GLM-5.3, DeepSeek Flash, and Kimi K3: Only Numbers from the Same Official Table
Rechecked against DeepSeek's pages on 2026-10-18. The old lines — unpublished parameter count, about ¥0.5–2 per million tokens, and no peak-price gap — do not hold. GLM and Kimi appear only as cells on DeepSeek's own card, marked official self-reported and self-selected. Older scores without a shared source are removed.
Checked on 2026-10-18. The pages reopened were DeepSeek’s:
Zhipu’s and Moonshot’s own pricing pages and model cards were not opened. This update does not quote GLM or Kimi monthly fees, parameter counts, or SWE-bench numbers published by those vendors. The old HumanEval / MBPP / LiveCodeBench / SWE-bench table had no shared source on the pages opened this round. The whole table is gone.
The old question, with the premise removed
The earlier draft asked: if you want the cheap option, do you stay on GLM or drop to DeepSeek Flash?
That question assumed Flash was the weaker model and that DeepSeek had no peak surcharge. The pages opened on 2026-10-18 do not support either assumption.
deepseek-flash is now DeepSeek-V4.1-Flash-1017, a minor update on 2026-10-17. Since 16:00 UTC on 2026-08-16, the official API has peak and off-peak prices. Off-peak is half of peak. Peak hours are weekdays 01:00–04:00 and 06:00–10:00 UTC, which is 09:00–12:00 and 14:00–18:00 Beijing time, excluding Chinese public holidays.
DeepSeek price and spec, as published
The official page is USD per 1M tokens. There is no CNY range, and no source for the old “about ¥0.5–2 per million tokens.”
| Off-peak | Peak | |
|---|---|---|
| Flash cache hit / cache miss / output | $0.003 / $0.15 / $0.60 | $0.006 / $0.30 / $1.20 |
| V4-Pro cache hit / cache miss / output | $0.022 / $0.66 / $1.98 | $0.044 / $1.32 / $3.96 |
Context is 1M. Max output is 384K. Flash supports vision. deepseek-v4-pro (V4-Pro-0813) does not. Concurrency limits are 2500 for Flash and 500 for V4-Pro.
Two parameter wordings, and DeepSeek does not explain the gap:
- The 2026-09-10 news post: a 552B MoE backbone, 8B active on input, 16B active on output.
- Both Hugging Face cards show a model size of 763B.
The V4.1-Flash card’s base-model table lists DeepSeek-V4-Pro-Base at 1.6T backbone parameters and 49B activated. The pricing page does not say the served V4-Pro-0813 has that count. The base-model row is not treated here as the API spec.
One table, official self-report
Every cell below is from the DeepSeek-V4.1-Flash-1017 Hugging Face card. DeepSeek ran the evals and chose the rows. Mark them official self-reported. The K3 and GLM-5.3 cells are that card’s numbers. They were not checked against Zhipu or Moonshot pages on 2026-10-18.
The dagger is the text-only subset of HLE. Flash-1017’s headline 42.6 and the 44.0† in parentheses are not the same subset. GLM-5.3 is given only as 42.0†. Do not rank 42.6 against 42.0.
| Benchmark | Flash-1017 | Previous V4.1-Flash | K3 | GLM-5.3 | Opus 5 |
|---|---|---|---|---|---|
| HLE, no tools | 42.6 (44.0†) | 36.8 (39.1†) | 43.5 | 42.0† | 56.3 |
| Terminal-Bench 4.0 | 42.3 | 31.2 | 12.6 | 37.9 | 51.8 |
| DeepSWE v1.1 | 77.0 | 74.2 | 67.5 | 66.9 | 74.0 |
| ProgramBench | 29.1 | 20.3 | 17.5 | 19.0 | 37.0 |
| NL2Repo-Bench | 68.2 | 65.4 | 58.0 | 58.0 | 75.3 |
| Automation-Bench v1.0.6 | 60.5 | 54.8 | 46.7 | 48.8 | 50.3 |
| Agents’ Last Exam | 35.4 | 31.8 | 27.6 | 28.5 | 28.6 |
| Apex Bench v1.1 | 72.1 | 64.6 | 47.8 | 56.6 | 65.8 |
Chartography, BabyVision, and ZeroBench have no GLM-5.3 cell, so they are not compared. K3 does have cells: Chartography 68.1 (Flash-1017 83.0, Opus 5 84.0), BabyVision 85.7 (92.5, Opus 5 94.1), ZeroBench 41.0 (51.0; Opus 5 52.0, GPT-5.6 Sol 53.0).
