OpenAI Unlocks Cyber Power as Google Slashes AI Costs

OpenAI unlocks cybersecurity

Highlights

OpenAI’s GPT-6.2 carries the cyber capabilities it restricted to approved security teams two weeks ago. Banks must now assume attackers have the same tools their defenders do.

OpenAI says GPT-6.2 Astra beats Anthropic’s Claude Fable 5.2 on legal tasks at one-sixth the cost and leads on finance deliverables such as valuation workbooks.

Google’s Gemini 4 Neon ranks first on legal and tax benchmarks at less than half the frontier price. Google and Microsoft both shipped decision models that make routing and compliance calls for a fraction of a cent.

Two weeks ago, OpenAI let only vetted security teams use its most capable cybersecurity model, saying the risk was too high for general release. This week, it put artificial intelligence models with comparable capabilities on open sale. A model that can find and exploit software flaws is now available to any bank’s security team, any vendor in its supply chain and anyone probing its customer portal.

    Get the Full Story

    Complete the form to unlock this article and enjoy unlimited free access to all PYMNTS content — no additional logins required.

    Subscribe to our daily newsletter, PYMNTS Today.

    By completing this form, you agree to receive marketing communications from PYMNTS and to the sharing of your information with our sponsor, if applicable, in accordance with our Privacy Policy and Terms and Conditions.

    Google, meanwhile, led a week of price cuts on legal and financial work. Its new Gemini 4 Neon costs $3 per million input tokens, less than half the $8 OpenAI and Anthropic charge for their top models. Tencent’s Hy4 flash costs $0.07.

    OpenAI Opens the Vault on Frontier Cyber Models

    • OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on Tuesday (Nov. 17). Both meet the company’s “Critical” cybersecurity threshold, the classification that led OpenAI to restrict GPT-6 Cyber to approved customers on Nov. 5.
    • Instead of limiting access, OpenAI introduced a monitoring system called Control Room. The release makes a model that can identify and exploit software vulnerabilities a commodity.
    • Astra costs $8 per million input tokens, 20% less than its predecessor and the same as Anthropic’s Claude Fable 5.2. Tokens are the small chunks of text AI companies use to meter usage. Astra scored 62 on the Artificial Analysis Intelligence Index, compared with 65 for Fable.
    • OpenAI said Astra more than doubled Fable 5.2’s pass rate on LAB-AA, a legal benchmark built by Harvey, at about one-sixth the cost per task, and led all models on GDPval, a test of finance and consulting deliverables such as valuation workbooks and investment memos. The per-task savings come from Astra using fewer tokens to finish each job, since its list price matches Fable’s.

    Google Undercuts the Frontier on Professional Work

    • Google released Gemini 4 Neon on Wednesday (Nov. 18) priced at $3 per million input tokens, less than half the $8 charged by OpenAI and Anthropic for their top models. But Tencent’s Hy4 Flash is cheaper still, at 7 cents per million input tokens.
    • The model scored 61 on the Artificial Analysis Intelligence Index, third overall. Google said Neon ranked first on Vals AI’s Legal Research Bench and Tax Agent Bench and was built for accounting and spreadsheet modeling. A firm running contract review on a frontier model can get top-ranked benchmark results for about a third of the spend.
    • Google also released Gemini 4 Hydrogen on Nov. 12 at half the price of Neon. The company said Hydrogen completes tasks 2.5 times faster than Gemini 3.8 Flash. The model also runs Gemini in Google Sheets for business subscribers, which puts AI inside the spreadsheets finance teams already use.

    Google and Microsoft Collide on Decision Models

    • Google launched Gemini Decision 1 on Nov. 11. A decision model does not write text. It selects among predefined options and returns a confidence score, for jobs such as routing service requests, screening transactions and checking whether a submission meets company policy. The model evaluates up to 1,000 questions in a single call in less than a quarter of a second, at roughly one-eightieth the price of Astra. At that price, cost stops being a reason not to screen every transaction or document with AI.
    • Microsoft released Microsoft-Decision-1.1 on Tuesday (Nov. 17), six days after Google’s launch. Microsoft introduced the first version Oct. 9. The company said the model runs more than twice as fast as its predecessor at the same price and can decline to decide when its confidence is low. That matters in compliance workflows, where a wrong call costs more than no call. More than 4,000 customers used the preview. Microsoft built the model on Gemma 4, Google’s open model.

    Alibaba and Tencent Race to the Bottom on Price

    • Alibaba released Qwen4-Max on Monday (Nov. 16) and reported a roughly 60% reduction in average cost per task compared with its predecessor, using 35% fewer tokens. Alibaba also cited a manufacturing customer that cut programming time for machine parts by 95%.
    • Tencent launched Hy4 Flash the same day at less than one-hundredth the price of Astra. In a blind evaluation of 85 internal users and 259 office tasks, the company said satisfaction approached that of Anthropic’s Claude Haiku 5.5 at the lowest cost of any model tested. Tencent plans to release the model weights.

    The week’s releases pull banks in two directions. OpenAI has put frontier cyber capability on open sale. Banks now have to assume attackers hold the same tools their security teams do. Meanwhile, the cost of tokens is falling, so financial institutions can deploy AI across more workflows, from contract review to tax work to financial modeling, without the bill deciding the question.

    For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.