DeepSeek Raises API Prices as OpenAI Makes ChatGPT Text Unlimited

The latest pricing decisions from DeepSeek and OpenAI point to a more segmented AI competition. DeepSeek is charging more for high-volume API workloads while introducing time-based billing; OpenAI is expanding free access to text chat to keep users inside ChatGPT. Together, the moves suggest that the US–China contest is no longer simply about which provider offers the cheapest model. It is also about capacity, distribution, developer adoption, and the economics of running AI at scale.

DeepSeek’s V4 API prices now depend on demand

DeepSeek announced higher prices for its V4-Flash and V4-Pro APIs on August 13. The new schedule began at 16:00 UTC on August 16—August 17 in Beijing—and replaced flat pricing with peak and off-peak rates.

Peak periods run from 01:00–04:00 UTC and 06:00–10:00 UTC. All other hours are off-peak, and peak prices are twice the new off-peak rates.

New DeepSeek V4 rates

Prices below are in U.S. dollars per 1 million tokens:

Model	Period	Cached-input hit	Uncached-input miss	Output
V4-Flash	Off-peak	$0.007	$0.22	$0.66
V4-Flash	Peak	$0.014	$0.44	$1.32
V4-Pro	Off-peak	$0.022	$0.66	$1.98
V4-Pro	Peak	$0.044	$1.32	$3.96

The increase varies by model, token type, cache status, and time of day. Reuters described the overall change as ranging from 50% to 1,100% above previous prices.

The practical effect is largest for workloads that generate large amounts of output during peak hours. V4-Pro output pricing moved from $0.87 per million tokens to $1.98 off-peak and $3.96 at peak. A coding agent generating 10 million output tokens per month would therefore pay roughly $8.70 under the former rate, about $19.80 if all output is off-peak, or about $39.60 if all output falls in peak periods—before input-token charges.

What the change means for developers

DeepSeek’s new structure makes scheduling part of API cost management. Developers running batch jobs, coding agents, evaluations, or other non-urgent workloads can attempt to move traffic into the 17 off-peak hours. Interactive applications, however, may have less flexibility because their requests arrive when users need responses.

The change also makes caching and routing more important. A system that improves cache-hit rates, sends routine work to V4-Flash, or routes selected requests to another provider can reduce exposure to the new rates. For enterprise teams, model choice is increasingly an operational decision rather than a one-time benchmark comparison.

The pricing does not eliminate DeepSeek’s cost appeal, but it does narrow the advantage created by exceptionally cheap inference. The introduction of peak billing is consistent with a provider attempting to manage demand and extract more value from scarce computing capacity. Publicly available material in this reporting does not establish that DeepSeek has reached profitability or provide a confirmed path to it, so that interpretation should remain tentative.

OpenAI is competing on access and habit

OpenAI took a different approach. On August 6, it announced that GPT-5.6 Luna would become the default model for ChatGPT Free and Go users, replacing GPT-5.5. Those users would also receive unlimited everyday text chats and a new Think option for harder questions, subject to abuse-prevention safeguards.

The word unlimited is narrower than it may sound. OpenAI says separate limits still apply to file uploads, image generation, and other tools. Free and Go users also do not receive GPT-5.6 Sol; Luna is the model assigned to their default experience and powers Think for those tiers.

Luna is positioned as the more economical member of the GPT-5.6 family rather than the flagship Sol model. Third-party reporting lists API pricing of $0.20 per million input tokens and $1.20 per million output tokens, but those figures are not confirmed here by the highest-authority sources in the supplied material. The important strategic distinction is clearer: Luna is effectively free for text use inside ChatGPT Free and Go, while API access remains a separate paid product.

Why OpenRouter usage matters

The pricing decisions arrived as Chinese models were gaining significant developer traffic. Bloomberg reported that Chinese models overtook U.S. platforms on OpenRouter in June and accounted for more than 60% of token share in the following month. OpenRouter is a useful indicator of developer model-routing behavior, but it is not a complete measure of global AI usage.

The shift involved more than DeepSeek. July OpenRouter rankings placed Chinese-developed models in all five leading positions by token volume: Xiaomi’s MiMo-V2.5, DeepSeek, MiniMax, Alibaba’s Qwen family, and Moonshot’s Kimi. Other analyses identified DeepSeek as the largest individual vendor on the platform, with a 17.6% share in one June ranking.

That pattern helps explain why pricing remains central. Token-heavy coding and agent workloads can favor models that are inexpensive enough to run repeatedly. At the same time, OpenAI’s response shows that usage can be won through product distribution as well as API price: putting a capable model in front of free users encourages habitual use, feedback, and future conversion to paid products.

Two different commercial pressures

DeepSeek and OpenAI are therefore addressing different sides of the same market pressure:

• DeepSeek is monetizing demand. Peak/off-peak billing can raise revenue during periods of high capacity use and encourage customers to schedule flexible workloads outside those windows.
• OpenAI is defending distribution. Unlimited text chat keeps Free and Go users engaged while preserving limits around more expensive features and access to higher-tier models.
• Developers face more complex optimization. Teams must evaluate not only model quality and list price, but also cache behavior, time-of-day pricing, latency, routing, and the cost of switching providers.
• The market is becoming multi-model. A company may use lower-cost models for repetitive extraction, translation, coding, or agent loops while reserving premium models for tasks where performance, controls, or ecosystem integration matter most.

This is why the AI price war is not necessarily ending even as some prices rise. Competition is splitting into several layers: low-cost inference and open-weight availability on one side, and consumer reach, product integration, premium capability, and enterprise controls on the other.

The broader takeaway

DeepSeek’s increase is a reminder that low API prices can attract enormous usage without guaranteeing permanently low costs. OpenAI’s free Luna rollout is the opposite kind of bet: subsidize a broad consumer experience to strengthen distribution, then differentiate through model tiers and non-text features.

For developers and enterprise buyers, the practical lesson is to avoid treating any single price sheet or benchmark as permanent. Costs can change with demand, model tier, cache status, and time of day. The most resilient architecture is likely to measure actual workloads, keep routing options open, and match each task to the model and access tier that fit its business risk and cost requirements.