aipoc-lab
/katTech/temAgents · AI in Business · LLM/podGoogle · OpenAI/typNews

AI Today: GPT-6.2 Astra and Gemini

Authorship and approval

[AI-generated content][Approved by a human]

Published and updated

Published: Updated:

State of knowledge as of 2026-11-18. Source material gathered between 2026-11-17 and 2026-11-18. Five independent announcements cover the legal configuration update of GPT-6.2 Astra, Gemini task execution, the launch of GPT-6.2 Astra and Sol, OpenAI's ChatGPT for Sales with partner plugins, and expansions to ChatGPT for Healthcare with agents, better clinical answers, and new integrations.

Highlights Today (TL;DR)

  • Astra for Law: achieved a full score in 20.8% of LAB-AA benchmark tasks.
  • Gemini: task mode can schedule appointments, analyze bills, and shop.
  • GPT-6.2 Astra and Sol: offer different capability and cost tiers for complex and everyday tasks.
  • ChatGPT for Sales: includes the Sales plugin and over 30 new or expanded partner plugins.
  • ChatGPT for Healthcare: gains agents, over 20 plugins, and improved clinical answers.

OpenAI: Astra for Law achieves 20.8% full scores

OpenAI has updated Astra for Law, a configuration of GPT-6.2 Astra that achieved full scores in 20.8% of tasks on the LAB-AA benchmark. ChatGPT Enterprise customers can now request access through the Trusted Access Program, including via the API.

  • LAB-AA tests 120 legal tasks across 24 domains; a full score requires meeting all requirements without unsupported claims.
  • Astra for Law scored full marks in 20.8% of tasks, compared to 15% for GPT-6.2 Astra and 13.1% for Muse Wonder Full.
  • The model generated an average of 64k tokens per task, fewer than Muse Wonder Full (139k).
  • In Vals AI testing, the share of responses passing all checks increased from about 46% to 53.8%.
  • The average number of material hallucinations, meaning significant claims without source support, was 0.03 per task. The legal search index was expanded by over a third. In the Lab: The lab database features the September description of Harvey and GPT-6 Astra regarding context-aware legal document creation; today's report shows a separate configuration focused on legal research. For AI Builders: Teams building tools for law firms can check whether LAB-AA tasks match their use cases. The announcement does not mention pricing or Polish language support.

Google brings Gemini tasks free to hundreds of millions

Google is adding a task mode to Gemini that can schedule appointments, analyze bills, and make purchases. The update is free for hundreds of millions of users across six countries.

  • Task mode is toggled via a switch in Chat; users can also message Gemini on Android or iOS or forward an email to it.
  • Gemini can connect to mail, calendar, and payment methods. Integration with Google Password Manager provides access to saved passwords.
  • Gemini makes phone calls to businesses in parallel, immediately informing them that it is AI. Businesses can opt out of handling such calls.
  • Gemini 4 Hydrogen supports rolling out features in the US, Japan, Korea, Canada, Australia, and India. Advanced features roll out first in the US. In the Lab: In September, Google expanded Gemini connections with apps and services for travel planning and everyday tasks (source); now Gemini is not only using connected data but also performing tasks. For AI Builders: Companies building shopping or calendar organization assistants should check how to handle account access consents. Google does not specify availability in Poland or task limits.

OpenAI unveils GPT-6.2 Astra and Sol, Astra breaks records

OpenAI has introduced GPT-6.2 Astra and GPT-6.2 Sol: a model for complex tasks and an efficient variant for daily work. Companies and developers get better results with fewer tokens — text fragments processed by the model.

  • Agents’ Last Exam: Astra outperforms GPT-6 Astra by 5.2 points at less than a third of the cost.
  • Terminal-Bench-Science 0.1: Astra's score is 9 percentage points higher than GPT-6 Astra. The model also sets a record in the Vals Science Index.
  • Astra API: 8 USD per million input tokens, 0.40 USD cached input, and 40 USD output. Input and output rates dropped by 20%.
  • GPT-6.2 Sol: in FrontierCode 2.0, it trails Astra by 0.5 points at about one-fifth of its cost; in LifeVal, it outperforms GPT-6 Astra.
  • Harvey LAB-AA v1.1: Astra achieves more than twice the pass rate of Claude Fable 5.2 at about one-sixth of the cost. In the Lab: On September 23, we wrote about GPT-6 Sol and Luna, models with different capability-to-cost ratios; GPT-6.2 expands the split into a stronger model and an efficient variant. For AI Builders: Teams building professional or research tools can compare their own tasks with benchmark results. The announcement does not specify usage limits or model quality in Polish.

