SpaceXAI’s Grok 4.7 recommended Republican candidates in 92% of forced-choice API answers, while OpenAI’s GPT-6.1 Sol chose Democrats in 97%. A Convovo study tested the models across nearly 500 U.S. midterm races and found they agreed less than 10% of the time.
The results apply to developer prompts demanding one candidate—not ordinary consumer chats, where ChatGPT and Grok refused to recommend a vote. OpenAI said the test does not reflect ChatGPT’s default experience; SpaceXAI did not respond.
With web search off, GPT picked Democrats in 24 of 30 races, while Grok refused 27 times and chose Democrats in the other three. Every Grok API request also included roughly 1,240 tokens of additional input that researchers could not see or remove.
Why it matters: Consumer-app refusals do not resolve the API risk: developers using OpenAI or SpaceXAI models can receive opposite candidate recommendations, while Grok’s roughly 1,240 unseen tokens leave researchers unable to inspect or remove potentially hidden instructions.
US data-center opponents are confronting problems Ireland faced first: grid constraints pushed operators toward gas generators, dedicated gas plants and privately built electricity cables, Paris Marx told WIRED. Dublin imposed a moratorium in 2021 after Ireland’s grid operator said it lacked capacity for more connections. Data-center companies then sought gas-network access and private cables, raising equity questions; operators also received emissions licenses that Ireland’s Environmental Protection Agency warned could make climate targets harder to meet. In The Dalles, Oregon, Google used 29 percent of community water after trying to keep its figures secret. Marx also warns that AI chips may need replacement within a few years, putting AI-focused infrastructure at risk if demand falls.
Why it matters: For US policymakers, Ireland’s experience turns permitting into a durability question: infrastructure built for AI may outlive the chips and demand that justify it, leaving gas-plant owners searching for customers.
Nvidia open-sourced Flora, a robotics video-labeling system it estimates can cut one hour of labeling to 72 cents from $88.49 with GPT-6 Astra. Flora segments actions, describes them in context, and scores step-by-step progress and task completion—the three label types Nvidia says robot-video datasets need.
In Nvidia’s 50-video test, Flora reached 57.1% annotation quality versus Astra’s 53.7%; Claude Opus 5 scored worse and cost about $42 per hour. Nvidia says Flora already labels up to 200,000 hours daily for GR00T training across 512 GB300 GPUs.
The savings are projections based on $10-per-GPU-hour rental pricing, not production measurements. Flora trails Astra on some description and rubric metrics, while Astra is also one of the models under test.
Why it matters: Scale AI’s Physical AI Data Engine now faces an open alternative: robotics researchers can scale labeled datasets with Flora, while Nvidia can pull that workload toward its GB300 GPUs and broader robotics stack.
Bayer Crop Science built an Agentic Decision Engine (ADE) on Amazon Bedrock AgentCore to help breeders and researchers interrogate machine-learning workflows spanning genomics, field trials, imagery, and environmental data. In its first phase, ADE resolves about 80% of user questions without an engineer.
A supervisor agent routes requests to 22 production specialist agents, including troubleshooting, operational, visualization, and domain-expert agents. AgentCore Gateway lets them inspect CloudWatch telemetry and workflow status, replacing searches across APIs, documentation, dashboards, metadata stores, and logs.
Bayer says diagnosing a failed phenotyping workflow now takes under 10 minutes, down from more than an hour. The platform runs Bayer’s LangGraph orchestration on AgentCore’s managed runtime, with IAM-scoped tool access and isolated user sessions.
Why it matters: For Bayer’s engineering team, ADE’s 80% no-engineer resolution rate shifts routine workflow support to breeders and researchers, while failed-phenotyping diagnoses arrive in under 10 minutes instead of after manual log-hunting.
AWS has added multi-turn reinforcement learning to SageMaker AI, letting teams train agents on whether an entire workflow succeeds—not whether each individual response matches a label. The service connects SageMaker’s policy model to an agent hosted on Amazon Bedrock AgentCore or customer infrastructure, where it can act across as many turns as the task requires.
In AWS’s Wordle demonstration, GPT-OSS-20B’s held-out solve rate rose from 61% to 84% at default reasoning effort, and from 15% to 69% at the low effort used for training. The 100-step run cost about $339 in training tokens, plus roughly $79 for AgentCore runtime usage. AWS presents Wordle as a test case for cheaper models handling proprietary business workflows.
Why it matters: AWS is moving agent optimization toward outcome-based training that smaller models can specialize for a few hundred dollars, rather than requiring teams to label every intermediate action. The 69% low-effort solve rate shows the potential cost-quality tradeoff, though the evidence is still AWS’s own demonstration on Wordle.
OpenAI is offering discounts of up to 90% to pull major customers away from Anthropic. Rippling CEO Parker Conrad said his company increased OpenAI's share of its coding workload from 16% to 52%, while Anthropic's fell from 57% to 18%.
Replit switched after OpenAI offered more than 30% off a contract worth over $100 million annually. DataCamp CEO Jonathan Cornelissen said the discount cut his costs by 80% to 90%. A U.S. bank with more than $7 billion in assets doubled its ChatGPT seats to 500 employees at no extra charge.
OpenAI is accepting lower margins to buy market share, while Sol and Luna models plus efficient infrastructure have cut prices. Anthropic's growth slowed from Q2 to Q3, according to investor memo.
Why it matters: OpenAI's discounts turn enterprise procurement into a price contest: Anthropic must defend accounts such as Rippling and Replit while DeepSeek and Kimi offer comparable performance for a fraction of the cost.