Claude Opus 5 at Half the Price: AI Competition Shifts Decisively Toward Cost Efficiency

Claude Opus 5: The Half-Price Flagship Reshaping AI Competition

According to reporting from Zhuyan AI Tech on Bilibili, Anthropic has officially launched Claude Opus 5 — a model that rivals the current top-tier flagship in performance but is priced at roughly half of what competitors charge via API. Across multiple independent benchmarks, the model delivers significantly improved performance at dramatically lower cost, earning it the industry nickname of "half-price flagship."

Interestingly, Opus 5 leaked ahead of its official launch, with users independently testing its 3D generation capabilities and finding performance that rivals industry leaders. Anthropic plans to officially release it to market as early as September and has begun instructing employees to execute on restricted stock schedules — a signal backed by expectations of Q3 profits potentially exceeding $1 billion.

This launch marks a pivotal shift in the large language model landscape: from a pure capability race to a head-on battle over price-to-performance. As model capabilities begin to converge, pricing becomes the decisive weapon in the fight for developers and enterprise customers.

Inference Cost Becomes the Deciding Factor: The Second Half of the Price War

If training capabilities defined AI competition over the past year, the battlefield has now moved to inference costs. Reports indicate that inference now accounts for two-thirds of total AI compute consumption — meaning whoever can break through on inference efficiency will win the second half of this game.

Open-source forces are playing an increasingly important role in this transformation. DeepSeek demonstrated that compute limitations can be overcome through engineering optimization. Following the release of its K3 model, the open-source and closed-source camps began competing head-to-head, drawing attention from the White House and Wall Street alike. Data shows that over the past year, 40% of model downloads on Hugging Face came from Chinese open-source models — a trend that is reshaping the balance of influence in the global AI ecosystem.

On the path to cost reduction, the industry has converged on three primary approaches, with every major player searching for the optimal solution. Every 1% improvement in inference efficiency translates directly into more leverage in the price war.

The Edge AI Chip Race: Fitting 100B-Parameter Models Into Desktop Devices

Competition at the chip level is equally intense. Origin Semiconductor raised over 1.2 billion RMB in three months, closing a 700 million RMB Series A alone, targeting edge AI chips with the ambition of building the "computing foundation" for the Agent era. Another company, Aquare, released a 5nm AI chip claiming to fit hundred-billion-parameter large models into desktop devices, with cumulative funding reaching $350 million.

AMD's market size projections are strikingly ambitious: by 2030, the AI accelerator market will reach $1.4 trillion, with the data center CPU market at approximately $220 billion — combined, approaching $2 trillion. Its Helios rack's launch customers include heavyweights like Microsoft and OpenAI, signaling an all-out intensification of competition between edge and cloud inference.

The Battle for AI Agent Entry Points: Control the Agent, Control the Traffic

Agents are becoming the next strategic battleground — particularly on mobile, where three competing approaches have emerged: device makers building their own models, third-party apps racing to own the entry point, and a three-way contest that has barely begun. The core logic is simple: whoever controls the Agent controls the traffic.

Cognition acquired Poke in an attempt to embed "AI personality" into developer tools. As model capabilities converge toward commodity, character and emotional connection are becoming the key differentiators for user retention. Major Chinese tech companies are also collectively pivoting their Agent strategies: shifting from the previous multi-team horse-race model toward single-point focus. ByteDance, Alibaba, Tencent, and Baidu have each identified their primary product lines, as competition moves into the deep water of scenario differentiation and data flywheel building.

The New AI Infrastructure Middle Layer: AI Managing AI

A landmark event: OpenRouter, the aggregated model API routing platform, was acquired by Stripe for approximately $10 billion. This transaction bridges the AI call layer with the payment routing layer, constructing a new middleware layer for internet infrastructure.

The deeper shift is that AI is now beginning to "manage itself" — automated scheduling and self-healing fault recovery can both be handled by AI, with engineers gradually moving into supervisory roles. This signals that AI system operations are transitioning from human-led to autonomous, and foreshadows yet another paradigm shift in software infrastructure.

Capital Euphoria and Hidden Risks: Cash Flow Warnings Behind Google's 40x Return

Google's investment in Anthropic has become a landmark case study: an initial $3 billion investment now holds a stake valued at $124 billion — a return exceeding 40x. This figure reflects both the enormous upside potential of the AI sector and the market's frenzied concentration of bets on leading players.

Yet beneath the boom, undercurrents are stirring. One company posted net profit exceeding 100 billion RMB for the first time in Q2, but its free cash flow was negative $5.8 billion. Another company, Mola, saw its CEO abruptly resign on the day of its earnings release, sending the stock down 15% and wiping out nearly $10 billion in market cap. These signals remind us that the path to AI commercialization is far from clear, and a massive chasm still exists between high investment and profitability.

Multimodal and Embodied Intelligence: New Unicorns Are Emerging

VIVX released the world's first real-time interactive multimodal models, A1 and W1 — approximately 30 billion parameters each — capable of generating video and enabling interaction simultaneously, creating a "conversational video world." The company recently closed a Series C round of 1.5 billion RMB in three months, with a valuation exceeding 10 billion, making it a new unicorn in the AI vision space.

Embodied intelligence is equally red-hot. One bionic humanoid face robot company completed seven funding rounds in two years — with Lei Jun investing twice — targeting reception and companionship scenarios through micro-expressions and flexible synthetic skin, though commercialization under high cost constraints remains a challenge. Meanwhile, the co-founders of LinkedIn and Inflection have partnered to launch a new AI lab and are in talks to raise $100 million, with top-tier capital remaining firmly bullish on the next paradigm.

Industry Leaders Speak: AI Bubble or Dawn of a New Era?

In a 70-minute interview, Jensen Huang expressed strong support for Chinese AI, praised the quality of open-source models, acknowledged that competitors have genuine opportunities, and flatly stated that "AI eliminating half of all jobs is nonsense" — concluding that we are not in a bubble. Elon Musk, meanwhile, offered a vivid analogy: AI is like "a rocket with a 20% chance of exploding" — but it still launches, and he predicts AI will fully surpass humans within five years.

The divergence between these two tech titans precisely reflects the complex picture of the AI industry today — opportunity and risk coexisting, enthusiasm and reason intertwined. Amid the multivariable pressures of a cost-efficiency war, inference cost battles, and Agent entry point competition, AI's most intense phase of competition is only just beginning.