Ask HN: Which open-source coding model comes closest to Claude Opus 4.6?

Developers on Hacker News are debating which open-source coding model holds up against Claude Opus 4.6 — a question with real money behind it for teams running LLM-based coding agents in production.

The models coming up most: DeepSeek-V3 and DeepSeek-R1, Alibaba's Qwen2.5-Coder-72B-Instruct, and Mistral's Codestral. DeepSeek has drawn the most enthusiasm. Its open weights and competitive scores on HumanEval, SWE-Bench Verified, and LiveCodeBench have convinced a chunk of the community that near-frontier coding performance is achievable without frontier API bills. Qwen2.5-Coder gets specific credit for multi-file reasoning — the kind of repository-level context-tracking that matters when you're hooking a model into Aider, OpenHands, or SWE-agent and asking it to navigate a real codebase.

The thread also surfaces the usual gap between leaderboard scores and day-to-day use. Practitioners keep noting that inference latency, context window limits, and whether something will actually run on hardware you own shape real-world value more than any single benchmark. Older models — Meta's CodeLlama successors, BigCode's StarCoder2, IBM's Granite Code — come up as alternatives, with Granite favored in enterprise environments where licensing terms matter. None of them top the current recommendations, but they remain in the conversation.

The pattern here is one the agent ecosystem has been watching for two years: the gap between the best open-weight models and the best closed ones keeps getting smaller. Some teams are already using open-source models to anchor full agent pipelines. That compression puts real pricing and roadmap pressure on Anthropic and its peers, and makes the model selection decision for a new coding project meaningfully harder than it was even six months ago.