Coding — AI That Writes Production Code

A generalist powerhouse that codes like a specialist — handling multi-file edits and long-horizon agents without the bloat. The decathlete who also holds the 100m record.

SWE-Bench Pro 57.7% (edges Codex's 56.8%); 1M context for massive repos; native tool-use cuts tokens 47%; 1.5x faster in Codex; GPQA Diamond 92.8% for reasoning-heavy code.

Higher API costs ($2.50/M in, $15/M out); Pro needed for peak performance; cyber blocks on sensitive prompts; 1M context counted at 2x rate in Codex.

The model that thinks before it codes. Opus 4.6 plans multi-step refactors, sustains context across sprawling codebases, and writes production code that reads like a senior engineer reviewed it — because, in a way, one did.

Anthropic's most capable model. 1M-token context window (beta) lets it hold entire repos in working memory. Top marks on agentic coding benchmarks — it plans, executes, and self-corrects across long tasks.

The most expensive model in its class. Long agentic sessions can amplify cost if you don't supervise — and it's slower than lighter models for quick questions.

The first open-weight model to hold the #1 spot on SWE-Bench Pro — and it's MIT licensed. GLM-5.1 doesn't just write code; it runs 8-hour autonomous engineering sessions with 655+ iterations, self-correcting across thousands of tool calls. The open-source answer to closed-model coding dominance.

SWE-Bench Pro SOTA at 58.4 — beating Claude Opus 4.6 (57.3) and GPT-5.4 (57.7). CyberGym 68.7 surpassing all closed models. 200K context window with 128K+ output length. Fully open weights under MIT license.

Text-only — no vision or multimodal input. ~754B total parameters means serious GPU requirements even with 40B active MoE. Western ecosystem tooling still less mature than Chinese-language resources.