The footnote for Terminal-Bench 4.0 is the Claude Code harness. The footnote for DeepSWE v1.1 is Mini-SWE-V2.
One mismatch inside DeepSeek’s own pages: the V4.1-Flash instruct table lists NL2Repo-Bench as 64.0. The 10 September changelog and the 1017 comparison table say 65.4. Both are official pages. The reason for the gap is unknown. Neither number is picked as the only one.
On this self-selected table, Opus 5 is still higher on HLE without tools, Terminal-Bench 4.0, ProgramBench, NL2Repo-Bench, Chartography, and BabyVision. Flash-1017 is higher on DeepSWE v1.1, Automation-Bench, Agents’ Last Exam, and Apex Bench. That is not “ahead of closed flagship models across the board.”
Whether to “drop to Flash”
The old answer does not carry over.
On DeepSeek’s table, Flash-1017 is ahead of the GLM-5.3 and K3 cells on most of the agent rows. It is not a slightly worse cheap tier. On HLE without tools, K3’s 43.5 is above Flash-1017’s 42.6. GLM’s 42.0† is a different subset. Opus 5 still leads the rows listed above.
Price is a separate axis. Flash’s official price is a USD peak and off-peak schedule, not pay-as-you-go with no peak gap. Daytime agent runs inside the two Beijing windows pay the peak rate. GLM and Kimi monthly fees were not reopened, so the old ¥49 / ¥149 / ¥599 figures are not compared with this USD table.
Vision is also no longer “Kimi only.” deepseek-flash supports vision. V4-Pro does not. Whether Kimi’s vision is still stronger was not checked on both vendors’ pages, so it is not claimed.
The current answer to the old question: on DeepSeek’s own table, Flash-1017 is not a downgrade from the GLM-5.3 or K3 columns. If the bill is estimated for Beijing weekday daytime, use the peak rate. Subscription prices and each vendor’s own scores wait until those pages are opened again.
Ready to pick your tool?
Check the full overview and find your perfect AI coding setup in 30 seconds.
View Full Overview →📖 Related Articles
5 China-Native AI Coding Subscriptions Compared: August 2026 Guide
An August 2026 comparison of Volcengine Ark Coding Plan, Alibaba Bailian Token Plan, MiniMax Token Plan, GLM Coding Plan, and Kimi Code: price, quota, models, API compatibility, and best-fit scenarios.
MiniMax Coding Plan vs Zhipu GLM Coding Plan: 2026 Deep Comparison
A comprehensive comparison of MiniMax and Zhipu AI's coding subscription plans: pricing, quota mechanisms, model ecosystems, bundled perks, and selection advice.
📚 Related Guides
China Coding API Roundup 2026: Ark vs Bailian vs MiniMax vs Zhipu vs DeepSeek
The most comprehensive 2026 comparison of Chinese AI coding API subscriptions: Volcengine Ark Lite ¥40/mo (¥9.9 first month), Alibaba Bailian ¥200, MiniMax ¥29, Zhipu GLM ¥49+, and DeepSeek pay-as-you-go. One big table, six scenarios, and a decision tree to find your fit in 30 seconds.
Cursor Composer 2 Deep Dive: Kimi K2.5 Base + Reinforcement Learning, 37% Better
In March 2026 Cursor launched Composer 2 — a proprietary coding model trained on top of Moonshot AI's open-source Kimi K2.5. Here's the full story: the technology, the attribution controversy, and what it means for developers.