Skip to content

Frontier AI, Security & Silicon

OpenAI Debuts GPT-6.2 Astra & Sol as Microsoft and Nvidia Surge

OpenAI releases GPT-6.2 Astra and Sol with 20% lower token pricing while Microsoft launches its Perception agentic security suite at Ignite 2026 and Nvidia reports record $115.8 billion third-quarter revenue.

By Lazy Tensor
Published · 4 min read

Restrained editorial print illustration on warm cream paper depicting geometric AI model nodes, security circuits, and data-center wafers.
Conceptual editorial print illustration of frontier model architectures, autonomous cybersecurity agent loops, and data-center silicon wafers. · Lazy Tensor Visual Desk
Your listening Queue & catch up
Unheard stories first, timed at 1×. Your selections stay first.

Add a story or build a catch-up queue to start listening.

Frontier model efficiency, autonomous cybersecurity defense, and hyperscale semiconductor earnings converged on Tuesday as OpenAI released its GPT-6.2 Astra and GPT-6.2 Sol models, Microsoft unveiled its generally available Perception agentic security suite at Ignite 2026, and Nvidia reported that third-quarter revenue more than doubled year-over-year to $115.8 billion.

OpenAI Launches GPT-6.2 Astra and Sol With Lower Token Costs and High Self-Improvement Safeguards

According to OpenAI News, OpenAI rolled out GPT-6.2 Astra for complex professional and scientific workflows and GPT-6.2 Sol for everyday agentic work across ChatGPT Work, Codex, and the OpenAI API, bringing Ultrafast inference to Pro 500 and Pro 200 subscribers. On Agents’ Last Exam, GPT-6.2 Astra scored 5.2 points higher than GPT-6 Astra at less than one-third the cost and 5.5 points ahead of Claude Fable 5.2 at 5% of the cost. It also gained 9 percentage points on Terminal-Bench-Science 0.1, set a new record on the Vals Science Index, more than doubled Claude Fable 5.2’s pass rate on Harvey LAB-AA v1.1 at one-sixth the cost (with Astra for Law scoring even higher), improved 6.2 percentage points on GDPval v2 using 22% fewer output tokens (while Sol gained 5.3 points), answered 62.5% of AA-AnalystAgent tasks correctly across all five attempts (including a 2031 Alderfen Field Systems valuation at an 8% discount rate and 3% perpetual growth), and led AA-3D human evaluations at 1,318 Elo ahead of Claude Fable 5.2 (1,271) and GPT-6.2 Sol (1,259). Meanwhile, GPT-6.2 Sol came within 0.5 points of Astra on Cognition’s FrontierCode 2.0 at one-fifth the cost and beat GPT-6 Astra on LifeVal at one-seventh the cost.

In the API, GPT-6.2 Astra cuts input and output token prices by 20% with a 60% cached-input discount—pricing at $8 for input, $0.40 for cached input, and $40 for output per million tokens—while GPT-6.2 Sol retains rates of $2 for input, $0.10 for cached input, and $10 for output per million tokens. Evaluated on DeepBinary (across 12 Windows binaries), PHP2Shell, and SecureSWE, both models meet the Preparedness Framework’s Critical threshold for cybersecurity (supported via OpenAI Daybreak Blue and Red) and High threshold for biological and chemical capabilities, while GPT-6.2 Astra is OpenAI’s first externally released model to reach the High threshold for AI Self-improvement relative to a 2024 baseline, governed by OpenAI’s Control Room monitoring system and Codex Blocker Deception mitigations.

Microsoft Unveils Perception Agentic Security Platform at Ignite 2026

At Microsoft Ignite 2026, Microsoft announced the general availability of Microsoft Perception and core Integrated Security Operations Center (ISOC) capabilities in Microsoft Defender, as detailed on the Microsoft Security Blog. Built atop a foundation of more than 500 connectors and 165 trillion daily signals across Defender, Entra, Intune, Purview, and Sentinel (with 90-day default Defender retention for E5 and E7 customers), Perception orchestrates red, blue, and green team security agents through automated playbooks.

Microsoft noted that more than 96,000 vulnerabilities have been published to the CVE Program in 2026—more than double 2024’s roughly 40,000 and 20 times 2010 levels—while only 26% of CISA Known Exploited Vulnerabilities were remediated in 2025 and roughly 29 million PHP servers remained exposed two weeks after PHP2Shell weaponization. Perception introduces four generally available blue team agents (Alert Triage, Attack Investigation, Detection Authoring, and Threat Intelligence), red team Recon and Web App Pen Test agents, green team Posture Prioritization and Risk Reduction agents across Microsoft, AWS, and Google Cloud, Codename MDASH repository scanning up to 5 GB powered by MAI Cyber 1.1 Flash and MAI-Cyber-2, January 2027 Purview and AI Readiness previews, discovery for 60 local and cloud agent types (up by 35) with Microsoft Entra Agent IDs, and the AI Transparency Initiative with NIST and over 40 partner organizations.

Nvidia Reports Q3 Revenue Up 103% Year-Over-Year to $115.8 Billion

On the hardware side of the AI boom, interim Nvidia CEO and executive vice president and CFO Colette Kress announced on Tuesday that Nvidia’s third-quarter fiscal 2027 revenue rose 103% year-over-year to $115.8 billion, driven by data center revenue that surged 110% year-over-year to $107.6 billion and a record $65.3 billion in net profit, according to The Verge — AI. Kress highlighted that the company’s Vera Rubin platform is off to a strong start without initially mentioning former CEO Jensen Huang in the written release ahead of the analyst earnings call.

Sources

Follow the coverage

Published reporting connected by OpenAI.

  1. NVIDIA Details Supply Chain as Apple Tests Siri Caps & OpenAI Grows
  2. OpenAI Teams With NATO on Cyber Defense as Microsoft Tests Remote Lock
  3. OpenAI Debuts GPT-6.2 Astra & Sol as Microsoft and Nvidia SurgeYou are reading this story
  4. OpenAI Launches ChatGPT for Sales With Over 30 Partner Plugins

Comments (0)

No comments yet. Be the first to share a thought on this story.

The Lazy Tensor briefing

Get the next story in your inbox.

Independent reporting on the people, companies and ideas shaping AI and technology. Free, once a week.

More from Lazy Tensor