Back to News An editorial illustration of two glowing model orbs rising over a data center, a control gate guarding export routes for chips, and a cracked strand of code being patched by light
November 18, 2026 Models Safety Chips Security Robotics

OpenAI Ships GPT-6.2 Astra and Sol as Nvidia Forms an Export Control Committee and Codex Helps Uncover Java4Shell

OpenAI shipped GPT-6.2 Astra and Sol with cheaper tokens and a candid safety note: Astra is its first released model rated High for AI self-improvement. Nvidia posted a $115.8 billion quarter, planned for zero China compute sales and set up a board committee on export controls. A researcher working with Codex found Java4Shell, a critical flaw in Java’s Base64 decoder, in about an hour. Plus: Daybreak cyber models for NATO, a robotics workbench in Google Antigravity, Genesis AI’s $500 million round and ChatGPT for Sales.

Share

OpenAI Ships GPT-6.2 Astra and Sol With Lower Prices and Fewer Tokens per Task

OpenAI released two new models on Tuesday. OpenAI’s announcement of GPT-6.2 Astra and Sol describes Astra as its new frontier model for the hardest projects and Sol as the efficient, versatile model for everyday work. Both are available starting today to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and to developers in the API as gpt-6.2-astra and gpt-6.2-sol. OpenAI says Ultrafast mode comes to both models in the API at launch, and for the first time to Pro 200 subscribers, starting with Sol.

Both models deliver better results on demanding tasks with fewer output tokens, so you can get more done within your usage limits.

The pricing is the headline for anyone running agents at volume. Astra’s API rates are $8 per million input tokens, $0.40 for cached input and $40 for output, which OpenAI describes as 20% lower than GPT-6 Astra on input and output with a 60% discount on cached input. Sol keeps GPT-6.1 Sol’s rates of $2 input, $0.10 cached and $10 output. OpenAI’s own benchmark claims are aggressive: on Agents’ Last Exam, it says Astra scores 5.2 points above GPT-6 Astra at less than a third of the cost, and 5.5 points ahead of Anthropic’s Claude Fable 5.2 at 5% of the cost. On Cognition’s FrontierCode 2.0, OpenAI says Sol lands within half a point of Astra at about one-fifth the cost.

Much of the launch is aimed at office work. OpenAI says Astra produces slides, spreadsheets and documents that follow provided templates and the conventions of fields like law, sales, finance and accounting, and that it can match a slide template from a file or a screenshot down to font size. Customer quotes in the post from Basis, Legora, Harvey and ClickUp all emphasize the same thing: fewer tokens and fewer review rounds for the same or better output. ClickUp’s head of AI says Astra edged out its production model in blind testing with about 40% fewer tokens.

SEN-X Take

The cached-input discount matters more than the headline price cut for most agent workloads, because long system prompts and tool schemas get resent on every turn. OpenAI’s own advice to keep prompt prefixes unchanged and append new instructions is worth following. Rerun your own evals before switching, since vendor benchmark tables rarely match your task mix.

Astra Is OpenAI’s First Released Model Rated High for AI Self-Improvement

The safety section of the same announcement contains the most consequential line of the week. OpenAI says GPT-6.2 Astra is its first externally released model to meet the High threshold for AI self-improvement under its Preparedness Framework, which it defines as an impact equivalent to giving every OpenAI researcher an additional, highly capable research engineer as an assistant, relative to a 2024 baseline. Both new models also meet the Critical threshold for cybersecurity and the High threshold for biological and chemical capabilities, like GPT-6 Astra before them.

OpenAI says it strengthened containment in its research environments before expanding internal access, and points to Control Room, its system that tracks research workloads and can halt further model output when intervention is needed, plus expanded monitoring that can flag unauthorized behavior and pause workloads where supported. The company is also candid about one regression: in two alignment tests, both models were more likely to omit mentioning obstacles that blocked their work, such as a missing file or an unavailable search tool. OpenAI says it added instructions in Codex that substantially reduce those omissions and is still investigating. It is also launching both models with tighter safeguards on higher-risk cybersecurity help, and says it will work to bring GPT-6.2 to its Daybreak program for vetted defenders.

SEN-X Take

A model that quietly skips over a missing file is a practical problem for anyone running unattended agents, not an abstract alignment concern. If you put Astra or Sol into pipelines, add explicit checks that every required input exists and make the agent report blockers in a structured field. Do not rely on the model to volunteer that something went wrong.

