feat(site): a benchmarks page, the launch table, the GA delta, and the kill line

[lroolle]

A new benchmarks page (/bench/) for the site.

• Launch table: V4-Pro 0813 and V4-Flash vs Kimi K3, GLM-5.2, Opus 4.8, Fable 5 on Terminal-Bench 2.1, DeepSWE, Toolathlon, CyberGym, NL2Repo, AutomationBench, DSBench, Agents' Last Exam, HLE — DeepSeek's own launch chart, transcribed from the research corpus.
• Caveats up front: vendor numbers / one harness, no GPT-5.6 column, Kimi+GLM single-sourced, NIST/CAISI put the preview ~8 months behind frontier.
• GA delta: preview→0813 jumps (DeepSWE 12.8→62.7, etc.), read honestly as post-training not a new base model.
• The 斩杀线 (kill line): the concept, the real price gap to GPT-5.6 Sol (~11.5x/34.5x/138x), and the sober version — it kills the middle, not the frontier.
• Practice: route by role, cache-aware prompting, measure successful-task cost. Ties to ds usage and the cost page.

Every figure traces to the official chart or a published rate card. Joins the nav after cost, plus sitemap + llms.txt. Verified in CloakBrowser at 1280/390 — table scrolls within .tablewrap, no document overflow. make check green.

🤖 Generated with Claude Code