The New Pricing Structure

DeepSeek is the first major AI API provider to introduce structural time-of-day surge pricing, modeled after electricity peak/off-peak billing . Off-peak rates are set at 50% of peak rates, meaning peak-hour API calls cost double . The peak windows (UTC) are reportedly 01:00–04:00 and 06:00–10:00 .

V4-Pro Pricing (per 1M tokens, USD)

V4-Flash Pricing (per 1M tokens, USD)

(Input cache-miss prices for the new tiers are estimated from the confirmed output multipliers. Output jumps are confirmed by multiple sources .)

Cached input tokens remain dramatically cheaper. For V4-Pro, cache-hit input stays at approximately $0.004 per 1M tokens, a 100x discount that incentivizes developers to optimize their caching strategies .

Model Specifications

Both V4 models are text-only, Mixture-of-Experts (MoE) architectures released under the MIT license, making them fully open-weight and self-hostable .

(Source: )

Both models support three reasoning modes: Non-think (fast, intuitive), Think High (logical analysis with chain-of-thought), and Think Max (deep reasoning) .

How DeepSeek Stacks Up Against Competitors

The new prices reshape DeepSeek's competitive positioning. The comparison below uses standard short-context API rates per 1M tokens for all providers.

(Sources: DeepSeek , OpenAI , Anthropic )

vs. OpenAI GPT-5.6 Luna

OpenAI cut GPT-5.6 Luna prices by 80% on July 30, 2026, to $0.20 input and $1.20 output per 1M tokens . The competitive picture flipped dramatically:

• V4-Pro peak output ($3.96) is 3.3x more expensive than Luna ($1.20).
• V4-Flash peak output ($1.26) is still roughly equivalent to Luna's output, just 5% more.
• V4-Flash off-peak output ($0.63) is 47% cheaper than Luna's output, preserving some of DeepSeek's historical cost advantage.

Where DeepSeek was once the clear budget option, it now faces a pricing leader in OpenAI's most affordable tier .

vs. Anthropic Claude Fable 5

Anthropic's flagship reasoning model is priced at $10 input and $50 output per 1M tokens . At those levels, DeepSeek remains the value play:

• V4-Pro peak output ($3.96) is 92% cheaper than Fable 5 ($50).
• V4-Flash peak output ($1.26) is 97.5% cheaper.

Even at their new higher rates, both DeepSeek models remain dramatically more affordable than Anthropic's top-tier offering .

vs. xAI Grok 4.6

xAI's Grok 4.6 pricing was not available from the provided sources, so a direct comparison cannot be made.

What's Driving the Price Increase

DeepSeek cited soaring demand straining inference capacity as the primary reason for the hike . The V4 models gained massive adoption after their preview launch in April 2026, creating GPU capacity bottlenecks that the company could no longer absorb at its previous ultra-low rates .

Several market forces converged:

• Adoption outpacing infrastructure: DeepSeek's combination of 1M-token context, MIT open-weight licensing, and low pricing attracted a flood of developers and enterprises .
• OpenAI's aggressive price cuts: On July 30, 2026, OpenAI slashed GPT-5.6 Luna prices by 80%, putting pressure on DeepSeek's value-for-money positioning . DeepSeek's price increases partially close the gap that OpenAI opened.
• Structural capacity management: DeepSeek described the change as enabling "more flexible workload scheduling," effectively pushing non-urgent traffic to off-peak hours to ease capacity constraints . This is the API industry's first structural time-of-day surge pricing model .

What Developers Should Do

• Shift non-urgent workloads to off-peak hours to benefit from 50% lower rates.
• Optimize prompt caching to take advantage of the ~100x discount on cache-hit input tokens. For V4-Pro, cached input is approximately $0.004 per 1M tokens versus a peak cache-miss input of ~$0.98 .
• Evaluate self-hosting: Both V4 models are MIT-licensed and available on Hugging Face. For high-volume workloads, self-hosting may offer better economics despite the upfront infrastructure cost .
• Re-benchmark against GPT-5.6 Luna: With Luna now at $0.20/$1.20, it may be the cheaper option for output-heavy workloads that don't require DeepSeek's 1M-token context or specific architectural advantages.

Key Takeaway

DeepSeek's price increases mark the end of its era as the overwhelmingly cheapest frontier API provider. The company is now closer to parity with OpenAI's most affordable model on some dimensions while still undercutting Anthropic's premium tier by an order of magnitude. For developers, the combination of peak/off-peak pricing, deep caching discounts, and MIT-licensed open weights means that workload architecture — rather than raw model choice — will increasingly determine the final API bill.

