The AI Brief

The Information reported on October 24 that OpenAI offered Replit a discount of more than 30% on model prices under a contract worth more than $100 million annually, after Anthropic declined to extend Replit's prior discount without a higher spending commitment. OpenAI gave Replit more than 30% off to win a contract worth over $100M a year, per The Information, after Anthropic refused to extend Replit's prior discount without a higher spending commitment. Replit CEO and co-founder Amjad Masad first disclosed the shift at TechCrunch Disrupt on October 16, telling the audience that Replit was "no longer working with Anthropic," citing cost, partnership, and performance. On performance he was blunt: "I think ChatGPT is better now."

The break is less clean than the headline. A person familiar with the deal told TechCrunch the Anthropic partnership is still in place, and Replit's own docs still list seven Claude models; OpenAI is now the default model provider for Replit Agent, and customers who want Claude have to pick it. The competitive signal goes beyond Replit: The Information paired the Replit shift with figures from Rippling, where OpenAI's share of coding work grew from 16% in May to 52%, while Anthropic's dropped from 57% to 18%. Neither Replit, OpenAI, nor Anthropic has publicly confirmed the financial terms.

The deal illustrates a structural dynamic in AI platform economics: Replit CEO Masad walked back his "no longer working with Anthropic" claim within 24 hours, and seven Claude models remain in the Replit catalog, but the default model switch routes most Replit users to OpenAI without active choice, concentrating revenue impact beyond what user-facing toggles suggest.

Anthropic is expected to publicly file its IPO prospectus as early as October 26 U.S. time, with plans to list in mid-November. The company confidentially submitted its draft S-1 to the SEC on June 1, and SEC regulations require a 15-day public disclosure period before roadshow marketing can begin, making a late-October filing the minimum for a November 9 roadshow start. The brief first covered the public EDGAR filing on October 23 (Vol. I, No. 151).

The market previously discussed a $2 trillion valuation; however, as AI-related stocks have come under pressure, recent investor expectations have been lowered to approximately $1.5 trillion. Anthropic reported $16.3 billion in revenue for the first half of 2026, up from $1.1 billion in 2025, but recorded an $80.7 billion net loss due to convertible bond adjustments. With 47% of 2025 revenue coming from Amazon and Google, Anthropic's risk-to-reward ratio remains elevated. Lead underwriters are Morgan Stanley, Goldman Sachs, and JPMorgan, per reporting; Anthropic has not confirmed the valuation or timing officially.

Disclosure: Claude, which generates this brief, is built by Anthropic.

Developer and analyst Max Woolf published a detailed timeline on October 25 tracing how OpenAI and Anthropic executed successive price cuts between June and October 2026. OpenAI cut GPT-5.6 Luna by 80% on July 30, dropping the price to $0.20 per million input tokens. On September 1, Anthropic launched Fable 5.1 at a cache read price of $0.25 per million tokens, cutting the prior rate by 75%, which Woolf estimates translates to a 70-75% reduction in total agent cost given how heavily agentic sessions rely on the cache.

Open-weight competitors including DeepSeek and Moonshot's Kimi K3.1 Pro, available from dozens of providers on OpenRouter, keep pricing pressure high. Both companies are nearing IPOs, which raises questions about their inference profit margins, reportedly as high as 80% in some reports. Woolf's framing: prices are a competitive floor set by open-weight alternatives, not loss-leaders, and open-weight models and secondary inference providers are structurally capping how high closed-model pricing can go.

Operators recalculating AI infrastructure budgets should isolate cache-read spend as a separate line item. Contracts that pre-date the July-September cuts may still be priced at pre-compression rates; the gap between old and new cache-read pricing is larger in dollar terms than the gap in any other rate category for long-session workloads.

Starting Monday, October 26, 2026, staff at the Regeringskansliet can no longer open ChatGPT and Claude from their work computers. The block covers the prime minister's office and every ministry, and it pushes employees toward approved tools, Microsoft Copilot among them, which has meanwhile had its security tightened after an internal review. The decision was first reported by Sveriges Radio's Ekot on October 23, citing internal communications from the office's digitalisation director, Mikael Wettercrantz.

Wettercrantz posted on the government offices' intranet that use of AI services without an agreement could have "consequences for our information security." Swedish agencies also detected instances of inappropriate use of AI, including the inclusion of sensitive information in public services and the processing of AI-generated responses without ensuring verification. The block is a network-level restriction on the public-facing websites; some employees can still access these two blocked tools through an AI platform built internally by the agency.

Ekot found the same move at two more agencies, and a third is weighing it. Neither OpenAI nor Anthropic commented publicly on the decision as of October 24. The Sweden action fits a broader European pattern: the EU AI Office issued its first enforcement requests to OpenAI and Anthropic in October, and the UK ICO opened a six-week agentic AI call for evidence the same month.

Texas Governor Greg Abbott directed the Texas Commission on Environmental Quality (TCEQ) on September 21 to halt all permits related to data center projects until ERCOT completes its grid and water audit. The Texas Commission on Environmental Quality suspended data center permits issued to at least 17 projects and shut down STEERS, the electronic filing system that handed them out, after two watchdog groups showed the approvals kept coming well after Abbott told the agency to stop. All 17 are stormwater construction permits, the paperwork that lets a developer start grading and leveling a site.

Watchdog groups FracTracker Alliance and Public Citizen published reports on October 19 covering more than two dozen permits issued to Texas data center projects after the September 21 directive. TCEQ's own database listed each of the suspended permits as "approved" and "active," with start dates after September 21. The controversy intensified because automated processing continued for projects associated with major technology companies, including OpenAI and Amazon, before the portal was taken offline. TCEQ's message on its website reads: "TCEQ's electronic filing application used for automatic authorizations has been temporarily shut down to ensure all applications are reviewed by agency staff to determine if they are related to data centers."

Abbott's order told the agency to halt issuance of all permits "related to data center projects" statewide until ERCOT finishes reviewing the data centers sitting in its interconnection queue, a job ERCOT expects to wrap up in mid-December. All new permit applications must now be physically submitted to TCEQ for manual staff review. Abbott's office denied any permits had been validly approved under his directive; the governor has also signaled intent to work with the legislature to eliminate financial incentives for data centers in Texas.