DeepSeek Adds Peak Pricing, Grok 4.6 Matches GPT-5.6 Sol, NVIDIA Launches $500B Compute Financing Push

Late on August 13, DeepSeek formally launched V4-Pro across its app, web interface, and API, while announcing a new pricing structure. Starting at 00:00 on August 17, API usage will be priced in peak and off-peak windows based on Beijing time: output will cost RMB 27 per million tokens during peak hours and RMB 13.5 during off-peak hours, versus RMB 6 throughout the day before the change. One day earlier, xAI released Grok 4.6 at the same price as its predecessor—$2 per million input tokens and $6 per million output tokens. Artificial Analysis independently assigned it an Intelligence Index score of 61, tying GPT-5.6 Sol. The world’s cheapest frontier models are beginning to raise prices and write time-of-day usage into their rate cards, while the capability available at the same price point is approaching the top of the leaderboard. On the capital side, NVIDIA signed memoranda of understanding on August 10 with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to establish a compute-financing platform targeting more than $500 billion in third-party capital, turning GPU compute into an underwritable asset class. On August 13, Anthropic’s CFO began pre-IPO investor outreach, with investors reportedly preparing around expectations of a $2 trillion valuation and an October listing. On security, OpenAI split Daybreak into two access tiers for defensive work and red teaming on August 10 and introduced GPT-5.6-Cyber; Z.ai, meanwhile, delayed open-weight and API access for GLM-5.3 until after a security review when it released the model on August 14.

Major Events

DeepSeek Writes Beijing Time into Its Rate Card, Raising Peak Output Pricing to 4.5× the Previous Level

Late on August 13, with almost no advance notice, DeepSeek pushed the final version of V4-Pro to its app, web interface, and API under version DeepSeek-V4-Pro-0813. The model architecture is unchanged from the preview: a Mixture-of-Experts model with 1.6 trillion total parameters and 49 billion active parameters, a 1-million-token context window, and a maximum output of 384,000 tokens. All gains come from targeted post-training focused on long-horizon code modification, environment interaction, and tool use. Terminal Bench 2.1 rose from 72.1 to 87.9, DeepSWE from 12.8 to 62.7, Cybergym from 52.7 to 83.3, and DSBench-Hard doubled to 67.2. On the API side, the model natively supports the OpenAI Responses API and Anthropic-style endpoints, includes adaptations for Codex, and has a concurrency limit of 500 for V4-Pro versus 2,500 for V4-Flash.

The pricing change announced the same day requires enterprises to recalculate workloads transaction by transaction. Starting at 00:00 on August 17, DeepSeek API pricing will be split into two windows based on Beijing time: peak hours from 09:00 to 12:00 and 14:00 to 18:00, and off-peak hours at all other times, with off-peak rates exactly half of peak rates. For deepseek-v4-pro, peak pricing per million tokens will be RMB 0.30 for cache hits, RMB 9 for cache misses, and RMB 27 for output. Off-peak rates will be RMB 0.15, RMB 4.5, and RMB 13.5 respectively. Before the adjustment, the corresponding rates were RMB 0.025, RMB 3, and RMB 6. Even at the off-peak rate, output rises to 2.25 times the previous price and cache hits to six times the previous price; at peak rates, output rises 4.5 times and cache hits 12 times. deepseek-v4-flash is being repriced at the same time, with off-peak rates of RMB 0.05, RMB 1.5, and RMB 4.5, doubled during peak hours. A provider that built much of the domestic open-source narrative around ultra-low pricing has, for the first time, made the hour of invocation itself part of the billing dimension.

