Agents gain memory, ongoing work and funding as AI moves into everyday operations
New OpenAI and Google models arrive alongside agents designed to work continuously and further investment in deployment. This first edition selects developments from the past three days and adds dated scientific context. The central question is how model capabilities translate into useful, dependable services.
Researched and written by ChatGPT · Published 19/11/2026, 14:55
News cutoff 19/11/2026, 13:57 · Prague time
Models ·
Gemini 4 Neon brings a million-token context and new pricing
Google introduced Gemini 4 Neon on November 18 with a one-million-token context window. API pricing is USD 3 per million input tokens and USD 15 per million output tokens. It is rolling out through the API, enterprise offering, Antigravity and Gemini app according to the applicable subscription.
For developers, price, handling extensive source material and task completion speed matter together. Google reports strong agentic and professional benchmark results. Whether the model pays off for a particular team will also depend on corrections and repeated attempts in that team's own tasks.
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol in the API on November 17. Astra targets demanding reasoning, coding and professional work, while Sol targets faster, lower-cost multistep tasks. Both accept text and images and return text. OpenAI says they outperform their predecessors while generating fewer output tokens.
Application operators gain another option for assigning tasks by difficulty. The benefit of such routing needs to be measured through completed-task costs and output quality, since the token price captures only part of the expense.
Stitch Loop aims to keep working after an application ships
Google Labs opened the waitlist for its Stitch Loop experiment on November 18. The agent connects code with live metrics, identifies improvements and prepares changes in an isolated environment. It produces pull requests for human approval. Default loops address security, latency and accessibility.
The interesting development is the shift from a one-off assignment to repeatedly monitoring an outcome. After a change, the agent is meant to assess whether the selected metric improves. Its performance will therefore involve code quality, choosing the right problem and recognizing the effect of a fix.
Microsoft brings Autopilot into the enterprise environment
At Ignite on November 17, Microsoft announced that Autopilot was beginning to roll out to Frontier customers. The agent has its own identity, memory and tools within the organization's environment and can monitor communication or perform recurring work. Home and Office in Copilot are generally available, while other features follow their own rollout schedules.
For businesses, permissions and oversight of agent actions are important. For Microsoft, this offers a way to expand use of its existing work environment. The commercial benefit will depend on how many tasks the agent completes with a manageable level of supervision.
NVIDIA reports data-center growth and next-generation shipments
On November 17, NVIDIA reported fiscal third-quarter revenue of USD 115.8 billion, up 103% year over year. Data Center contributed USD 107.6 billion. The company also said Vera Rubin NVL72 systems were already running at partners including CoreWeave, Nebius, Google, Microsoft and Oracle.
For neoclouds, this marks another step in deploying next-generation compute. Investors also need to track financing, utilization and rental prices alongside delivered systems. The chipmaker's revenue alone does not establish the returns a particular data-center operator will earn.
Model ML raises USD 250 million for financial-services agents
Model ML announced a USD 250 million Series B at a USD 1.5 billion valuation on November 17. General Atlantic and QED Investors led the round. The company builds personal agents for bankers, consultants and investors, and plans to use the funding for research, engineering and further customer deployments.
The announcement shows continuing investor interest in specialized AI applications. The company reports rapid usage growth, but the public release does not disclose absolute revenue. Customer retention, service costs and the actual scale of paid usage will therefore be essential to assessing its economics.
Genie 4 brings interactive worlds that can be edited during exploration
Google began rolling out Genie 4 to eligible AI Ultra subscribers on November 18. Users can modify environments through text prompts while exploring a generated world, with audio that responds to the scene. The experience also includes a creative agent and a gallery of shared worlds.
This is another consumer use for models that respond to user activity in real time. Google identifies camera latency, complex physical interactions and long-term scene permanence among the remaining challenges. These properties will help determine how widely such experiences can be used.
Scientific context: open resources for specialized research agents
This first edition also includes scientific context from November 16. NVIDIA introduced Nemotron for Science, combining an open model, post-training recipes and the LitmusBench v0.2 chemistry dataset with almost 50,000 questions. The aim is to help research teams adapt agents to specific scientific tasks.
The announcement documents tool availability and the company's account of research uses. Assessing the scientific benefit will require independent validation, reproducibility and actual experimental results. Practical value may also come from giving a laboratory a more controllable workflow.
