Top AI Product

DeepSeek pushed its flagship out of preview on August 13 and hit number one on HackerNews the same day — 799 points, 309 comments. It’s a 1.6T-parameter MoE with ~49B active, hybrid attention to keep long-context inference cheap, 32T+ pretraining tokens, 1M context and 384K max output.

The numbers that got it to HN #1

Terminal Bench 2.1 went from 72.1 to 87.9 over the preview. CyberGym 52.7 → 83.3. DeepSWE 12.8 → 62.7. On DeepSeek’s own agent and coding benchmarks it edges past Opus 4.8, and SWE-bench Verified lands at 80.6 — a hair behind Claude. These are vendor-reported, so treat them as a direction, not a ranking. Independent frontend and 3D work still looks weaker than Opus.

What you actually plug it into

It’s a model, not an app: an API you point your agent loop at. DeepSeek’s official endpoint keeps the same deepseek-v4-pro model id, so old integrations inherit the new weights with zero code change. OpenAI-compatible format, also on OpenRouter. $0.435/M in, $0.87/M out, $0.003625/M cached input — roughly an order of magnitude under Opus for long-horizon terminal agents, SWE-bench-style repo work, and million-token codebase reads.

DeepSeek has already warned a big price hike is coming.

You Might Also Like

• Deepseek v4 Flash 0731 Hits 82 7 on Terminal Bench Chasing Opus 4 8 at 0 14 m Tokens
• Grok 4 5 Spacexai Solves swe Bench pro Tasks With 4 2x Fewer Tokens Than Opus 4 8
• Kimi k2 6 Beats gpt 5 4 and Claude Opus 4 6 on swe Bench pro
• Deepclaude Lets Claude Code run on Deepseek v4 pro 0 87 vs 15 per Million Tokens
• Herdr Puts Every Coding Agent in one Terminal and Lets Them Orchestrate Each Other