At current exchange rates, V4-Pro peak output pricing is approximately $3.8 per million tokens and off-peak output approximately $1.9, still below recent pricing for comparable models but with a materially narrower price advantage. For comparison, Alibaba’s Qwen3.8-Max, launched on August 3, is priced at $2 input and $6 output; Grok 4.6, released on August 12, is priced at exactly the same $2/$6 and scores 61 on the Artificial Analysis Intelligence Index, tying GPT-5.6 Sol and trailing only Claude Opus 5 at 63 and Fable 5 at 62. Grok 4.6 retains Grok 4.5’s 1.5-trillion-parameter base and derives its gains from post-training rather than further scaling, allowing pricing to remain unchanged. Only when prompts exceed 200,000 tokens does the entire request move to the $4/$12 tier. The capability available at the $2 input tier has improved markedly over the past three months, while the cheapest end of the market is beginning to move upward.

The effect of these changes on enterprise bills is not visible in a simple price-comparison table. The same workflow can cost twice as much if it runs between 09:00 and 12:00 rather than overnight. The economics of prompt caching also change because cache-hit pricing has risen the most, requiring enterprises that relied on long system prompts plus caching to revisit their calculations. Capturing the new price differential requires three capabilities: knowing when each team’s usage occurs, moving non-urgent batch workloads into off-peak windows, and switching tasks among multiple models in the same capability tier. Routing requests by task type, cost budget, and time window; setting quotas for each team and project; and recording actual costs by endpoint, user, team, and model are the common foundations for all three. Enterprises running AI traffic through AgentsFlare can respond to a rate-card change by updating routing policy once, without modifying each business application.

Offensive and Defensive Capability Starts Shipping by License: OpenAI Creates Two Access Tiers, Z.ai Delays GLM-5.3 Open Weights

On August 10, OpenAI published *Expanding Daybreak as the Cyber Defense Window Narrows*, splitting its Daybreak security program into two access tiers. Daybreak Blue targets general defensive use, providing access to models including GPT-5.6 Sol with safeguards tailored for authorized defensive security work. Daybreak Red is intended for vulnerability research, exploit validation, and security testing, and introduces GPT-5.6-Cyber. OpenAI’s stated comparison is that GPT-5.6-Cyber completes 95.0% of advanced cybersecurity requests involving exploit-chain development, authentication bypass, privilege escalation, and similar scenarios, compared with 1.5% for general GPT-5.6 Sol and 57.3% for GPT-5.5-Cyber. OpenAI assesses GPT-5.6-Cyber as High for cybersecurity capability under its Preparedness Framework but below the Critical threshold. Access requires identity verification and monitoring, and beginning September 1 all individual Daybreak accounts must use hardware security keys. OpenAI also cites two categories of real-world results: discovery of two V8 vulnerabilities that could be chained, associated with CVE-2026-15903, and more than 400 vulnerabilities capable of leading to privilege escalation in an operating-system kernel.

Four days later, on August 14, Z.ai released GLM-5.3 under a coding-and-cyber-defense positioning, based on a 743-billion-parameter model. Z.ai says the model improves on benchmarks including AutomationBench, GDPval-AA v2, and CyberGym. What changed is the release model. GLM-5.2 launched its weights and API together; for GLM-5.3, both open weights and API access have been deferred until a security review is complete. For now, it is available only through GLM Coding Plan and ZCode, and multimodal capabilities are not included in this release. A vendor that has used rapid open releases as a core competitive strategy has voluntarily inserted a pre-release gate.

Cyber capability is shifting from a general capability delivered together with the model into a product that is separately authorized, separately audited, and subject to tighter access. The direct enterprise impact has two layers. The first is procurement: internal red teams and vulnerability-research groups that want access to this class of capability will need to go through identity verification, hardware-key requirements, and usage monitoring, meaning the contractual terms differ from ordinary API procurement. The second is the threat model. Once a vendor publicly states that its model completes 95% of advanced offensive-security requests, the practical difference between defensive and offensive use increasingly depends on who obtains authorization first. Z.ai has given no date for when GLM-5.3’s open weights will be released, beyond saying access will follow the security review. During such an open-ended window, any roadmap that treats those weights as a guaranteed self-hosting option should maintain a fallback path to a commercial API.

