Open Models
New Open Model Prices, August 2026 Snapshot: DeepSeek V4, Kimi K3, GLM-5.2, MiniMax M3, Qwen 3.6
At the 22 August 2026 recorded snapshot, DeepSeek V4 Flash costs $0.44 input and $1.32 output per million tokens, Kimi K2.5 costs $0.60 and $3.00, GLM-5.2 costs $1.40 and $4.40, MiniMax M3 costs $0.30 and $1.20, and Qwen 3.6-Flash lists $0.1875 and $1.125. Kimi K3 is the outlier at $3.00 input and $15.00 output. Provider pages remain the final word, since this snapshot is not a fresh verification.
Snapshot scope
Every rate in this post comes from one committed record, data/ai-provider-pricing.json, generated at 2026-08-22T10:54:07.443Z. GPU rental figures come from data/gpu/gpu-pricing.json, generated at 2026-08-22T14:22:27.451Z. Measured throughput comes from data/gpu/gpu-throughput.json, generated at 2026-07-09T12:35:09.683Z. Architecture facts come from data/gpu/open-model-specs.json, generated at 2026-08-04T06:48:51.735Z. Nothing here was rechecked against a live provider page at writing time, so confirm any rate on the provider page before signing a contract.
New open model prices
All rates are USD per 1M tokens. The provider and model ids name the exact committed record each row reproduces.
| Model | Provider record | Input | Output | Cache read | Context |
|---|---|---|---|---|---|
deepseek-v4-flash | deepseek | 0.44 | 1.32 | 0.014 | 1,000,000 |
deepseek-v4-pro | deepseek | 1.32 | 3.96 | 0.044 | 1,000,000 |
kimi-k2.5 | moonshotai | 0.60 | 3.00 | 0.10 | 262,144 |
kimi-k2.6 | moonshotai | 0.95 | 4.00 | 0.16 | 262,144 |
kimi-k2.7-code | moonshotai | 0.95 | 4.00 | 0.19 | 262,144 |
kimi-k3 | moonshotai | 3.00 | 15.00 | 0.30 | 1,048,576 |
glm-5 | zhipuai | 1.00 | 3.20 | 0.20 | 204,800 |
glm-5.2 | zhipuai | 1.40 | 4.40 | 0.26 | 1,000,000 |
MiniMax-M2.5 | minimax | 0.30 | 1.20 | 0.03 | 204,800 |
MiniMax-M3 | minimax | 0.30 | 1.20 | 0.06 | 1,000,000 |
qwen3.5-397b-a17b | alibaba | 0.60 | 3.60 | unknown | 262,144 |
qwen3.6-flash | alibaba | 0.1875 | 1.125 | unknown | 1,000,000 |
unknown means the committed record carries no cache-read rate for that model, not that caching is free. The moonshotai, zhipuai, and alibaba provider records carry docs URLs but no dedicated official pricing URL in the snapshot, while deepseek and minimax records link official pricing pages directly.
Equal-workload math
Use one shared unit, 1M input tokens plus 1M output tokens, so versions compare on equal volumes:
deepseek-v4-flash:0.44 + 1.32 = $1.76kimi-k2.5:0.60 + 3.00 = $3.60MiniMax-M3:0.30 + 1.20 = $1.50glm-5.2:1.40 + 4.40 = $5.80qwen3.6-flash:0.1875 + 1.125 = $1.3125kimi-k3:3.00 + 15.00 = $18.00
A monthly coding workload of 10M input and 5M output tokens shows the same ordering at scale:
deepseek-v4-flash:10 × 0.44 + 5 × 1.32 = 4.40 + 6.60 = $11.00kimi-k2.7-code:10 × 0.95 + 5 × 4.00 = 9.50 + 20.00 = $29.50MiniMax-M3:10 × 0.30 + 5 × 1.20 = 3.00 + 6.00 = $9.00
Output tokens drive each total because every output rate here exceeds its input rate. A lower price never implies higher quality, so route by evaluation results and keep price as a separate column.
What is still unknown
Three gaps stay visible. First, the committed throughput file holds 26 measured rows and none for kimi-k3, kimi-k2.6, kimi-k2.7-code, glm-5.2, or qwen3.6-flash, so self-host unit cost for those versions cannot be derived from committed measurements yet. Second, the Moonshot, Zhipu, and Alibaba records have no official pricing URL in the snapshot, which weakens provenance for the Kimi, GLM, and Qwen rows relative to the DeepSeek and MiniMax rows. Third, open-weights flags differ by record: Kimi K3 and the K2 series are flagged open weights in the pricing record, while glm-5.2 and deepseek-v4-pro are not, even though the architecture file lists a Hugging Face entry for DeepSeek V4 Pro under MIT. Confirm license terms in the model repository before planning a self-host.
One more narrow note: glm-4.5-flash lists 0 input and 0 output in the committed record. Treat that as a record value to confirm on the provider page, not as a promise of free inference.
Self-host sketch, modeled
Measured throughput and modeled cost stay separate here. The committed throughput file records MiniMax-M3 at 2,814.84 decode tokens per second per GPU on b200 with vLLM, fp4, 1k input and 1k output tokens, and concurrency 512. The GPU catalog records a $6.79 median hourly rate per B200 GPU across 201 records. Modeled cost follows:
2,814.84 × 3,600 = 10,133,424 output tokens per GPU hour, about 10.13M.
6.79 × 730 = $4,956.70 per GPU for a 730-hour month.
10,133,424 × 730 = 7,397,399,520 tokens per month at full use, about 7,397M.
4,956.70 ÷ 7,397.40 = $0.67 per 1M output tokens at full use.
At 30 percent sustained use the modeled figure rises to 0.67 ÷ 0.30 = $2.23 per 1M, above the $1.20 API output rate for MiniMax-M3. The break-even sits near 0.67 ÷ 1.20 = 56 percent sustained use. This sketch excludes staff time, replicas, power, and long-context slowdown, and the lab measurement used short 1k contexts, so production cost per token may be higher.
These throughput figures are normalized per GPU across a serving cluster. They do not imply the model fits on one GPU; total rental spend scales with the number of GPUs required. These output-only sketches exclude API input charges and are not complete service-cost comparisons.
Frequently asked questions
Which newly tracked model is cheapest per equal workload?
Qwen 3.6-Flash at $1.3125 per 1M input plus 1M output tokens, followed by MiniMax-M3 at $1.50 and DeepSeek V4 Flash at $1.76. Kimi K3 at $18.00 is the most expensive row in this snapshot.
Why does Kimi K3 cost so much more than Kimi K2.5?
The committed record prices kimi-k3 at $3.00 input and $15.00 output with a 1,048,576-token context, against $0.60 and $3.00 with 262,144 context for kimi-k2.5. The record gives no quality comparison, so price the context and generation mix separately.
Should a team self-host one of these models now?
Only with a measured throughput number for the exact model, GPU, precision, and context length, plus sustained use above the modeled break-even. The open model token cost calculator and the self-host versus API calculator turn those inputs into a per-1M comparison.
Sources
New Open Model Prices, August 2026 Snapshot: DeepSeek V4, Kimi K3, GLM-5.2, MiniMax M3, Qwen 3.6. ByteCosts. Updated 2026-09-05. https://bytecosts.com/blog/new-open-model-pricing-watchlist/