AI Inference Price War Explained [What It Changes]

The inference price war is the accelerating competition among AI providers to cut the cost of running large language model queries — measured in dollars per million tokens — with each major reduction forcing rivals to respond or lose market share.

The most recent escalation: DeepSeek made a 75% price cut on its flagship V4-Pro model permanent on May 22, 2026, after originally positioning it as a promotion. That permanence signal changes how businesses should think about AI infrastructure costs — not as a one-time discount but as a new structural floor.

TL;DR: On May 22, 2026, DeepSeek announced that the 75% promotional discount on V4-Pro — originally set to expire May 31 — would become permanent. Input tokens now cost $0.435 per million (down from $1.74), output tokens $0.87 per million (down from $3.48), and cached input tokens $0.003625 per million. This is frontier-class model capability at a quarter of the launch price. The move directly pressures OpenAI, Anthropic, and Google to respond with their own pricing adjustments and accelerates the trend toward AI-powered features in SMB software becoming cheaper or free. The limit: data-residency and compliance concerns prevent direct use of Chinese-hosted models in healthcare and legal workflows.

Key Takeaways

• DeepSeek V4-Pro input pricing dropped 75% permanently — from $1.74 to $0.435 per million tokens, as announced May 22, 2026 (Engadget).

• Output tokens dropped from $3.48 to $0.87 per million, and cached input tokens fell to $0.003625 per million (apidog).

• The original 75% promotional discount was set to expire May 31 but was made permanent before that date (Engadget).

• The permanent price cut directly pressures OpenAI, Anthropic, and Google to respond, accelerating the price compression across all frontier models.

• For SMBs, the practical outcome is that AI-powered features inside CRMs, phone systems, and back-office tools will get cheaper or move toward inclusion in base subscriptions.

• Compliance limit: Data-residency and regulatory concerns make direct use of Chinese-hosted models (DeepSeek's infrastructure) unsuitable for healthcare, legal, and financial workflows subject to US data-protection requirements.

What Happened: The DeepSeek V4-Pro Price Event

The Original Launch and the Promotion

DeepSeek launched V4-Pro with a promotional 75% discount applied from the start. The promotion was originally set to expire May 31, 2026. The expectation in the market was that prices would revert to the full rate after that date.

On May 22, 2026, DeepSeek announced the discount would not expire — it would become the permanent price. According to Engadget, DeepSeek permanently reduced the price of its flagship V4 model by 75 percent. That announcement removed the expiration date from the market's mental model and reframed DeepSeek's positioning from a promotional launch to a structural price point.

The Numbers: Before and After

Source: apidog.

At $0.435 per million input tokens, DeepSeek V4-Pro is priced below most mid-tier commercial models and dramatically below flagship US-hosted models, according to Apidog. For context: running one million input tokens through a typical frontier US model costs several dollars at current rates. The gap is not marginal — it is structural. According to DeepSeek's official API pricing page, V4-Pro is permanently priced at $0.435 per million input tokens and $0.87 per million output tokens — a cost floor that makes it cheaper than most mid-tier Western models and forces comparable providers to respond or cede price-sensitive API customers.

The Mechanism: Why Price Cuts Like This Spread

How Inference Pricing Works

Every AI model query consumes tokens — units of text that go in (input) and come out (output). Providers charge per million tokens processed. The cost of running a model at scale is a function of compute (GPUs), energy, infrastructure overhead, and desired margin. When one provider cuts price significantly, they are either:

1. Operating at a lower cost base (more efficient training, cheaper infrastructure, government subsidy), or

2. Accepting lower margin to capture market share and force competitors to respond.

DeepSeek's position likely involves both. The permanent price cut signals that DeepSeek can sustain V4-Pro operations at the reduced rate — this is not a loss-leader desperation move but a deliberate market positioning. The fact that it was made permanent rather than extended-then-lapsed underscores that framing.

The Competitive Response Loop

When DeepSeek prices V4-Pro at $0.435/M input, it creates pressure on every other provider whose comparable model is priced higher. OpenAI, Anthropic, and Google must either:

• Match or beat the pricing on comparable-quality models, or

• Differentiate on non-price dimensions (safety, compliance, ecosystem integration, latency).

According to Apidog, this move resets the cost floor for what frontier-class AI capability costs to operate. According to Apidog, the permanent discount positions DeepSeek V4-Pro at $0.435 per million input tokens — a price point that forces OpenAI, Anthropic, and Google to respond or cede cost-sensitive API customers. According to The Decoder, V4-Pro output tokens at $0.87 per million are at least 34x cheaper than GPT-5.5's $30 per million output rate — a gap wide enough to reshape cost planning for any token-intensive agentic workload. The price war timeline shows an escalation pattern across 2025-2026:

What This Means for Business AI Costs

The Direct Effect: Cheaper API Runs

For businesses that query AI models directly through APIs — for document processing, customer-response automation, data extraction, or content generation — the per-query cost drops when they can use a lower-priced model of equivalent quality. Not every use case can use DeepSeek (see compliance limits below), but the pricing pressure it creates eventually reaches US-hosted models too.

