Anthropic Launches Claude Opus 5 — Matching Fable 5 at Half the Cost
Anthropic made the week's biggest move on Thursday, releasing Claude Opus 5 — a model that comes remarkably close to the company's own flagship Fable 5 while costing exactly half as much. Priced at $5 per million input tokens and $25 per million output tokens, Opus 5 slots in at the same price point as the outgoing Opus 4.8, making it an immediate upgrade for existing users.
The benchmark numbers are striking. On Frontier-Bench v0.1, Opus 5 scored 43.3% compared to Fable 5's 33.7%, and it won five of nine head-to-head evaluations. On CursorBench 3.2, it performed within 0.5% of Fable 5's peak score at half the cost per task. Perhaps most notably, it posted the lowest misalignment score of any recent Claude model on Anthropic's automated behavioral audit.
Opus 5 ships with a 1M-token context window and a new low/medium/high effort toggle, letting developers trade cost for capability on a per-request basis. It's now the default model on Claude Max and the strongest model available on Claude Pro.
The release underscores a broader trend in the AI industry this month: the cost of frontier-level intelligence is collapsing. With GPT-5.6, Grok 4.5, and now Opus 5 all competing aggressively on price, developers have never had more options at the high end of the capability curve.
Source: Anthropic
Stripe in Talks to Acquire OpenRouter for Nearly $10 Billion
In what could become one of the year's landmark AI deals, Stripe is reportedly in advanced talks to acquire OpenRouter, the AI model-routing platform, for approximately $10 billion. The deal, first reported by the Wall Street Journal on July 24, would represent a massive jump from OpenRouter's roughly $1.3 billion valuation following a May funding round.
Founded in 2023, OpenRouter lets developers access, compare, and switch between dozens of AI models — both proprietary and open-source — through a single API. It has become critical infrastructure for companies that don't want to be locked into a single AI provider. Stripe already processes OpenRouter's payments, making the acquisition a natural vertical integration play.
If completed, the deal would mark Stripe's biggest move beyond payments and signal that AI model routing and orchestration is now considered core infrastructure worth paying a premium for. Discussions remain ongoing and could still fall apart, but sources suggest an announcement could come within a month.
Sources: Wall Street Journal via Yahoo Finance, Benzinga
DeepSeek V4 Fully Live as Legacy API Endpoints Go Dark
As of July 24 at 15:59 UTC, DeepSeek officially pulled the plug on its legacy API model names — deepseek-chat and deepseek-reasoner — forcing all users onto the new V4 model family. The transition completes a rollout that began with V4's general availability on July 20.
DeepSeek V4 brings a 1-million-token context window and a new peak/off-peak pricing structure, with input tokens priced as low as $0.14 per million during off-peak hours. The aggressive pricing continues to put pressure on Western competitors and makes DeepSeek one of the most cost-effective frontier-class options available.
Developers who haven't migrated their API calls will find their applications broken as of Thursday. DeepSeek has been warning about the deprecation for weeks, but the hard cutoff caught some teams off guard, particularly those running on older SDK versions.
Sources: DeepSeek API Docs, WaveSpeed Blog
Nvidia and SK Group Unveil $500 Billion Korea AI Package
Nvidia CEO Jensen Huang and SK Group Chairman Chey Tae-won announced a sweeping $500 billion AI infrastructure initiative in South Korea, combining chip supply agreements with SK Hynix and Samsung, AI cloud infrastructure buildouts, and direct Nvidia investments in Korean companies.
The announcement is part of South Korea's broader $880 billion 10-year national plan covering semiconductors, AI infrastructure, and robotics. Samsung and SK Hynix have committed $518 billion toward new chip fabrication sites, positioning the country as a global hub for AI hardware manufacturing.
The scale of the investment reflects the escalating AI infrastructure arms race, where compute capacity and advanced memory chips have become strategic national assets. With the U.S., China, and now South Korea committing hundreds of billions, the race to build the physical backbone of AI is accelerating faster than the models themselves.
Source: The Korea Herald
Quick Hits
Delhi Court Sides with OpenAI on Copyright: India's Delhi High Court rejected a news agency's injunction against OpenAI, ruling that ChatGPT's training constitutes "fair dealing" under Indian copyright law — a significant precedent for AI training data disputes in one of the world's largest markets. (LiveLaw)
Meta Launches AI-Powered Seller App: Facebook head Tom Alison unveiled Seller, a standalone iOS app for Facebook Marketplace that uses Meta AI to auto-generate listings — titles, descriptions, pricing, and shipping labels — from uploaded photos. (Meta)
FTC Seeks Input on AI Accuracy Rules: The Federal Trade Commission opened public comments on a policy statement addressing AI accuracy requirements, with responses due by July 31. The move signals the agency is preparing enforceable standards for AI output truthfulness. (FTC)