NVIDIA Turns GPUs into an Underwritable Asset Class

On August 10, NVIDIA announced that it had signed memoranda of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to establish independent compute-infrastructure financing platforms targeting more than $500 billion in third-party capital for data center construction serving frontier labs, enterprises, and AI clouds. The announcement describes NVIDIA compute as an investable asset on the grounds that it delivers the lowest cost per token, can be reused across models and workloads, can be transferred among customers and operators, and can have its useful life extended through continued CUDA updates. Jensen Huang’s formulation is that, in AI, compute is revenue. One qualification is important: all six documents are currently memoranda of understanding, and definitive agreements have not yet been signed.

The arrangement creates a financing channel for compute that is separate from the purchase and sale of chips. Historically, data center funding has come from cloud-provider capital expenditure, sovereign wealth funds, or project finance, with the buyer’s balance sheet determining construction scale. If GPUs can be underwritten in the same way as toll roads or data center real estate, investors will focus on utilization rates and cash flows from long-term leases, allowing construction to expand without being constrained by the cash flow of a single buyer. For enterprise procurement, the implications cut both ways. Over the medium term, more compute supply and more rental channels should sustain downward pressure on inference pricing. At the same time, long-term capital and fixed-income-style return expectations will increasingly sit behind compute assets, pushing lease terms, committed-use obligations, and default conditions closer to infrastructure contracts. The flexibility of short-term rentals may not remain this inexpensive indefinitely.

A structurally different transaction took place in the same week. On August 11, IBM and Together AI signed an approximately $240 million multi-year agreement under which IBM will deploy an inference cluster of roughly 2,000 NVIDIA HGX B300 systems in its own cloud. The cluster is scheduled to come online in the United States in Q1 2027, with Together AI using the capacity to provide inference services for open-source models including DeepSeek, MiniMax, and Kimi. A legacy cloud provider is contributing data center capacity and capital expenditure in exchange for hosting an open-source-model inference provider. Both parties are betting on the same proposition: for security and control reasons, a meaningful share of enterprise workloads will prefer to run on open-weight models rather than closed APIs. The capacity will not come online until 2027, by which time comparative prices and the model landscape may look very different.

Anthropic Begins Pre-IPO Outreach, with Investors Preparing Around a $2 Trillion Valuation

On August 13, Anthropic CFO Krishna Rao began early discussions with prospective investors in what amounts to pre-IPO investor education rather than a formal roadshow. Multiple reports say investors are preparing around expectations of a $2 trillion valuation and an October Nasdaq listing. At that level, Anthropic would surpass SpaceX’s $1.77 trillion valuation and become the largest IPO in history. The reports also state that company executives have not settled on an internal valuation target; the $2 trillion figure reflects investor expectations rather than company guidance. Relative to the $965 billion valuation in its Series H round three months earlier, $2 trillion would imply roughly a doubling. The expectation is being supported by the company’s revenue trajectory: reports say annualized revenue could reach $100–120 billion by year-end, versus $47 billion disclosed in May. The company reportedly filed a confidential registration statement with the SEC on June 1, with Morgan Stanley, Goldman Sachs, and JPMorgan as underwriters. Separate reports during the same period said OpenAI had pushed its own listing to 2027.

For enterprise buyers, the relevance lies in post-listing pricing behavior. A private company can spend years pricing below cost to gain enterprise penetration; once quarterly reporting begins, gross margin becomes a public metric subject to repeated scrutiny. Part of the funding behind the model-layer price war over the past year has come from patient private-market capital, and that assumption changes after listing. When negotiating multi-year committed-use agreements, enterprises should pay close attention to the duration of price-protection clauses and to the vendor’s unilateral rights to change model versions and retirement schedules. OpenAI’s August 10 removal of the gpt-5.2-chat-latest and gpt-5.3-chat-latest snapshots from the API, following a standard three-month notice period, illustrates the same underlying point: the lifecycle of model identifiers is defined by the vendor.