Teams already running automated workflows — such as those built through the US Tech Automations platform — will find that model selection has become a more frequent optimization decision. When a new price floor emerges, the question is whether you can swap the model underlying your workflow without rebuilding the workflow itself. That is the operational advantage of model-agnostic pipeline architecture: the business logic stays the same while the model underneath gets cheaper.

The Indirect Effect: Cheaper SaaS AI Features

Most SMBs do not pay for AI tokens directly — they pay for software that uses AI under the hood (CRM systems with AI summaries, phone systems with AI call scoring, accounting tools with AI categorization). When inference costs drop, those SaaS providers either:

• Expand the AI features included in base plans, or

• Launch AI tiers at lower price points than previously viable.

This is the channel through which the inference price war reaches businesses that have never seen an API invoice. According to Engadget, the move pressures OpenAI, Anthropic, and Google pricing across the board — and that pressure flows downstream to every SaaS product built on top of those APIs.

The Compliance Limit: When You Cannot Use DeepSeek Directly

DeepSeek operates Chinese infrastructure. For businesses subject to US data-protection requirements, the inference price war has a meaningful asterisk:

For regulated verticals, the value of the DeepSeek price cut is indirect: it pressures US-hosted providers to reduce prices, and those US-hosted providers are the ones compliant businesses can actually use. The downstream effect is real even if the direct use is not.

Signal vs Speculation

What is documented fact (as of June 2026):

• DeepSeek V4-Pro launched with a 75% promotional discount that was subsequently made permanent (apidog).

• Permanent price cut announced May 22, 2026: input $0.435/M, output $0.87/M, cached input $0.003625/M (apidog).

• The cut directly pressures OpenAI, Anthropic, and Google pricing (Engadget).

• Data-residency concerns limit direct use in regulated US verticals.

Our read (forward-looking interpretation):
The inference price war is not over — it is in its early innings. DeepSeek's permanent cut removes the safety valve that allowed competitors to wait for the promotion to expire. The most likely 12-24 month trajectory: US-hosted frontier models will compress pricing on at least one tier in response, SaaS products will absorb the savings through expanded AI feature sets rather than price cuts (margin normalization), and the compliance differentiation between US-hosted and foreign-hosted models will become a product feature rather than just a regulatory checkbox. Businesses that build now on model-agnostic workflow infrastructure are better positioned to capture each successive price reduction without rebuilding their automation stack. The risk: if inference costs drop faster than business adoption cycles, the competitive advantage of early automation builds may compress — meaning the operational moat is in the workflow design and data integration, not in being first to pick a cheap model.

FAQ

What is the inference price war?

The inference price war is the competition among AI model providers to cut the cost of running large language model queries, measured in dollars per million tokens. Each major price cut — like DeepSeek's permanent 75% reduction on V4-Pro — forces competitors to respond with their own cuts or risk losing API customers to cheaper alternatives.

What did DeepSeek actually change on May 22, 2026?

DeepSeek announced that the 75% promotional discount on V4-Pro — originally set to expire May 31, 2026 — would be made permanent. Input tokens dropped from $1.74 to $0.435 per million, output from $3.48 to $0.87 per million, and cached input to $0.003625 per million. This transformed a time-limited promotion into a structural price point.

Can my business use DeepSeek V4-Pro directly?

It depends on your industry and data type. General SMBs without regulated data can evaluate it with standard vendor review. Businesses subject to HIPAA, GLBA, or government contracting requirements should not use DeepSeek for regulated data without legal and compliance review — Chinese-hosted infrastructure creates data-residency exposure under US frameworks.

How does the inference price war affect businesses that don't use AI APIs?

Indirectly but materially. SaaS tools (CRMs, phone systems, accounting software) built on top of AI APIs absorb lower inference costs over time, either by expanding AI features in base plans or launching lower-priced AI tiers. The price compression in the API market flows downstream to end-user software over 6-18 months.

What is the right business response to the inference price war?

Two moves: (1) If you use AI APIs directly, audit your current model spend and evaluate whether a model swap is viable for your use case — the compliance table above is your starting framework. (2) If you use SaaS AI tools, watch for pricing tier changes from your vendors over the next 6-12 months. For teams building automation workflows, the structural response is model-agnostic architecture — design the workflow around the task, not around a specific model, so you can swap models as prices fall. The small businesses implications, marketing agencies breakdown, and accounting firms analysis cover the vertical specifics.

The Structural Shift Under the Price Numbers

The inference price war matters less as a one-time cost reduction and more as a signal that commodity inference is coming — the same way cloud storage went from a premium service to a line item too small to optimize.

When inference approaches commodity pricing, the competitive differentiation in AI moves entirely to:

1. Data access and integration (which business data the model can see)

2. Workflow design (how reliably the model's outputs plug into actual operations)

3. Compliance architecture (which regulated data can legally run through which models)

That is the layer where US Tech Automations works with business operations teams — not on which model to use, but on how to connect model outputs to the actual systems of record where business decisions happen.

The accounting firm, small business, and marketing agency specific implications of the inference price war are covered in the spoke posts linked in the FAQ above. For the broader question of how to build automation workflows that remain cost-efficient as model pricing continues to fall, the agentic workflows platform is the practical starting point.