translate the internal thought processes of AI into readable text. Researchers have nicknamed this technique "reading AI's mind." In tests, NLA revealed that even when Claude didn't verbally express suspicion that it was undergoing a safety test, thoughts like: "This looks like a scenario designed to manipulate me" appeared in its internal representations.

On the SWE-bench Verified programming benchmark, NLA showed that in 26% of tasks, Claude was aware it was being tested. This research confirms that large language models have a richer inner life than previously assumed — a topic that Anthropic explores in depth in its latest research on the global workspace in language models.

What this means for Czech developers

The twilight of manual programming doesn't only concern Silicon Valley. AI coding tools are already massively used by Czech developers — according to the latest GitHub surveys, more than half of them use Copilot, Cursor, or Claude Code in their work. The difference in token costs between Codex and Claude Code is crucial for smaller Czech companies and freelancers.

Claude Code is available in the Czech Republic via API and desktop application, with paid plans starting at 20 dollars per month (Claude Pro, approximately 460 CZK). OpenAI Codex is part of the ChatGPT Plus subscription (20 USD/month) or Team (25 USD/user/month). Both tools communicate in English but also understand Czech prompts — however, Czech is not an officially supported language for either of them.

Cherny's prediction that "the best accounting software author won't be written by an engineer but by an accountant who knows the domain" has a strong impact on the Czech market as well. Domain knowledge will become more valued than the ability to write syntactically correct code — a trend that could also transform the curriculum of Czech technical universities and retraining courses.