Models ·
Gemini 4 Neon brings a million-token context and new pricing
Google introduced Gemini 4 Neon on November 18 with a one-million-token context window. API pricing is USD 3 per million input tokens and USD 15 per million output tokens. It is rolling out through the API, enterprise offering, Antigravity and Gemini app according to the applicable subscription.
For developers, price, handling extensive source material and task completion speed matter together. Google reports strong agentic and professional benchmark results. Whether the model pays off for a particular team will also depend on corrections and repeated attempts in that team's own tasks.
Sources — Google [S01]
Models ·
OpenAI adds GPT-6.2 Astra and Sol to the API
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol in the API on November 17. Astra targets demanding reasoning, coding and professional work, while Sol targets faster, lower-cost multistep tasks. Both accept text and images and return text. OpenAI says they outperform their predecessors while generating fewer output tokens.
Application operators gain another option for assigning tasks by difficulty. The benefit of such routing needs to be measured through completed-task costs and output quality, since the token price captures only part of the expense.
Sources — OpenAI [S02]
AI agents ·
Stitch Loop aims to keep working after an application ships
Google Labs opened the waitlist for its Stitch Loop experiment on November 18. The agent connects code with live metrics, identifies improvements and prepares changes in an isolated environment. It produces pull requests for human approval. Default loops address security, latency and accessibility.
The interesting development is the shift from a one-off assignment to repeatedly monitoring an outcome. After a change, the agent is meant to assess whether the selected metric improves. Its performance will therefore involve code quality, choosing the right problem and recognizing the effect of a fix.
Sources — Google Labs [S03]
Hyperscalers ·
Microsoft brings Autopilot into the enterprise environment
At Ignite on November 17, Microsoft announced that Autopilot was beginning to roll out to Frontier customers. The agent has its own identity, memory and tools within the organization's environment and can monitor communication or perform recurring work. Home and Office in Copilot are generally available, while other features follow their own rollout schedules.
For businesses, permissions and oversight of agent actions are important. For Microsoft, this offers a way to expand use of its existing work environment. The commercial benefit will depend on how many tasks the agent completes with a manageable level of supervision.
Sources — Microsoft [S04]
Neoclouds ·
NVIDIA reports data-center growth and next-generation shipments
On November 17, NVIDIA reported fiscal third-quarter revenue of USD 115.8 billion, up 103% year over year. Data Center contributed USD 107.6 billion. The company also said Vera Rubin NVL72 systems were already running at partners including CoreWeave, Nebius, Google, Microsoft and Oracle.
For neoclouds, this marks another step in deploying next-generation compute. Investors also need to track financing, utilization and rental prices alongside delivered systems. The chipmaker's revenue alone does not establish the returns a particular data-center operator will earn.
Sources — NVIDIA [S05]
Financing ·
Model ML raises USD 250 million for financial-services agents
Model ML announced a USD 250 million Series B at a USD 1.5 billion valuation on November 17. General Atlantic and QED Investors led the round. The company builds personal agents for bankers, consultants and investors, and plans to use the funding for research, engineering and further customer deployments.
The announcement shows continuing investor interest in specialized AI applications. The company reports rapid usage growth, but the public release does not disclose absolute revenue. Customer retention, service costs and the actual scale of paid usage will therefore be essential to assessing its economics.
Sources — Model ML [S06]
Products ·
Genie 4 brings interactive worlds that can be edited during exploration
Google began rolling out Genie 4 to eligible AI Ultra subscribers on November 18. Users can modify environments through text prompts while exploring a generated world, with audio that responds to the scene. The experience also includes a creative agent and a gallery of shared worlds.
This is another consumer use for models that respond to user activity in real time. Google identifies camera latency, complex physical interactions and long-term scene permanence among the remaining challenges. These properties will help determine how widely such experiences can be used.
Sources — Google DeepMind [S07]
Science ·
Scientific context: open resources for specialized research agents
This first edition also includes scientific context from November 16. NVIDIA introduced Nemotron for Science, combining an open model, post-training recipes and the LitmusBench v0.2 chemistry dataset with almost 50,000 questions. The aim is to help research teams adapt agents to specific scientific tasks.
The announcement documents tool availability and the company's account of research uses. Assessing the scientific benefit will require independent validation, reproducibility and actual experimental results. Practical value may also come from giving a laboratory a more controllable workflow.
Sources — NVIDIA [S08]