大模型网摘Flagship LLM Lineup, November 2026: Benchmarks and Pricing at a Glance
Who Holds the Flagship Spot at Each Lab
As of November, the "current lineup" across major labs has rotated several times. On DeepSeek's side, the API exposes two model names, deepseek-flash and deepseek-v4-pro: the former maps to V4.1-Flash-1017, launched October 17 with optimizations for coding and agent tasks; its Search and Fetch endpoints opened earlier and entered billing in November. Context window is 1M tokens with a 384K output cap. The next-generation V5 remains a supply-chain rumor, unconfirmed by the company.
OpenAI's GPT-6 family is split by purpose: Astra is the general flagship, Sol and Luna cover lightweight tiers, and GPT-6.1 Terra focuses on visual reasoning and computer use. On November 5 the company also released GPT-6 Cyber, a model for authorized security work, initially available only through the restricted Daybreak Red channel. Public reporting cites an ExploitGym score of 48.5% (versus 32.1% for the previous GPT-5.6 Cyber), and its security risk rating moved to Critical in the company's Preparedness Framework. See our earlier GPT-6.1 Terra coverage for the lineup context.
In the open-weight camp, Moonshot AI's Kimi K3.1 Pro, released October 23, is a 3T-scale open-weight model with a 1M-token context window; full weights were promised before November 4. Zhipu's GLM-5.3 follows a post-training upgrade path on 753 billion parameters — details in our model deep-dive — and Alibaba's Qwen3-Max also appears in the open-weight flagship spec table. Our October spec comparison lays out all four side by side.
Benchmarks: Three Labs, One Test
On Terminal-Bench 4.0 (coding-agent work in a real terminal), a benchmark cited by both official posts and public reporting: Claude Opus 5.5 scores 66.4%, Kimi K3.1 Pro 65.7%, and GPT-6 Astra 58.2% — the first two sit on nearly the same line. Note that part of K3.1 Pro's results are self-reported; third-party verification is still pending, and cross-vendor benchmarks should be read as rough references only.
Pricing: Off-Peak and Caching Move the Needle
On API pricing, DeepSeek's official peak/valley scheme since September: for deepseek-flash, per million tokens, cache-miss input is $0.15 off-peak and $0.30 at peak; output is $0.60 off-peak and $1.20 at peak; cache hits drop to about $0.003. The full table is in our pricing update explainer. By contrast, pricing for newer models such as GPT-6 Cyber has not been published, and flagship tiers at Claude and OpenAI remain well above the leading domestic offerings. For a traffic-side view, see our OpenRouter November token-share note.
How to Choose by Scenario
For cost-sensitive batch work with long context, deepseek-flash combined with caching and off-peak calls is among the leaner combinations on paper. For open-weight self-hosting, K3.1 Pro, GLM-5.3, and V4-Pro are first-tier candidates. Among cloud-only flagships, Opus 5.5 and K3.1 Pro lead the coding-agent pack, while visual-reasoning and computer-use tasks point to GPT-6.1 Terra. All figures reflect public information as of publication; verify current versions and prices against official docs before committing. This article is compiled from publicly available materials.