Anthropic Says Claude Now Writes 80% of Its Code. The Speedup Chart Is the Real Story.

Anthropic published internal data claiming recursive self-improvement is within view. The benchmark jumps are striking — but they reframe how builders should think about model upgrades this quarter.

Anthropic spent June 4, 2026 doing something vendors rarely do: publishing the internal benchmarks they use to decide whether their own engineers should keep writing code. The numbers are loud — over 80% of code merged into Anthropic's codebase is now written by Claude, per Alex Albert's post — but the more interesting signal for working developers is what the company is measuring, not what it's claiming.

What Anthropic actually published

The main thread from Anthropic walks through three internal evaluations. On a code-speedup test where a skilled human takes 4-8 hours to make a small training script 4x faster, Claude Opus 4 averaged a ~3x speedup as of May 2025 (Anthropic corrected the original date from 2024). The Mythos Preview model hit ~52x on the same task in April 2026. On open-ended coding problems with unclear answers, Claude's success rate is now 76% — a 50-point jump in six months. Anthropic engineers ship 8x more code per quarter than the 2021-2025 baseline.

The framing is recursive self-improvement, but the practical read is narrower: Anthropic is publicly anchoring the next model generation to coding-agent performance, not chat quality. That's a planning signal for anyone choosing a stack right now.

What it means alongside the rest of the week

OpenAI spent the same week on a different vector. The new ChatGPT memory system rolled out to Plus and Pro users in the US on June 4, 2026, automatically tracking context across conversations with 2x more memory than the previous system. Separately, OpenAI confirmed an incident on June 5 that incorrectly suspended user accounts — a reminder that consumer-tier reliability still bites enterprise pilots. And GPT-5.2 and GPT-5.3-Codex were sunset in Codex for ChatGPT subscribers earlier in the week, with users in the developer forum noting they preferred 5.3-Codex over the newer generic models.

The contrast matters. Anthropic is publishing coding benchmarks. OpenAI is shipping memory features and pruning coding-specific SKUs. If your team's AI budget is going toward agentic coding tools — and Simon Willison noted that Brex now allows $2,000/month per engineer for these tools, with Uber at $1,500 — the divergence in vendor focus is going to show up in your stack choices.

Implications for builders

For engineering leads: the 76% success rate on open-ended coding tasks is the number to test against your own evals before the next procurement cycle. For PMs evaluating vendor lock-in: Anthropic's public benchmarks are auditable in a way that vendor demos aren't, which makes them useful as a negotiating anchor. For anyone running a Codex-based workflow: the sunset of 5.3-Codex is a reminder that consumer-tier model availability is not a stable foundation for production tooling.

The bottom line: Stop reading vendor threads as marketing and start reading them as roadmap leaks — Anthropic just told you what next quarter's models will be optimized for, and OpenAI just told you what they won't.

Sources