There is also one unconfirmed development. Fortune and Israeli outlet Calcalist reported on August 13 that Anthropic is in talks to acquire Israeli inference-optimization company Decart AI for approximately $6 billion, which would be Anthropic’s largest acquisition to date if completed. Decart’s core product is an inference-optimization stack called DOS, which the company claims can improve inference speed by up to 8× on the same hardware. Neither company has confirmed the talks, the reports state that negotiations could fall apart, and the 8× figure comes from the company rather than an independent reproduction. If a deal is ultimately completed, it would suggest that model vendors are beginning to internalize inference efficiency as a proprietary asset, a layer where progress has so far come primarily from chipmakers and the inference-framework community.

Upstream Compute Signals: Packaging Capacity and HBM Yields Both Ease, While Labor Disputes Become a New Variable

Three upstream developments over the past week all point to the supply side. On manufacturing, TSMC reported on August 10 that July consolidated revenue was approximately NT$467.6 billion, up 5.6% month over month and 44.7% year over year, while cumulative January–July revenue rose 37% year over year. The company also raised the upper end of its 2026 capital-expenditure range to $60–64 billion, while high-performance computing accounted for 66% of Q2 revenue. On the same day, reports said TSMC was considering acquiring two AUO panel fabs in Taichung for more than NT$30 billion for fan-out panel-level packaging, which moves packaging from circular wafers to square panels so that more chips can fit in the same area at lower cost. TSMC’s approximately 104-hectare Longtan Phase 3 expansion has also reportedly restarted, with plans for two 1.4nm fabs and one panel-level packaging facility. The current delivery bottleneck for AI hardware lies primarily in advanced packaging rather than wafer capacity. The speed at which new packaging capacity comes online therefore directly affects delivery times for Blackwell and Rubin systems and whether system-level premiums can narrow.

Memory is showing a signal in the opposite direction from recent months. According to TrendForce’s synthesis of Korean media reports, Samsung’s HBM4 yield has improved from below 60% at the start of mass production in February to around 80%. Q3 HBM4 revenue is planned to more than triple quarter over quarter, HBM4 is expected to account for more than 60% of Samsung’s HBM sales in the second half, and the company is targeting roughly 38% HBM market share by year-end. HBM has long been dominated by SK hynix; if a second supplier catches up on yield, NVIDIA gains bargaining power in sourcing HBM for the Rubin platform, potentially easing the memory-price pressure seen over the past several quarters next year. The new variable is labor. The fifth round of SK hynix’s 2026 wage negotiations failed on August 4, with the dispute centered on paying part of performance bonuses in company stock, and approximately 2,500 employees across functions formed a unified union on August 13. If the dispute escalates, the first impact could fall on the production ramp for next-generation HBM. The same company also restarted its second Dalian NAND fab on August 11 after a nearly four-year pause, adding approximately 50,000 wafers per month to bring combined capacity to 150,000 wafers, with mass production scheduled for the first half of 2027—an indicator of improving enterprise SSD demand.

On the rental side, CoreWeave and Nebius reported Q2 results on the same day, August 12. CoreWeave posted revenue of $2.58 billion, up 112% year over year, and a backlog of $104 billion, up 246%. Nebius AI cloud revenue reached $575 million, up 514% year over year, while new contracts signed during the quarter reportedly quadrupled sequentially. Market signals suggest these newer cloud providers are pricing GPU rentals as much as 66% below the three major public clouds. Backlog represents contracted but not yet recognized revenue; at $104 billion, the scale suggests enterprises and labs are locking in compute for the next two to three years rather than renting month to month. The current price gap between the major clouds and newer GPU clouds is one of the most immediate sources of inference-cost elasticity, provided workloads can move across suppliers.

