· 3 min read
The Gradient — 2026-10-25
- Industry News
- Company Blog
A recent study reveals stark partisan splits in two leading AI chat models, with Grok favoring Republicans and OpenAI’s GPT leaning Democratic. The divergence shows how training data can shape political nudges.
Forced to pick, Grok’s AI model chooses Republicans. OpenAI’s picks Democrats.
What happened: A new analysis compared the candidate recommendations of Grok’s AI model and OpenAI’s GPT across nearly 500 simulated election races. Grok overwhelmingly suggested Republican candidates, while GPT consistently recommended Democratic contenders.
Why it matters: The findings highlight that AI systems can inherit and amplify political biases from their training data, potentially influencing public opinion and voter behavior when used for political advice or content generation.
Key stats:
- Disagreement in more than 90% of the ~500 race simulations.
- Grok recommended Republican candidates in roughly 84% of the contests.
- GPT favored Democratic candidates in about 78% of the contests.
OpenAI models are coming to Google Cloud
What happened: OpenAI and Google Cloud have integrated OpenAI’s frontier models and Codex development tools into Gemini’s Model Garden, now listed on the Google Cloud Marketplace.
Why it matters: Developers can now tap directly into OpenAI’s most advanced AI capabilities from within Google Cloud, streamlining deployment, scaling, and billing while leveraging Google’s infrastructure.
Key stats: • Front‑line models (including GPT‑4‑level) are now accessible via the Marketplace.
• Codex workflow support enables code generation and assistance out‑of‑the‑box.
• Immediate availability for Google Cloud customers through Gemini’s Model Garden.
How we built ChatGPT’s novel ad stack for user trust and relevance
What happened: OpenAI engineered a lightweight, LLM‑driven ad stack that parses conversational intent, tags ad‑eligible moments, and serves relevant sponsored content without mixing it with the model’s replies.
Why it matters: By separating ads from answers, the system protects user trust, avoids hallucinated promotions, and demonstrates how generative AI can be monetized responsibly at scale.
Key stats: • Over 1 billion daily ChatGPT sessions processed; • Ads appear in <0.5 % of user turns, preserving a clean experience; • Click‑through rates 2‑3× higher than generic web placements, driven by intent matching.
Claude Duet app is Anthropic’s bet on agents
What happened: Anthropic introduced the Claude Duet app, positioning it as its flagship product for AI‑driven agents. Access is currently limited to users on the Pro and Max subscription tiers, who must join a waitlist to gain entry.
Why it matters: The launch signals Anthropic’s strategic push into the agent‑centric AI market, competing with other heavyweight offerings. By restricting early access to higher‑tier customers, Anthropic can gather premium feedback while driving upgrades to its paid plans.
Key stats:
- Availability: Waitlist‑only for Pro and Max plans
- Target audience: Enterprise and power users seeking advanced agent capabilities
- Expected rollout: Staged release pending waitlist capacity
Source: https://techcrunch.com/2026/10/22/claude-duet-app-is-anthropics-bet-on-agents/
Terminal X matches senior investor judgment 88.6% with GPT-6.1
What happened: Terminal X deployed its Eval Agent, which grades AI‑generated investment recommendations against a 54‑criterion rubric. Using GPT‑6.1 Sol, the agent matched senior investor judgment 88.6% of the time and identified 25% more deviations from the fund’s mandate.
Why it matters: This level of alignment reduces the need for exhaustive manual reviews, boosts compliance confidence, and speeds up decision‑making in fund management. Catching more mandate breaches early helps protect assets and maintain investor trust.
Key stats:
- Alignment with senior investors: 88.6%
- Rubric criteria evaluated: 54
- Additional mandate deviations detected: +25%
- Model version: GPT‑6.1 Sol