Search

DeepSeek Makes Its 75% API Discount Permanent

On May 22, DeepSeek quietly updated its pricing documentation with a single line that should alarm every Western AI lab: the 75% discount on V4 Pro API pricing will not expire. When the promotional period ends on May 31, the price will be officially adjusted to one-quarter of the original rate - permanently. For a model that already sits at rank 8 on LMRank with a 9.0 score, this is not a promotional stunt. It's a declaration that frontier-class AI inference should cost less than a cup of coffee per million tokens.

The new permanent price? $0.50 per million input tokens and $2.00 per million output tokens. To understand why that matters, you need to see it next to the competition.

The Price Landscape at a Glance

Model Score Input $/M Output $/M Value Ratio
Claude Opus 4.7 9.6 $15.00 $75.00 0.64
GPT-5.5 9.4 $12.50 $50.00 0.75
Gemini Ultra 2 9.2 $10.00 $40.00 0.92
GPT-5 Turbo 9.1 $2.00 $10.00 4.55
Qwen3.7 Max 9.0 $2.50 $7.50 3.60
DeepSeek V4 9.0 $0.50 $2.00 18.0

Value Ratio = Score ÷ Input Price per Million. Higher is better. DeepSeek V4's ratio of 18.0 is the highest of any paid frontier model on the leaderboard.

What "Permanent" Actually Means

DeepSeek's pricing page spells it out explicitly: after the promotional period ends on May 31, 2026, the V4 Pro API price will be "officially adjusted to 1/4 of the original price." That means the current $0.50 per million input tokens and $2.00 per million output tokens are not a limited-time acquisition play. They are the new baseline.

This matters because the entire AI API market has been running on a psychological anchor: frontier models cost $10–15 per million input tokens. Anthropic charges $15 for Claude Opus 4.7. OpenAI charges $12.50 for GPT-5.5. Google charges $10 for Gemini Ultra 2. These prices have been treated as immutable - the cost of doing business at the frontier.

DeepSeek is proving that anchor was fiction. A 9.0-scoring model with strong coding, math, and reasoning capabilities does not intrinsically require $10+ per million tokens to run. DeepSeek's MoE architecture, efficient inference stack, and willingness to operate on thinner margins have shattered the pricing floor.

The Competitive Pressure Is Now Unavoidable

For teams running production AI workloads, the math is brutal. A typical application consuming 100 million input tokens per month would cost:

The gap between DeepSeek V4 and Claude Opus 4.7 is 30x on input pricing. Even against the already-aggressive GPT-5 Turbo, DeepSeek undercuts by 4x. For cost-sensitive deployments - chatbots, content pipelines, coding assistants, data extraction - that difference is the margin between "profitable product" and "expensive experiment."

Western labs now face a choice. They can maintain premium pricing and cede the mid-market to DeepSeek, Qwen, and other Chinese labs. Or they can cut prices and compress their own margins. Either way, the era of $15-per-million frontier APIs is ending.

The Trade-Offs Haven't Disappeared

DeepSeek V4 is not a flawless clone of Claude Opus 4.7 at one-thirtieth the price. Three limitations remain relevant:

Context window: At 128K tokens, DeepSeek V4 is competitive but not class-leading. Gemini Ultra 2, Grok 4.3, and Qwen3.7 Max all offer 1M tokens. For workloads that need to process entire codebases, legal documents, or biomedical literature in a single prompt, DeepSeek's 128K can be a constraint.

Instruction drift: DeepSeek V4 occasionally deviates from strict instructions in ways that Claude and GPT-5.5 rarely do. For applications requiring exact formatting, precise structured output, or rigid guardrails, this can require additional prompt engineering or post-processing.

Ecosystem maturity: Anthropic's tool-use SDK, OpenAI's function calling, and Google's Vertex AI ecosystem are more mature than DeepSeek's surrounding tooling. If your stack is deeply integrated with a Western provider, switching costs are real.

Who Should Switch - and Who Shouldn't

Three profiles should strongly consider moving workloads to DeepSeek V4:

  • High-volume API consumers - If your monthly token bill exceeds $500, the savings are immediate and substantial. A $1,500 Claude Opus bill becomes a $50 DeepSeek bill for input tokens.
  • Cost-sensitive startups - For early-stage products where every dollar of burn matters, DeepSeek V4 delivers frontier-adjacent quality at bootstrap-friendly pricing.
  • Self-hosters exploring open-weight alternatives - DeepSeek V4 is open-weight. Teams already running Llama 4 405B or Qwen3.6 35B A3B should evaluate V4 for its stronger general reasoning at a similarly permissive license.

Conversely, teams requiring absolute maximum reasoning quality, 1M+ token contexts, or deep integration with Claude's creative writing ecosystem should probably stay put. The 0.6-point gap between DeepSeek V4 (9.0) and Claude Opus 4.7 (9.6) is real on the hardest tasks.

Scores at a glance: DeepSeek V4 (9.0) · Qwen3.7 Max (9.0) · GPT-5 Turbo (9.1) · Gemini Ultra 2 (9.2) · Claude Opus 4.7 (9.6). Compare pricing, context windows, and full rankings on the LMRank leaderboard.

The Bottom Line

DeepSeek's decision to make its 75% discount permanent is the strongest signal yet that frontier AI capability is commoditizing faster than the incumbents expected. When a model scoring 9.0 - within striking distance of the absolute best - costs $0.50 per million tokens, the entire pricing psychology of the industry shifts.

The Western labs still hold quality leads at the very top. Claude Opus 4.7 and GPT-5.5 are better models. But they are not 30x better. And for the vast majority of production workloads, "slightly better" does not justify "30x more expensive." DeepSeek just made that calculation explicit, permanent, and impossible to ignore.

At lmrank.com, we track live scores, pricing, and context windows for every major LLM. Follow DeepSeek V4's detail page for updated benchmarks and pricing as the permanent rates take effect.

See also: Cheapest Models