AI Vendors Grapple With the Price of Practical Intelligence
OpenAI's latest research models can solve math puzzles, but those headline moments don't keep the lights on. The real grind happens where companies must squeeze value from "just enough" intelligence-enough to automate routine work, not chase scientific fame.
On October 6, OpenAI ran internal models through a round of mathematical research, burning three hours of ChatGPT Pro compute for each result. Anthropic followed with Haiku 5.5, a compact model built for repetitive jobs like summarizing and classifying. OpenAI then started rolling out GPT-6 in ChatGPT, splitting access: Plus, Pro, Business, and Enterprise users got GPT-6 Sol, while free and Go users received the lighter Luna. The official OpenAI announcement set October 8 as the start date for free users, with changes limited to ChatGPT mode. Work and Codex models stayed put. The pattern is clear. The most advanced models chase research headlines, but daily business runs on cheaper, stripped-down AI.
OpenAI stated that GPT-6 is designed for over 1.2 billion weekly ChatGPT users and introduces 'Intelligent UI'-interactive interfaces embedded within responses.
The Economics of Sufficient Intelligence
Most companies don't need to crack unsolved math. They need to automate routine work without blowing their budgets. The gap between superintelligent models and what businesses actually deploy is wide. Alibaba's CEO Wu Yongming calls AGI a starting point, but buyers want models they can afford to run at scale. The price gap is hard to ignore: GPT-6 Astra costs $50 per million output tokens, Luna just $0.5. Anthropic's Haiku 5.5 clocks in at $0.1 and $0.5 per million input and output tokens. At the scale of tens of millions of daily queries, these numbers decide who stays in the game.
Anthropic claims Haiku 5.5 is about 75% cheaper to operate than Haiku 4.5. For requests up to 100,000 tokens, the price drops by 90%. Longer queries see a 50% cut. Anthropic's official pricing details put Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output tokens for the most common request sizes. Around 90% of previous Haiku requests fit this lower-cost window, so most customers see the savings.
But a low sticker price doesn't guarantee savings. If a bargain model forces users to retry or fix mistakes, costs pile up fast. Tencent's Marvis team cut daily user costs to a tenth of launch levels by tuning file indexing and model selection. Qianwen Office slashed average token use by 75% in office pilots. The only way to scale is to make each small job cost what it should-no more, no less.
OpenAI described GPT-6 Sol and GPT-6 Luna as models with different balances of capabilities and cost: Sol is intended for paid plans, while Luna serves free and Go tiers. This confirms OpenAI's strategy of distributing models according to workload and available budget.
From Model Specs to Business Impact
As model prices fall, developers can't charge a premium for raw AI muscle. Features like meeting record sorting are now baseline. The real edge comes from weaving AI into business workflows, permissions, and company data. Alibaba and Tencent spent 67.678 billion yuan and 52.8 billion yuan last quarter, showing the scale of investment needed. To make AI pay, those costs must be spread across products people actually use, not just hyped in press releases.
OpenAI's new "intelligent interface" for ChatGPT-able to spin up buttons, forms, and charts on the fly-signals a shift. If users can state what they want and let AI handle the rest, traditional software risks fading into the background. Tencent's Appbao, once valued for distributing apps, now faces a squeeze as user needs converge. Marvis, for instance, now tackles jobs like speeding up a slow computer or finding a video file, aiming to bake professional software features directly into the workflow.
Office platforms like DingTalk and Feishu aren't immune. Third-party AI agents can draft reports, but lack access to the deep company context-permissions and business data-that incumbents control. Qianwen Office's "Enterprise Context" and Feishu's Doubao Work integration both try to lock AI into the company's operational core, making old assets like permission systems newly valuable. But keeping these integrations running isn't cheap, and the work goes far beyond a simple software update.
The Hard Truth About AI Monetization
Turning intelligence into revenue is a tough sell. Individual users expect free system tools and may resist paying for AI-powered disk cleaning or file search. Enterprises need a reason to boost budgets beyond what they already spend on DingTalk or Feishu. Qianwen Office's seat-based pricing, with shared quotas and point bundles, tries to bridge the gap between old contracts and usage-based billing. The real test is whether AI can automate jobs that were too slow or costly before-like batch legal checks or handling logistics customer queries.
