Overview
On November 12 and 13, a set of consistent signals emerged across the large-model industry at home and abroad: the centre of competition is moving from the strongest single capability toward unit output cost and dispatch efficiency. According to IT Home and Wallstreetcn, Google launched Gemini 4 Hydrogen, dropping the Flash naming and positioning it as the efficiency tier for everyday agentic execution, priced at 1.5 US dollars per million input tokens and 7.5 dollars per million output tokens across the full one-million-token context with no segmented billing. It scores 51 on the Artificial Analysis Intelligence Index at the default high-thinking level, slightly above the flagship Argon’s 50, and reaches 54 at the highest setting, while completing tasks about 2.5 times faster than the previous Flash generation with substantially fewer fabricated statements. An accompanying Orchestrator switches dynamically between Carbon and Hydrogen within a single session, with the company claiming roughly 24% lower quota consumption and 19% faster execution.
Domestic moves in the same window point toward building market mechanisms for model distribution and invocation. Per CNR, Baidu World 2026 unveiled a model-trading platform that admits models through fingerprint comparison, capability and usability testing, and assigns credit ratings on service quality, fulfilment, security compliance and commercial performance; the company says that within a month of launch the platform hosted more than 100 supply resources, with over 50 million daily calls and more than 700 billion daily tokens. On the infrastructure side it released a 1,024-card supernode. ByteDance’s Seedance 2.5 flash cut 720P text- and image-to-video generation to about 0.99 yuan per second, roughly 34% below the flagship. CCTV Finance reported that e-commerce platforms opened token zones selling more than 20 trillion tokens in a week, AI office monthly passes were promoted at near-zero prices, and national online services retail sales exceeded 5.6 trillion yuan in the first three quarters — tokens are becoming a retail product category.
Capital is repricing the distribution layer too. Per IT Home and TechCrunch, payments infrastructure firm Stripe acquired model-routing platform OpenRouter for roughly 7.5 to 8 billion US dollars; the platform aggregates over 80 providers and more than 400 models for over 10 million developers. A leading foundry’s board approved up to 1 billion US dollars of investments across two frontier model companies, and OpenAI’s third-quarter revenue topped 10 billion US dollars for the first time. Together these threads suggest the scarce link in the AI value chain is migrating from training compute toward whoever decides which model gets called.
Background & Interpretation
1. Technical Evolution and Context
For the past two years, model competition was measured on capability leaderboards: parameter count, context length, reasoning scores and multimodal coverage. As flagship capabilities converge and open-weight models catch up quickly, raw capability leadership alone struggles to sustain pricing power. At the same time, agentic applications scale invocation frequency from human-conversation levels to machine-driven automatic execution, multiplying single-call cost by millions of repetitions; cost elasticity becomes the deciding factor in whether a product can be deployed at scale. The technical response is tiering: high-reasoning models handle long-horizon complex tasks, efficiency-tier models absorb high-frequency routine execution, and a routing and orchestration layer allocates tasks dynamically between them. This is not simple discounting but turning model selection itself into a built-in cost-optimisation layer, forming a complete cost-reduction stack alongside edge compression models, quantised inference and cache reuse.
2. Core Drivers and Underlying Mechanism
Efficiency-tier models deliver low price and usable quality through three mechanisms. First, distillation, quantisation and sparse activation cut compute per inference, while flat pricing across the full context avoids unpredictable costs from segmented long-context billing. Second, orchestrator-style dynamic dispatch assigns tasks by difficulty to different tiers, bringing average cost close to the lightweight tier while still allowing flagship calls for hard tasks — in essence trading routing precision for overall cost-efficiency. Third, structured output on the decision side, such as emitting confidence scores rather than generating token by token, further compresses cost in high-frequency judgement scenarios. Trading and routing platforms solve the other half: with hundreds of models available, information asymmetry and selection cost rise, so admission testing, continuous monitoring and credit ratings convert service quality into comparable metrics. Whoever controls routing and usage-based billing controls the entry point to model distribution — the underlying reason capital is willing to pay a high valuation for a routing platform.
3. Market Structure and Industry Chain Impact
For model vendors, a narrowing capability gap shifts competition from leaderboard scores to fulfilment, stability and unit price, with pricing transparency and security-compliance records entering procurement evaluation; the room to charge a premium on parameter scale alone shrinks. For cloud and compute providers, a rising share of inference load changes cluster demand — HBM capacity, energy efficiency and interconnect bandwidth gain importance while peak training FLOPS lose marginal weight; a foundry taking stakes in two model companies also reflects upstream attempts to bind demand certainty with capital. For application developers, multi-model routing reduces single-vendor lock-in but introduces cross-platform consistency, observability and data-compliance costs. For payments and infrastructure firms, metering model calls is forming a new clearing layer analogous to cloud-resource billing. On the consumer side, tokens becoming a retail category and AI services reaching mass users through low-priced monthly passes show demand moving from early-adopter trials to scaled, usage-based consumption, which in turn requires more granular cost accounting on the supply side.
Implications & Outlook
Near term, watch whether efficiency-tier pricing holds and whether the capability boundary between flagship and efficiency tiers stays stable; if the gap keeps narrowing, price pressure will transmit to the mid and high end. Medium-term variables include whether routing platforms’ billing standards and neutrality gain industry acceptance, whether credit ratings and admission testing become de facto standards, and the evolution from per-token toward per-task or per-outcome billing. Long term, the industry may form a three-layer structure of foundation models, routing distribution and vertical applications, with value concentrating in the middle layer that controls invocation decisions and metering settlement; training compute remains strategic, but its excess returns are gradually diluted by the cost-optimisation layer. Risks include low-price competition squeezing model vendors’ R&D budgets and routing concentration creating new single-point dependency and data-compliance challenges.
Takeaways for Industry Participants
Model vendors should turn verifiable fulfilment, stability and security-compliance records into comparable metrics for buyers rather than publishing leaderboard scores alone, and build task-level cost accounting to avoid volume growth at a loss under flat pricing. Application firms should adopt multi-model routing architectures while building invocation logging, outcome evaluation and data-boundary management so cost optimisation does not come at the expense of observability. Cloud and compute providers should tune clusters for inference workloads, focusing on energy efficiency and memory bandwidth rather than simply stacking training compute. For buyers, treating model services as replaceable inputs and specifying service quality and metering terms in contracts is more robust than locking into one vendor; when judging industry direction, pricing structures, routing share and invocation growth carry more information than any single model release.
This column compiles industry information and shares technical perspectives; it does not constitute investment advice.