大模型网摘Picking models by scenario (November 2026): coding, writing, self-hosting and agents
November's model market moved in three places: OpenAI split its lineup by scenario (GPT-6 Cyber for security work, GPT-6.1 Terra for mid-tier and computer use); Kimi K3.1 Pro pushed open weights to the 3T class with 1M context; and on the DeepSeek side, deepseek-flash (V4.1 Flash) became the de facto workhorse, with V4 Pro officially scheduled for orderly retirement and requests temporarily routed to V4.1 Flash billing. Here is the current list, organized by four scenarios.
Coding
On the subscription side, Claude Code paired with Claude Opus 5.5 ranks near the top of the coding category on several leaderboards; budget-sensitive teams keep calculating between subscriptions and metered APIs, and our earlier Kimi Code subscription vs DeepSeek API comparison walks through the math. On the metered side, deepseek-flash and Kimi K3.1 Pro are the two open-weight anchors, while GLM-5.3 holds a steady reputation for agentic coding. The old advice stands: benchmark on your own repository first, then consult the leaderboards.
Writing and long documents
1M-context windows are now shared equipment for Kimi K3.1 Pro and DeepSeek V4.1 Flash; tasks that read whole books or whole codebases can be picked by window size. V4.1 Flash ships native vision input, so mixed image-and-text material can be processed in one pass without a separate OCR step. For working methods, see our 1M context guide.
Self-hosting
The entry bars differ widely: full V4.1-Flash is 510GB of original weights, and the community project Promethevs can run it with 64GB of RAM plus 12GB of VRAM, at the cost of 5 to 8 tokens/s; server-grade deployments can follow the official vLLM recipe on multi-GPU machines; the official reference bar for large-scale deployment is a 2,000-GPU cluster with storage. For the full on-premises route where data never leaves the building, see DeepInfer self-hosting scenarios.
Agents and computer use
GPT-6.1 Terra sits at the mid tier with vision and computer-use capabilities, suited to desktop automation; GPT-6 Cyber focuses on authorized security work and remains access-restricted. On the open side, DeepSeek Harness and dsh-superagent offer self-hostable agent frameworks. Put batch runs in off-peak hours, when prices are half.
Closing notes
Prices change frequently on all sides; treat official pricing pages as the source of record. This piece maps scenarios to version numbers rather than issuing verdicts: test on your own workload. For version and availability context, see our November lineup overview and the GPT-6.1 Terra pricing quick reference.