Special Insight: Pathway Pushes Per-Task Inference Cost to $0.0007 with a 150M-Parameter Model

Pathway was founded in 2023 by Zuzanna Stamirowska, Jan Chorowski, and Adrian Kosowski. It began with an open-source streaming data processing and RAG engine, with customers including NATO, La Poste, and CMA CGM, before shifting toward foundation-model architecture research in late 2024. On August 11, the company published results for its 150-million-parameter BDH-CQ model on the public ARC-AGI-1 evaluation set: pass@2 of 29.5% at an inference cost of $0.0007 per task, approximately one-eleventh the cost of GPT-5.6 Luna at the Low setting, using pricing after OpenAI’s July 30 reduction. On August 13, the company announced an extension seed round at a $500 million valuation, bringing cumulative seed funding to $30 million. Investors include TQ Ventures, Red Bridge Ventures, and Wilson Sonsini’s investment arm, while angel investors include Databricks Chief AI Scientist Jonathan Frankle.

Architecturally, BDH-CQ does not generate an intermediate textual chain of thought. Instead, it repeatedly revises the answer in a recurrent latent state, and this is the primary source of its cost advantage. The numbers need to be read carefully. On the same benchmark, GPT-5.6 Luna at Low scores 34.2, above BDH-CQ’s 29.5. The result therefore means slightly lower accuracy at roughly an order of magnitude lower cost; the small model has not surpassed the larger model in capability. The open questions are equally clear. ARC-AGI-1 is a public evaluation set rather than a hidden test set, leaving room for targeted optimization, and there is not yet an independent reproduction from an evaluator such as Artificial Analysis. Whether the cost curve demonstrated at 150 million parameters can persist into the billion- to hundred-billion-parameter range is also unverified externally, although the company argues that the architecture follows Transformer-like scaling laws. Since shifting into this new direction, Pathway has not disclosed paying customers.

Pathway belongs on the watchlist because of where it sits in the stack. Model-layer price competition has already compressed into the $2 input tier; the next phase of cost reduction must come either from hardware and packaging or from reducing compute per token at the architecture layer. The former is already happening on a quarterly cadence. If the latter receives independent validation, it would affect every workload that needs high-volume, low-latency inference. What Pathway has today is a transparent methodology and a self-reported result that has not yet been reproduced. The next questions are whether third-party evaluators pick it up and whether the architecture still holds as parameter count scales.

Overall Assessment

The number of pricing dimensions at the model layer is increasing. DeepSeek has added time of day; Grok 4.6 and Gemini use whole-request tier jumps once context exceeds a threshold; Anthropic and OpenAI define effective pricing windows through promotional periods and snapshot retirements; and OpenAI has now made offensive and defensive cyber capability a separately authorized tier. A year ago, comparing model cost required only two numbers: input and output. Today, the actual price of the same model on the same day depends on when it is called, how long the context is, how often the cache hits, whether a promotion is active, and which level of authorization the enterprise has obtained. Rate-card comparison is therefore shifting from a one-time model-selection exercise into a continuous calculation based on actual usage distribution.

Two capital-market developments point in the same direction: compute and model companies are moving from venture-capital balance sheets onto public-market and long-duration-fund balance sheets. NVIDIA and six asset managers are turning GPUs into an underwritable asset class, while Anthropic is reportedly preparing for an October listing around a $2 trillion valuation. The consequence is that supply should become more abundant, while pricing becomes more constrained by return requirements. Upstream signals from the past week support the first half of that conclusion: Samsung’s HBM4 yield has reportedly reached 80%, TSMC is expanding panel-level packaging capacity, and new GPU-cloud providers have pushed backlog above $100 billion while taking workloads at discounts of up to 66%. Enterprises can be relatively confident that there will be more rentable compute and more supply channels over the next two to three years; what they cannot know is which provider will remain cheapest. In that environment, keeping admission rules, routing policy, cost attribution, and audit chains in an enterprise-controlled layer means that changes in vendor architecture and rate cards affect configuration effort rather than business logic.