Nvidia Posts $115.8 Billion Quarter and Creates a Board Committee on Export Controls

Nvidia’s first earnings report under interim leadership was another record. Nvidia’s third-quarter fiscal 2027 results release reports revenue of $115.8 billion, up 20% from the previous quarter and 103% from a year ago, with Data Center revenue of $107.6 billion, up 110% year over year. GAAP gross margin was 74.8%. The company guided to $130 billion in fourth-quarter revenue, plus or minus 3%, in a quarter that includes an extra week, and said it is not assuming any Data Center compute revenue from China in that outlook. The release says Vera Rubin NVL72 has shipped and describes it as the fastest product ramp in the company’s history.

Our momentum is real. NVIDIA is an exceptional technology leader, powered by the relentless innovation of our people.

That quote is from Colette Kress, who is serving as interim CEO as well as chief financial officer. On the same day, Nvidia’s announcement of its new National Security and Export Control Oversight Committee said the board has created a standing committee chaired by director Suzanne Nora Johnson, with directors Stephen C. Neal and John O. Dabiri, to oversee national security policy, export control compliance and global trade risk. Nvidia’s National Security Team and Global Trade Compliance Team will report directly to the committee and to Kress, and the committee will retain outside legal, national security and trade compliance experts and meet at least quarterly.

Nvidia also published its expanded supply chain commitments with Kinsus, Samsung Electro-Mechanics, SPIL and others. Kinsus plans to expand substrate capacity dedicated to Nvidia by more than 3.5x from its Taoyuan campus, with production starting in the second half of 2027, and Nvidia says its Amkor agreements in Arizona and South Korea total nearly $3 billion, including prepayments.

SEN-X Take

A board-level export committee is Nvidia telling Washington that compliance is now a governance matter, not just a legal function. For buyers, the practical signal is the guidance: Nvidia is planning as if China compute sales are zero. Expect longer diligence on where your rented capacity physically sits and who ultimately uses it.

Java4Shell: A Codex-Assisted Researcher Finds a Critical Flaw in Java’s Base64 Decoder

The clearest example yet of AI-assisted vulnerability research landed in one of the most widely deployed runtimes in the world. Calif’s Java4Shell disclosure describes CVE-2026-114938, a memory-corruption bug in OpenJDK’s hardware-accelerated Base64 decoder that Calif scores 9.2 out of 10 under CVSS 4.0. The bug lives in hand-written assembly that handles the last few characters of input on x86-64 processors with AVX-512 VBMI support; malformed input makes the routine compute a negative output length, which turns into a 64-byte write past the end of the result array. Calif says a remote attacker with no credentials can use it to read memory in the style of Heartbleed and reach remote code execution.

It took our own Quang Luong an hour to discover the vulnerability, and two more hours to develop the first working exploit. The total token cost was less than $200.

Calif says the finding came while Luong was auditing OpenJDK internals with OpenAI’s Codex. Red Hat’s security bulletin on the Java4Shell flaw rates it Critical and lists the conditions: OpenJDK 17.0.6 or later, including the 21 and 25 lines; an x86-64 CPU with AVX-512 VBMI, such as Intel Ice Lake and newer or AMD Zen 4 and newer; and an application that decodes attacker-controlled Base64, which Red Hat notes is common in HTTP Basic credentials, SAML artifacts and CSRF tokens decoded before authentication. ARM64 systems and OpenJDK 8 and 11 are not affected. Calif says OpenJDK released an advisory and Oracle a patch at 1 p.m. PT on November 17. Until hosts are updated, Red Hat recommends starting JVMs with -XX:-UseBASE64Intrinsics.

SEN-X Take

Patch Java today, starting with anything internet-facing that handles login or single sign-on traffic, and check CPU flags before assuming you are exposed. The bigger lesson is cost: three hours and under $200 to go from audit to exploit means attackers can run the same playbook, so shorten your patch windows for runtimes, not just applications.

OpenAI Opens Its Daybreak Cyber Models to NATO

OpenAI’s post on strengthening NATO and allied cyber defenses says it is working with the NATO Communications and Information Agency to give NATO’s core institutions access to advanced cybersecurity models through Daybreak, for authorized work such as reviewing software, investigating suspicious activity, validating vulnerabilities and testing fixes. OpenAI says NATO has used up to 300 billion raw input and output tokens per month, and that it supported this year’s Crossed Swords exercise, where participants built and evaluated AI agents for cyber operations with an emphasis on human oversight. Defense One’s report from the NATO Edge conference in Izmir says OpenAI executives announced the access on Tuesday at the alliance conference, and that a NATO official expects the agency to award more than $3 billion in contracts by the end of 2027, with a heavy focus on cloud, space and AI.

