DeepSeek-V4-Pro vs Kimi K3: Price and Output Length

All figures below are from the China AI Hub database, last verified 2026-09-22. Where a field is not publicly disclosed, we say so rather than estimating.

At a glance

Dimension	DeepSeek-V4-Pro	Kimi K3
Status (database)	deprecated	active
Context window	1,048,576 tokens	1,048,576 tokens
Maximum output	393,216 tokens	1,048,576 tokens
Input price (per 1M)	$0.66	$3.00
Output price (per 1M)	$1.98	$15.00
Open weights	Yes (MIT)	Yes (Kimi K3 License)
API available	Yes	Yes
Reasoning	Yes	Yes
Coding	Yes	Yes
Vision / Video	No	Yes / Yes
Tool calling	Yes	Yes
Agent capability	Not listed	Yes
Benchmark records in DB	4	6

Pricing

DeepSeek-V4-Pro lists $0.66 input / $1.98 output per 1M tokens. Kimi K3 lists $3.00 input / $15.00 output — roughly 4.5x the input price and 7.5x the output price. Both are pay-as-you-go API prices from official pages. Because output price dominates reasoning workloads, the per-task cost gap can be larger than the input-price gap suggests.

Context and output

Both list a 1,048,576-token context window. Kimi K3 lists a 1,048,576-token maximum output — the full window — versus 393,216 for DeepSeek-V4-Pro. For extreme-length generation tasks this is the largest listed-output difference in our database.

Capabilities

Both list reasoning, coding and tool calling. Kimi K3 additionally lists vision, video, structured output and agent capability. DeepSeek-V4-Pro lists function calling and structured output, and no vision/video. Agent capability is listed for Kimi K3 and not for DeepSeek-V4-Pro.

Openness and licensing

Both are open weight, but the licenses differ materially. DeepSeek-V4-Pro is MIT. Kimi K3 uses the Kimi K3 License: permissive MIT-style terms, with additional obligations — Model-as-a-Service operators above $20M aggregate revenue must sign a separate agreement, and products above 100M MAU or $20M monthly revenue must display “Kimi K3” in the UI. Read the license text before commercial self-hosting.

Benchmarks and verification

The database holds 4 benchmark records for DeepSeek-V4-Pro and 6 for Kimi K3, each labeled with source type and version. Different benchmarks cover different models; counts are not scores. We do not rank the two models on this page.

Benchmark comparability is limited: the two models’ scores come from different test configurations — different benchmark versions, evaluation harnesses and tool setups — and several benchmarks cover only one of the two. Do not rank on headline scores alone.

Why each difference matters

China AI Hub analysis indicates the following per-dimension implications, drawn from the listed facts above.

• Cost: DeepSeek-V4-Pro has an advantage at ~4.5x lower input and ~7.5x lower output price. Because reasoning workloads multiply output tokens, the output-price gap has a larger real-world effect on per-task cost than the input gap alone suggests.
• Long-context: both list a 1M-token input window; Kimi K3 has an advantage in extreme-length generation because it lists the full-window 1M-token maximum output.
• Reasoning: both list reasoning, so neither has a listed advantage on this axis — the output-price differential, not capability, drives the reasoning-workload decision.
• Coding: both list coding and tool calling; DeepSeek-V4-Pro’s price is more relevant for high-volume code generation, while Kimi K3 adds vision/video for multimodal coding.
• Agent: Kimi K3 is more relevant when agent orchestration is core — it lists agent capability where DeepSeek-V4-Pro does not.
• Deployment / license: both are open-weight and self-hostable, but the licenses differ materially. MIT (DeepSeek) is permissive with no thresholds; the Kimi K3 License has a limitation for large operators — Model-as-a-Service above $20M revenue needs a separate agreement, and products above 100M MAU / $20M monthly revenue must display attribution.
• API: both expose an API; DeepSeek-V4-Pro is listed deprecated, which is the operational limitation to verify before committing.

Trade-off summary

• Price: DeepSeek-V4-Pro lists substantially lower prices on both input and output.
• Output length: Kimi K3 lists a 1M-token maximum output, the highest in our database.
• Media and agent: Kimi K3 lists vision, video and agent capability; DeepSeek-V4-Pro does not.
• License: MIT (DeepSeek) versus Kimi K3 License with revenue-triggered obligations.
• Status: DeepSeek-V4-Pro is listed as deprecated; Kimi K3 as active.

Choose by workload: long-form generation and media favor Kimi K3’s listed capabilities; price-sensitive text workloads favor the DeepSeek pricing, with the caveat of its deprecated status. Verify current prices and status on the official pages before committing.

Decision context

China AI Hub analysis indicates the following decision-context implications, drawn from the listed facts above.

For API developers. Price is the sharpest split: DeepSeek-V4-Pro lists ~4.5x lower input and ~7.5x lower output price, which compounds on reasoning workloads that multiply output tokens. Both list tool calling; Kimi K3 additionally lists structured output. DeepSeek-V4-Pro is listed deprecated, which is the operational caveat to verify before committing. Latency is not publicly documented for either model on this page.

For self-hosting. Both are open weight but under different licenses — MIT (DeepSeek-V4-Pro, no revenue thresholds) versus the Kimi K3 License (a separate agreement above $20M revenue, and UI attribution above 100M MAU / $20M monthly revenue). Hardware requirements and quantization are Not publicly documented on this page.

For coding agents. Both list coding and tool calling. Kimi K3 lists agent capability where DeepSeek-V4-Pro does not, and adds vision/video for multimodal coding. Both list a 1M-token context. SWE-benchmark and terminal scores are not broken out on this page (aggregate counts 4 vs 6).

For enterprise. Region, SLA, data-residency and compliance terms are not publicly documented — confirm with the vendor. Deployment and license differ materially: MIT has no revenue thresholds, while the Kimi K3 License imposes obligations for large operators, which may matter in enterprise legal review. DeepSeek-V4-Pro’s deprecated status is the other operational risk to verify.

What is uncertain

• Benchmark scores are vendor-reported and not independently re-measured.
• DeepSeek-V4-Pro is listed deprecated with conflicting official pages on its post-2026-09-14 status.
• No third-party benchmark evidence is recorded for either model.

Sources

• DeepSeek API pricing
• DeepSeek V4 Preview announcement
• Moonshot AI — Kimi K3 (GitHub)
• Kimi platform — Models documentation

Labels used above: Official fact (prices, context windows, capabilities and license terms from primary provider sources), Vendor-reported claim (benchmark scores), and China AI Hub analysis (the “why each difference matters” reasoning, introduced as analysis). No third-party benchmark evidence is currently recorded for these models.