The New Pricing Structure

DeepSeek is the first major AI API provider to introduce structural time-of-day surge pricing, modeled after electricity peak/off-peak billing . Off-peak rates are set at 50% of peak rates, meaning peak-hour API calls cost double . The peak windows (UTC) are reportedly 01:00–04:00 and 06:00–10:00 .

V4-Pro Pricing (per 1M tokens, USD)

V4-Flash Pricing (per 1M tokens, USD)

(Input cache-miss prices for the new tiers are estimated from the confirmed output multipliers. Output jumps are confirmed by multiple sources .)

Cached input tokens remain dramatically cheaper. For V4-Pro, cache-hit input stays at approximately $0.004 per 1M tokens, a 100x discount that incentivizes developers to optimize their caching strategies .

Model Specifications

Both V4 models are text-only, Mixture-of-Experts (MoE) architectures released under the MIT license, making them fully open-weight and self-hostable .

(Source: )

Both models support three reasoning modes: Non-think (fast, intuitive), Think High (logical analysis with chain-of-thought), and Think Max (deep reasoning) .

How DeepSeek Stacks Up Against Competitors

The new prices reshape DeepSeek's competitive positioning. The comparison below uses standard short-context API rates per 1M tokens for all providers.

(Sources: DeepSeek , OpenAI , Anthropic )

vs. OpenAI GPT-5.6 Luna

OpenAI cut GPT-5.6 Luna prices by 80% on July 30, 2026, to $0.20 input and $1.20 output per 1M tokens . The competitive picture flipped dramatically:

• V4-Pro peak output ($3.96) is 3.3x more expensive than Luna ($1.20).
• V4-Flash peak output ($1.26) is still roughly equivalent to Luna's output, just 5% more.
• V4-Flash off-peak output ($0.63) is 47% cheaper than Luna's output, preserving some of DeepSeek's historical cost advantage.

Where DeepSeek was once the clear budget option, it now faces a pricing leader in OpenAI's most affordable tier .

vs. Anthropic Claude Fable 5

Anthropic's flagship reasoning model is priced at $10 input and $50 output per 1M tokens . At those levels, DeepSeek remains the value play:

• V4-Pro peak output ($3.96) is 92% cheaper than Fable 5 ($50).
• V4-Flash peak output ($1.26) is 97.5% cheaper.

Even at their new higher rates, both DeepSeek models remain dramatically more affordable than Anthropic's top-tier offering .

vs. xAI Grok 4.6

xAI's Grok 4.6 pricing was not available from the provided sources, so a direct comparison cannot be made.

What's Driving the Price Increase

DeepSeek cited soaring demand straining inference capacity as the primary reason for the hike . The V4 models gained massive adoption after their preview launch in April 2026, creating GPU capacity bottlenecks that the company could no longer absorb at its previous ultra-low rates .

Several market forces converged:

• Adoption outpacing infrastructure: DeepSeek's combination of 1M-token context, MIT open-weight licensing, and low pricing attracted a flood of developers and enterprises .
• OpenAI's aggressive price cuts: On July 30, 2026, OpenAI slashed GPT-5.6 Luna prices by 80%, putting pressure on DeepSeek's value-for-money positioning . DeepSeek's price increases partially close the gap that OpenAI opened.
• Structural capacity management: DeepSeek described the change as enabling "more flexible workload scheduling," effectively pushing non-urgent traffic to off-peak hours to ease capacity constraints . This is the API industry's first structural time-of-day surge pricing model .

What Developers Should Do

• Shift non-urgent workloads to off-peak hours to benefit from 50% lower rates.
• Optimize prompt caching to take advantage of the ~100x discount on cache-hit input tokens. For V4-Pro, cached input is approximately $0.004 per 1M tokens versus a peak cache-miss input of ~$0.98 .
• Evaluate self-hosting: Both V4 models are MIT-licensed and available on Hugging Face. For high-volume workloads, self-hosting may offer better economics despite the upfront infrastructure cost .
• Re-benchmark against GPT-5.6 Luna: With Luna now at $0.20/$1.20, it may be the cheaper option for output-heavy workloads that don't require DeepSeek's 1M-token context or specific architectural advantages.

Key Takeaway

DeepSeek's price increases mark the end of its era as the overwhelmingly cheapest frontier API provider. The company is now closer to parity with OpenAI's most affordable model on some dimensions while still undercutting Anthropic's premium tier by an order of magnitude. For developers, the combination of peak/off-peak pricing, deep caching discounts, and MIT-licensed open weights means that workload architecture — rather than raw model choice — will increasingly determine the final API bill.