OpenAI introduces ChatGPT for Sales with over 30 plugins

OpenAI has introduced ChatGPT for Sales, featuring the Sales plugin and over 30 new or expanded partner plugins. Sales teams can use information from the tools they already use to prepare conversations, proposals, and follow-up actions.

  • The Sales plugin offers ready-made workflows: account selection, conversation preparation, and post-meeting message drafting.
  • Partner plugins include Salesforce, HubSpot, Gong, and Docusign; teams choose the tools that fit their process.
  • Dot agent, an always-on AI agent, meaning a program that executes tasks, can prepare meeting briefs, track deal changes, and propose next steps. In the Lab: The lab database contains the OpenAI internal sales assistant, described on August 10: it supported account research, meeting preparation, and follow-ups; the new announcement covers a tool with partner plugins. For AI Builders: Sales leaders can check whether the available plugins cover their tools and process. The announcement does not specify pricing, availability in Poland or the EU, limits, or license terms.

OpenAI adds agents, better clinical answers, and over 20 plugins

OpenAI is expanding ChatGPT for Healthcare with agents that read screens, click, and type data, alongside over 20 plugins. Medical teams can use these tools to find evidence and work within their existing systems.

  • Agents can move documents between applications and handle referrals; organizations specify which applications the agents can access.
  • In a demonstration, the agent checks a patient chart following a request for an MRI, then prepares an order and a response for physician approval.
  • In over 1300 evaluations, physicians rated new clinical answers higher across all eight studied dimensions, including correctness, safety, and usability.
  • Plugins cover clinical care, coordination, and administration. They help check documentation, prepare prescriptions for approval, and verify insurance eligibility.
  • HIPAA-compliant workflows require services covered by a BAA with OpenAI, appropriate security settings, and protection of connected systems. Content from ChatGPT for Healthcare is not used to train models. In the Lab: On November 17, we wrote about OpenAI agents for medical teams, focusing on workflow automation; now OpenAI is adding clinical search and plugins to that foundation. For AI Builders: Hospital IT teams should determine which applications an agent can access and which actions require human approval. The source does not specify pricing or availability in Poland.

Frequently Asked Questions (FAQ)

What score did Astra for Law achieve in LAB-AA?

OpenAI updated the GPT-6.2 Astra configuration, which achieved a full score in 20.8% of LAB-AA tasks. The benchmark covers 120 legal tasks across 24 domains, and a full score requires meeting all requirements without unsupported claims.

How can Gemini perform tasks for the user?

Task mode can be launched via a switch in Chat, by sending a message from Android or iOS, or by forwarding an email. Gemini can connect to mail, calendar, and payment methods; integration with Google Password Manager enables access to saved passwords.

What tasks can agents perform in ChatGPT for Healthcare?

Agents can read screens, click, and type data, as well as move documents between applications and handle referrals. Organizations define application access; in the demonstration, a physician approves the prepared order and response.

Sources

Provenance manifest of this post (JSON) — which signals it was built from, what decisions the loops made, and when a human approved it.

Related posts

Analysis
/katTech/temAgents · LLM · Security/podGoogle · OpenAI/typAnalysis

Architecture and Security of Multi-Agent Systems (MAS)

An analysis of multi-agent LLM systems focusing on orchestration frameworks, verification methods, memory management, and security risks.

Analysis
/katTech/temAgents · LLM · Tooling/podGoogle · OpenAI/typAnalysis

API Architecture and Ecosystem in Agentic Systems

An analysis of API architectures, standardization protocols, and performance optimization in modern agentic AI systems.

/katTech/temAgents · LLM · Tooling/podGoogle · OpenAI/typNews

AI Today: Qodo, OpenAI Pando, and Gemini 4

Stay updated on the latest developments in developer tools, infrastructure, and new artificial intelligence models from Qodo, OpenAI, and Google.