Coding — AI That Writes Production Code

The agentic coding model that doesn't just autocomplete — it plans, tools up, debugs across files, and finishes the messy repo task while you walk the dog. Terminal-Bench 82.7% isn't a typo.

Terminal-Bench 2.0 82.7% (crushes Opus 4.7's 69.4%); Expert-SWE 73.1% on 20-hour human tasks; FrontierMath Tier 4 35.4%; ~40% fewer output tokens; 1M context with native tool use and Codex integration.

2× API price ($5/$30 per 1M tokens); trails Claude Opus 4.7 on SWE-Bench Pro (58.6% vs 64.3%); API not live at launch; early hallucination reports need verification.

Anthropic's hybrid reasoning monster — the model that doesn't just write code, it *engineers* it. SWE-Bench Pro 64.3% obliterates every other model on the hardest real-world coding benchmark. CursorBench 70%. High-res vision that reads your screenshots. And an 'xhigh' effort mode that lets it think harder than any model before it. This isn't an incremental update — it's a category break.

SWE-Bench Pro 64.3% (new SOTA — beats GPT-5.4's 57.7% and Kimi K2.6's 58.6% by a chasm). CursorBench 70% in real IDE sessions. OSWorld 78%. High-res vision up to 3.75 MP for screenshots and diagrams. Same $5/$25 pricing as Opus 4.6. Available everywhere: Claude.ai, API, Bedrock, Vertex, GitHub Copilot.

Not all sugar and roses. Token usage is noticeably higher (new tokenizer inflates costs 15–35% on code-heavy prompts). Adaptive reasoning makes it feel 'lazier' on simple prompts unless you force high effort. Some users report regressions in long-context recall beyond 100K tokens. This is a specialist — brilliant at hard coding, occasionally frustrating on easy tasks.

Alibaba's agentic coding flagship — purpose-built for the kind of coding tasks that take hours, not minutes. Qwen 3.7 Max ran a 35-hour kernel optimization session with 1,158 tool calls and zero human intervention. SWE-Bench Pro 60.6%, a 1M-token context window, and cross-harness compatibility that lets it slot into Claude Code or any standard agent framework out of the box.

SWE-Bench Pro 60.6%, Terminal-Bench 2.0 Terminus 69.7, Code Arena WebDev ~1541 Elo (top 4). The first Chinese proprietary model to consistently match Western frontier models on production coding benchmarks. 210+ output tokens/sec makes it one of the fastest frontier models available.

API-only with no open weights (yet). Heavy agent sessions get expensive fast — one user reported $43 in 15 minutes of autonomous coding. Independent evaluations show more variance than official benchmarks suggest. Not the strongest option for pure UI/design work.