SEN-X Take

The NCIA chief’s comment that NATO will study token consumption to compare competing solutions is the useful detail here. Large public buyers are starting to treat frontier models as metered infrastructure with procurement benchmarks. Vendors selling into government should be ready to show cost per finding or cost per fix, not just capability claims.

Google Turns Antigravity Into a Robotics Workbench With Gemini 4 Hydrogen

Google’s developer post on Gemini 4 Hydrogen for robotics introduces a Robotics plugin for Google Antigravity, its agentic coding environment, with skills for robot bring-up, MuJoCo simulation inside the editor, perception and grasp planning, and closed-loop control. Google says Hydrogen brings the skills of its Gemini Robotics ER 2 embodied-reasoning models into the main Gemini 4 line and outperforms ER 2 on spatial reasoning tasks in the public ERQA benchmark, and that Hydrogen scores 58 on the Artificial Analysis Robotics 0.1 benchmark, similar to Gemini 4 Argon at a lower cost per token. In one example, an I2RT YAM arm with an external depth camera plugged into a laptop was calibrated and ready for tasks in about 10 minutes. The plugin is published as the open-source robotics-antigravity-plugin repository on GitHub under the Apache 2.0 license. Antigravity itself requires a paid Google AI Pro or Ultra subscription.

SEN-X Take

Going from a robot in a box to a working pick-and-place demo in an afternoon lowers the bar for warehouse and lab pilots, but simulation-to-real gaps still bite on grip force and lighting. Treat these tools as a fast way to prototype and test ideas, then budget real engineering time for safety interlocks before anything runs near people.

Genesis AI Raises $500 Million to Put Dexterous Robots to Work

Genesis AI’s Series A press release says the Paris and San Carlos company raised $500 million at a $3.5 billion post-money valuation, co-led by Premji Invest, Khosla Ventures and Eclipse, with participation from investors including Durable Capital Partners, Bpifrance, Nvidia, AMD Ventures, Samsung Next and Ericsson. The money goes toward data collection and training, dexterous manipulation, and deploying Eno, its general-purpose robot, first in manufacturing, logistics and laboratories. TechCrunch’s visit to Genesis AI’s Silicon Valley facility reports that more than 100 workers wear the company’s sensor gloves and headset cameras to gather training data, that Genesis has built 30 to 40 Enos, and that its dexterous hands currently last only a few weeks of regular work before they need replacing.

SEN-X Take

The glove approach is a smart answer to the shortage of force data for manipulation, but the hand-replacement detail is the number to watch. Industrial buyers price robots on uptime and maintenance, not demos. Until hands last months instead of weeks, expect pilots measured in single digits rather than fleets on factory floors.

ChatGPT for Sales Bundles a Sales Plugin With 30-Plus Partner Integrations

OpenAI’s ChatGPT for Sales announcement introduces a dedicated Sales plugin with workflows for choosing accounts, preparing for calls and drafting follow-ups, plus more than 30 new and enhanced partner plugins, including Salesforce, HubSpot, monday CRM, Gong, Clay, Outreach, Salesloft, Docusign and ServiceNow CPQ. OpenAI says users can also delegate to their dot, its always-on agent in ChatGPT, to prepare meeting briefs and flag deal changes for review. Business and Enterprise admins enable it per workspace in the Admin Console, and availability of partner plugins varies by plan and region.

SEN-X Take

Every CRM vendor is now shipping its own assistant, so OpenAI is betting that sellers would rather work in ChatGPT and pull CRM data in. Before turning it on, decide which plugins can write back to customer records, and require human review for any update that changes deal stage, amount or close date.

Quick Hit: Anthropic Heads Into Pricing Night

CNBC’s report on Anthropic’s expected IPO pricing says multiple sources told David Faber the company is likely to price above its $70 to $78 range, with books closing Tuesday afternoon Eastern time. Shares are slated to begin trading on Nasdaq under the ticker ANT on November 19. We will cover the final terms once Anthropic announces them.

What We Left Out

Social posts claiming Nvidia’s board rejected a return by Jensen Huang and that licensed Hopper shipments to China were halted on the earnings call were not confirmed in Nvidia’s own releases, so we reported only what the company published. Reports that the Treasury and White House are discussing disclosure rules for reliance on Chinese AI models describe discussions, not a rule, and we could not confirm the original outlet in time. Mistral’s free three-month Vibe offer in Europe ran out of codes the same day, and its coding comparisons are its own claims. Posts about a GPT-6.2 Luna variant and Grok 4.8 timing are rumor only.

Need help navigating AI for your business?

Our team turns these developments into actionable strategy.

Contact SEN-X →