• NATURAL 20
  • Posts
  • GPT-6.2 Arrives as Google Makes AI Agents Free

GPT-6.2 Arrives as Google Makes AI Agents Free

PLUS: Perplexity launches Multiagent, Qwen4-Max debuts, and Hugging Face adopts OpenRouter routing.

In partnership with

The Dell Pro keeps your day moving.

Built for those who stay in the flow.

Handle the busiest days by staying in the flow with built-in intelligence on the Dell Pro laptop powered by Intel® Core™ Ultra with Intel vPro®.

Today:

  • OpenAI launches GPT-6.2 Astra and Sol for more efficient professional and coding work

  • Google makes Tasks in Gemini free across six countries

  • Isomorphic Labs raises $5 billion to advance AI-driven drug discovery

  • Perplexity Computer adds Multiagent for complex assignments

  • Alibaba releases the full Qwen4-Max model after its earlier preview

AI Is Getting More Capable and Easier to Put to Work

OpenAI has released two updated frontier models, Google is taking a major step toward bringing AI agents to free users, and Isomorphic Labs has secured $5 billion to advance the search for new medicines with AI.

The developments reflect two priorities driving the industry: making advanced models more effective at work while reducing the friction of actually using them. At the same time, new research and safety programs are examining what happens when AI systems work across longer tasks and more consequential environments.

Agents’ Last Exam v1.1 chart comparing AI model performance and cost, with GPT-6.2 Astra achieving the highest score.

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol with improvements in professional work, coding, computer use and efficiency. Astra targets the most demanding multi-step reasoning and software projects, while Sol is designed to handle everyday coding and agent workflows more quickly and affordably.

On Agents' Last Exam, a benchmark built around complex professional assignments, OpenAI says Astra scores 5.2 percentage points above GPT-6 Astra at less than one-third of the task cost. The company also reports a 5.5-point advantage over Claude Fable 5.2 at approximately 5% of the task cost under its evaluation setup.

The models are intended to make practical work more reliable, from analyzing information and using desktop applications to editing code across repositories. OpenAI says defensive coding was a particular focus, introducing training signals that reward writing more secure software and repairing vulnerabilities without breaking expected behavior.

The release also includes an Ultrafast processing option for customers prioritizing response speed. OpenAI is positioning the two models as different choices rather than insisting that the most powerful model is necessary for every task; users can select reasoning depth and performance according to the work.

GPT-6.2 Astra and Sol are rolling out to eligible Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and developers can access both models through the API. Ultrafast availability and usage terms vary by product and plan.

The performance and cost comparisons are based on OpenAI's reported evaluations, where differences in reasoning effort, task mix and token usage can substantially affect results. The company's accompanying safety assessments also note that the new models can still omit known obstacles when explaining their actions.

Gemini mobile interface showing settings for starting AI tasks through text messages and email.

Google announced that Tasks in Gemini is beginning to roll out free of charge in the United States, Japan, South Korea, Canada, Australia and India. Previously associated with paid access, the feature uses Gemini 4 Hydrogen to carry out multi-step requests instead of simply producing an answer.

Users can assign a task such as organizing a schedule, preparing an itinerary, looking for savings opportunities or finding appointments that fit an existing calendar. Gemini can work with connected Google services, gather information and carry out supported steps, while users can steer or change instructions before the job is complete.

Google also introduced additional ways to start tasks. Beginning in the U.S., people can forward an email to Gemini or send a request through Google Messages or iMessage after setting up the connection. They can reply to adjust a task while it is running rather than starting over.

Each task runs inside Google Cloud's Agent Compute Engine, using a hardware-isolated virtual machine assigned to the account. Google says this keeps the task environment separate from those of other users as the agent navigates websites and works through connected services.

The free rollout is limited to the six announced countries and eligible adult users. Some additional capabilities begin in the United States, and higher-tier plans or other countries may have different access and usage limits.

The shift makes autonomous task completion available to a much larger audience, but outcomes still depend on service permissions, the websites involved and whether the agent correctly interprets the request. Users should review consequential changes and completed actions.

Isomorphic Labs announced a $5 billion Series C funding round as it works to translate AI-driven drug discovery into potential medicines. The company is a Google DeepMind spinout focused on using advanced computational models to accelerate the design and development of drug candidates.

Greenoaks and Sequoia Capital co-led the round, with new participation from Andreessen Horowitz and continued backing from Alphabet. Existing investors including Thrive Capital, MGX, GV, CapitalG, Temasek and the UK Sovereign AI Fund also participated.

The funding will support research infrastructure, the company's AI-driven drug design platform and programs intended to advance its first therapeutic candidates toward clinical trials. Isomorphic is trying to integrate predictive modeling with laboratory experimentation and drug development rather than treating molecule generation as an isolated software task.

Building a promising molecule is only an early step toward a medicine. Researchers must still test whether candidates are effective, safe, manufacturable and suitable for human studies, and many compounds do not progress through those stages.

The scale of the financing signals investor confidence in applying frontier AI to biology. It also provides Isomorphic Labs with significant resources to undertake the slower and more expensive experimental work needed after initial computational discovery.

The company's announcement describes funding and development ambitions, not newly approved treatments or evidence of successful human clinical trials. Its early programs remain subject to scientific and regulatory validation.

🧠RESEARCH

Envisioning Agentic Robotics outlines a framework for robots that perform sustained tasks rather than isolated movements. It pairs foundation models with an embodied harness for context, tools, planning, memory, and verification. The authors propose ways to evaluate long-term autonomy and safety; this is a research roadmap, not a demonstrated robot.

Gamer trains an AI agent to play long, interactive games using screen observations and keyboard-and-mouse actions. Researchers combine gameplay trajectories, task supervision, memory, and reinforcement learning to improve multi-step performance. They report gains on tested game benchmarks, but results do not establish reliable play across every real-time game or environment.

Sage S2.1 Flash Technical Report presents a 320-billion-parameter mixture-of-experts system for scientific reasoning and interactive work. It combines general language abilities with specialized scientific encoders to interpret evidence such as molecular structures and spectra. The authors report results across 33 benchmarks covering scientific tasks, challenging coding tasks, and agent workflows.

📲SOCIAL MEDIA

🗞️MORE NEWS

Perplexity introduced Multiagent, a new effort level in Computer that lets a lead agent divide difficult assignments among other agents using different frontier models. The agents share findings, cross-check one another and contribute to a common deliverable, with a separate Model Council option that compares independent model answers. Multiagent is available on the web for Perplexity Max subscribers.

Alibaba launched Qwen4-Max on November 16, moving beyond the limited Qwen4-Max Preview covered in our November 13 edition. The new model has 2.4 trillion total parameters, native visual understanding and a one-million-token context window; Alibaba reports improvements on its Work360 benchmark while using 35% fewer tokens at high reasoning effort and reducing average task costs by about 60% compared with Qwen3.8-Max. Those efficiency figures are from Alibaba's own evaluation.

OpenAI announced agreements with Apollo Research, METR, Transluce and Goodfire to give outside specialists deeper access to models, internal systems and safety evidence. The planned assessments cover alignment, monitoring, safeguards and whether training and deployment processes surface important risks. The projects are at different stages, with additional work planned into 2027, so signing the agreements does not mean every assessment has already been completed.

Hugging Face's Inference Providers now uses OpenRouter to route chat-model requests across available inference services, choosing providers according to live price, speed and reliability. The integration adds provider selection and failover while preserving existing Hugging Face authentication and API workflows. Image, audio and video generation continue to use Hugging Face's separate routing rather than OpenRouter.

What'd you think of today's edition?

Login or Subscribe to participate in polls.

Reply

or to participate.