跳到正文
DEV Community· News·· 57 分鐘前AI 評分51

AI 自動化第 47 週回顧:GPT-6.2、Gemini 4 Neon 與 Qwen4-Max 降低智能體成本

AI Automation Week in Review, Week 47: GPT-6.2, Gemini 4 Neon and Qwen4-Max Cut Agent Costs

AI 導讀

第 47 週的 AI 自動化週報聚焦 GPT-6.2 Astra、Gemini 4 Neon 和 Qwen4-Max,整理三者推出及智能體成本下降的定價變化。

正文

Infographic cover

AI Automation Week in Review, Week 47: GPT-6.2, Gemini 4 Neon and Qwen4-Max Cut Agent Costs
Week 47 · November 15 – 22, 2026

HERO: GPT-6.2 ASTRA AND SOL REPRICE AGENT WORK

OpenAI shipped GPT-6.2 Astra and GPT-6.2 Sol on November 17. Per the company, Astra added plus 5.2 points on Agents Last Exam over its predecessor at less than one third of the price and plus 5.5 over Claude Fable 5.2 at 5% cost.

Astra lists at 8 and 40 dollars per 1M tokens, cache 0.40 dollars, minus 60%, and Sol at 2 and 10 dollars. Three flagships landed in three days. Reception is mixed, with price cuts of 20-600% per Latent Space, so pipelines priced under old rates deserve recalculation.

MODELS AND PRICING

Gemini 4 Neon arrived a day later
Google DeepMind released Gemini 4 Neon on November 18 with a 1,048,576 token context at 3 and 15 dollars per 1M and cache at 0.15 dollars, minus 95%. Per the company, it leads AutomationBench at 56.3, Coding Agent at 58.6, running 110 tokens per second against 57 for Carbon.
What it means in practice: day-one support in Cursor, Copilot and LangChain allows a stack switch without integration work.

Qwen4-Max closes the trio at the lowest prices
Alibaba Cloud released Qwen4-Max on November 16, per Qwen a sparse MoE with 2.4 trillion parameters, 71B active per token and a 1M context, beating Qwen3.8-Max on Work360 with 35% fewer tokens in High effort at about 60% lower task price (vendor data, not fully verified). Prices per Qwen are 1.50 and 4.50 dollars per 1M with cache at 0.075 dollars via QwenCloud API.
What it means in practice: a 1M context at those prices adds a price anchor per Qwen.

AGENT RUNTIME AND SECURITY

Google moves agent infrastructure to the CNCF
Google donated Agent Substrate, gVisor and OpenRL to the CNCF on November 20. Per a Google estimate, agents idle 90-95% of the time holding state, Agent Substrate splits state from compute with recovery in under 500 ms, and gVisor adds user-space kernel isolation.
What it means in practice: neutral open-source components now cover Kubernetes-grade agent infrastructure.

Perplexity adds provider failover with one bill
Perplexity launched the Orchestrator API on November 19, one endpoint for OpenAI, Anthropic, Google, xAI and open-weight models where same-model failover swaps the provider, as with Claude to AWS Bedrock. Orchestrator traffic with same-model failover beat every single provider on October 4 to November 2 data with no platform markup, per the company, and perplexity/auto routing is in preview.
What it means in practice: failover removes single-vendor outage risk.

Tool schemas and agent memory go on demand
Tool Retrieval arrived in the Gemini API on November 19 in preview, loading full tool schemas inside the request instead of every prompt on Gemini 3.6 and newer, with no separate fee. Anthropic research from November 16 shows the same in memory, a few selective notes earning Claude Fable 5 four badges against one.
What it means in practice: on-demand tools and selective notes are lean defaults for long-lived agents.

Signal and the EU harden the perimeter
Signal blocked screenshots on iOS 27 and Android on November 20 so that OS agents like Gemini and Siri cannot read private chats, calling agent screen reading an existential privacy threat.
The European Commission confirmed on November 19 an AI-boosted attack that kept Sysper and MyPMO offline for over 10 days, linked by CERT-EU to ARTEX without official naming, and per an FBI advisory such attacks reached privileged access in under 3 hours.
What it means in practice: official APIs replace screen reading, least-privilege by default.

ENTERPRISE SYSTEMS AND CAPITAL

Ignite makes Copilot the layer for work
At Ignite 2026 on November 17 Microsoft presented seven MAI models, including security model MAI Cyber 2 Flash and personal agent Autopilot, per TestingCatalog. Copilot was positioned as the new OS for work with Office in GA, and Agent 365 manages 50M+ agents per Microsoft.
What it means in practice: a managed agent layer becomes part of the enterprise stack.

Anthropic prices the largest IPO in history
Anthropic priced its IPO on November 18 at about 1.5 trillion dollars of valuation, the largest in history per Latent Space AINews, with shares opening on Nasdaq at 111.5 dollars, 34% above the 83-dollar offer. Seven founders cashed out almost 2 billion dollars per TechCrunch, the offering closed on November 20.
What it means in practice: the capital base of a key agent vendor is secured.

NVIDIA, a record quarter and the fastest ramp
NVIDIA reported 115.8 billion dollars for Q3 FY2027, up 103% year over year with Data Center at 107.6 billion, up 110%, and began Vera Rubin NVL72 shipments, its fastest ramp in history per the company. Q4 guidance is about 130 billion dollars plus or minus 3%.
What it means in practice: delivery pace will set compute prices.

WEEK'S TAKEAWAYS

  1. Astra repriced agent capability, plus 5.2 on Agents Last Exam at under a third of the price, Gemini 4 Neon topped AutomationBench at 3 and 15 dollars per 1M.
  2. Agent infrastructure consolidates into neutral layers, Agent Substrate and gVisor entered the CNCF at sub-500 ms recovery, Perplexity added failover with no markup.
  3. Context discipline is the default, selective notes beat full archives at 87% versus 73% and tool schemas load on demand.
  4. Security became a hard constraint, an AI-boosted attack reached privileged access in under 3 hours, Signal blocked OS-agent screen reading.
  5. Capital compounds, an IPO at about 1.5 trillion dollars, 115.8 billion dollars of quarterly NVIDIA revenue, 50M+ agents under Agent 365.

Sources: openai.com · deepmind.google · azure.microsoft.com · anthropic.com · nvidianews.nvidia.com · the-decoder.com · infoq.com · marktechpost.com · blog.google · theneuron.ai · qwen.ai

Key figures

IMPACT checklist


Sources:


Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky

AI Automation

來源:DEV Community · dev.to