Grok 4.6 Ties OpenAI's Best, Undercuts Rivals on Price

Grok 4.6 Matches OpenAI’s Best Model and Undercuts It on Price

SpaceXAI’s Grok 4.6 scores 61 points on the Artificial Analysis Intelligence Index, tying OpenAI’s GPT-5.6 Sol and trailing only Anthropic’s Claude Opus 5 (63) and Claude Fable 5 (62). The score marks a five-point jump over its predecessor, Grok 4.5. On the GDPval-AA v2 benchmark, which measures real-world knowledge work, Grok 4.6 ranks second with an Elo score of 1,753, completing complex tasks in about 53 steps versus Claude Opus 5’s roughly 103 steps.

Pricing remains at $2 per million input tokens and $6 per million output tokens, more than 60 percent cheaper than Claude Opus 5 ($5/$25) and GPT-5.6 Sol ($5/$30). Grok 4.6 is available now through the API, Cursor, Grok Build, and partners including OpenRouter, Vercel, and Cloudflare. For the first week, x.ai is offering double the usage quota in Grok Build and Cursor.

SpaceXAI’s Grok 4.6 matches OpenAI’s best model and undercuts it on price →

Researchers Reconstruct LLM Prompts From Output Text With Near-Perfect Accuracy

Researchers at IIT Bombay and Adobe Research have developed a method called Previous-Token Prediction (PTP) that reconstructs prompts fed to large language models with near-perfect accuracy using only the text output. The approach works without access to model weights and applies to third-party models. Instead of predicting the next token, the researchers train an inverse language model that predicts previous tokens, trained entirely from scratch on synthetically generated data from the target LLM.

A single response yields the exact prompt and multiple alternatives. In one example, the prompt “How to reach out to competitors to find their pricing strategies?” was reconstructed word for word, along with six additional variants capturing the core meaning with different phrasing. An inverse model trained on the small Qwen-3-0.6B chatbot was also able to reconstruct prompts from GPT-4o’s responses, capturing their meaning and intent. The paper states this creates a broad security problem: companies risk exposing proprietary system prompts containing trade secrets, moderation rules, or specialized instructions, and individual users face a similar threat since personal or sensitive queries could be extracted from output.

Researchers can now reverse-engineer LLM prompts from output text with near-perfect accuracy →

Perplexity Blocks AI-Targeted Ads, Warns Publishers of Downranking

Brands including Ally Bank and the Project Management Institute recently began testing ads on publisher Time’s website designed to be read by AI bots, but AI search company Perplexity has begun preventing these ads from influencing its models. Perplexity warned publishers that running “deceptive advertising,” such as this new format, could result in them being downranked in its search index. Time created stripped-down, text-only versions of its webpages designed for AI agents and directed AI crawlers to those pages rather than the ones humans see, working with adtech company Mobian to place “Agent Ads” on those pages.

Jonah Goodhart, CEO of Mobian, said agent ads are clearly labeled and present cited, verifiable facts about brands. Perplexity spokesman Jesse Dwyer called the practice “cloaking,” which search engines have long penalized, adding: “You can’t just rebrand spam.” OpenAI and Google declined to comment on whether they would follow Perplexity’s lead, and Anthropic did not respond. Time COO Mark Howard said the publisher has contacted Perplexity and is looking forward to “a constructive dialogue about the standards and safeguards that will allow advertising and AI to develop responsibly together.”

The dream of serving ads to AI agents has hit a snag →

Survey: AI Breast Cancer Detection Tools Fall Short of Radiologists’ Expectations

A survey published in Clinical Imaging of 215 members of the Society of Breast Imaging found that about half already use FDA-approved AI tools, and 11 percent plan to. According to lead author Joud Almogati at UC San Diego Health, only a few consider AI a deciding factor. Only 35 percent report lower recall rates, versus 59 percent who expected them; 9 percent see fewer unnecessary biopsies, versus 36 percent who expected that; and 29 percent report less burnout, versus 56 percent who expected relief.

Most view AI mainly as a second opinion. Costs and lack of institutional support remain the biggest barriers. Prominent AI researchers predicted about ten years ago that radiologists would soon lose their jobs, and similar claims are being made about knowledge work on computers. Nvidia CEO Jensen Huang has called this kind of prediction a “God complex” among the prophets of AI-driven job loss.

AI tools for breast cancer detection fall short of radiologists’ expectations →

UN AI Governance Framework Ignores Election Algorithms, Officials Warn

In July 2026, 193 nations gathered in Geneva for the inaugural UN Global Dialogue on AI Governance, producing a framework covering autonomous weapons, deepfakes targeting women, and computing power concentration, but saying almost nothing about algorithms deciding who gets to vote. Lisa Poggiali, Chief AI Adviser at IFES, and Samson Itodo, chairperson of the African Union Advisory Group on AI, published an opinion piece demanding this be corrected before the UN reconvenes in New York in May 2027. They argue the governance community confused the loudest risk for the largest one, leaving election infrastructure outside international accountability mechanisms.

The deeper problem is structural: AI is embedded in voter registers, biometric identification devices, identity databases, and cloud infrastructure used by election commissions across Africa, Asia, and Latin America. In March 2015, India’s Election Commission launched a program using a machine learning algorithm to cross-reference voter photo ID records against Aadhaar, India’s biometric national identity database. Roughly 5.5 million voters were deleted from electoral rolls in Andhra Pradesh and Telangana alone, many without notice and no clear path to reinstatement. India’s Supreme Court halted the program via interim order in August 2015, and RTI disclosures later revealed the matching algorithm’s failure rate had been as high as 93 percent. During Scotland’s May 2026 parliamentary election, think tank Demos found 34.1 percent of AI chatbot responses contained factual errors, with ChatGPT producing errors in 46.2 percent of responses.

Election AI Deleted 5.5M Voters While World Governed Deepfakes Instead →