A question most "switch your model" discussions skip: after a 350% price hike, is DeepSeek actually still cheap?
On August 17, DeepSeek raised V4-Flash peak-hour output pricing from 2 RMB to 9 RMB per million tokens - a 350% increase - and hiked cache-hit input pricing by 400% in the same move. Nine days later, on August 26, Zhipu launched and open-sourced GLM-5.3-Flash: output at 2.8 RMB per million tokens, which against DeepSeek V4-Pro's peak-hour 27 RMB is exactly one-tenth. Lanjing News ran an unsparing headline: "Nine Days After DeepSeek Raised Prices, Zhipu Comes In at One-Tenth the Price."
The product band where these two are fighting has an industry name: the "lite flagship" - MoE models with two to three hundred billion total parameters but only ten-odd billion activated per token. Inference costs are squeezed to the floor while intelligence sits at the flagship line: GLM-5.3-Flash scores 57 on the Artificial Analysis composite intelligence index, level with Claude Opus 4.8. So every piece of selection advice looks obvious: switch to GLM.
Not so fast. Once you break this ledger down by scenario, the conclusion flips once: in coding-agent workloads with cache-hit rates above 90%, DeepSeek V4-Flash's off-peak tier is still about 32% cheaper than GLM-5.3-Flash (inferred basis) - the company that hiked prices 350% may still be the cheaper one for its heaviest users.
Scope first: this site has already covered GLM-5.3-Flash's capability story (The Open-Source Lite-Flagship Backed by 100,000 Domestic GPUs), so this article skips the model review and only does the money: unit-price provenance, peak/off-peak spreads, cache-hit pricing, the real cost across three scenario types, and the industry ledgers behind this price war. The five contenders are GLM-5.3-Flash, DeepSeek V4-Flash, DeepSeek V4-Pro, and Kimi K3, with Claude Opus 4.8 as the international reference. Every number is tagged with its basis (official pricing / media report / inferred / converted); overall this is a representative comparison - not our own stress test, not investment advice - verify against official pages before purchasing.
1. The Pricing Baseline: How Five Models Bill
The master table first. Note that provenance varies wildly: first-party official pricing, media-reported figures, inferred prices, and converted prices - each tagged line by line (prices as of the 2026-08-27 snapshot):
| Model | Total/active params | Input (RMB/M tokens) | Output (RMB/M tokens) | Cache-hit input | AA intelligence index | Basis |
|---|---|---|---|---|---|---|
| GLM-5.3-Flash (Zhipu) | 320B / 18B | 0.8 | 2.8 | 0.23 | 57 (level with Claude Opus 4.8) | Official pricing, live 2026-08-26 |
| DeepSeek V4-Flash | 284B / 13B | Off-peak ~1.5 (inferred) | Peak 9 / off-peak 4.5 | Off-peak 0.05 (after the 400% hike) | Not collected | Lanjing News 2026-08-27; off-peak input reverse-derived from "GLM input is 53% of it" |
| DeepSeek V4-Pro | Not collected | Not collected | Peak 27 (~$3.96, vs $0.87 pre-hike) | Not collected | 53 | Lanjing News 2026-08-27 |
| Kimi K3 (Moonshot) | 2.8T | 2-20 | 100 (more than 3x the previous gen) | Not collected | Not collected | Tencent Cloud developer community, 2026-07-17 |
| Claude Opus 4.8 (intl. reference) | Not collected | ~32 (converted) | ~112 (converted) | Not collected | 57 | Reverse-derived from Zhipu's official "GLM is ~1/40 of it"; not the official list price |
Three first-glance conclusions:
- Peak/off-peak pricing is a new variable. DeepSeek's hike introduced time-of-day pricing: 9am-12pm and 2pm-6pm are peak hours, with V4-Flash output at 4.5 RMB off-peak and 9 RMB peak - a 2x spread within a single day; V4-Pro peaks at 27 RMB. For the first time, a domestic API price table has "time of day" written into it, so buyers now have one more dimension to model: can your workload run off-peak?
- The media's "one-tenth" holds - depending on what you compare against. GLM output at 2.8 RMB is exactly 1/10 of V4-Pro's peak 27 RMB, about 1/3 of V4-Flash's peak 9 RMB, and about 62% of its off-peak 4.5 RMB. "How much cheaper is Zhipu" has three different answers - which is precisely why this ledger needs to be broken apart.
- The AA index and price have clearly decoupled. GLM-5.3-Flash's 57 is level with Opus 4.8 at roughly 1/40 the price (Zhipu's official framing); V4-Pro scores 53 with peak output nearly 10x GLM's. In this cohort, "capability score" and "value for money" are two independent axes.
One more detail worth recording: GLM-5.3-Flash is the only model in the table with an ultra-low cache-hit price (0.23 RMB/M, about 29% of its normal input price), paired with a 1M-token context, 128K max output, MIT open weights, and inference running on a domestic 100,000-GPU cluster. DeepSeek's hike, meanwhile, has already shown its cost: its call volume on the coding tool OpenCode fell by half (LatePost) - one end of the pricing lever is revenue, the other is churn, and both companies are betting their own ledgers.
2. The Core Ledger: Three Scenarios, Three Answers
A single price table cannot answer "which one should I use." Break the ledger into three typical scenarios (prices as of the 2026-08-27 snapshot; DeepSeek's off-peak input price is on an inferred basis):
| Scenario | GLM-5.3-Flash | DeepSeek V4-Flash | Notes |
|---|---|---|---|
| Coding agent (900K of 1M input tokens served from cache) | 0.8×0.1+0.23×0.9=0.287 RMB/M input | Off-peak 1.5×0.1+0.05×0.9=0.195 RMB/M input | V4-Flash ~32% cheaper; inferred basis |
| Chat / RAG / one-shot long documents (cache almost never hits) | Input at 53% of V4-Flash off-peak, output at 62%; 27%/31% during the half-price launch window | Off-peak tier as the baseline | GLM clearly cheaper |
| Peak hours (9-12, 14-18) | No peak/off-peak spread | Output doubles to 9 RMB/M (V4-Pro at 27 RMB/M) | GLM's biggest paper advantage |
Scenario one: coding agents - run the ledger in reverse. Coding-agent inputs are highly repetitive - system prompts, codebase context, tool definitions get re-sent every round, and cache-hit rates routinely exceed 90%. Counting 900K of every 1M input tokens as cache hits: GLM-5.3-Flash costs 0.8×0.1 + 0.23×0.9 = 0.287 RMB per M input; DeepSeek V4-Flash off-peak costs 1.5×0.1 + 0.05×0.9 = 0.195 RMB, about 32% cheaper (off-peak input price on an inferred basis). In other words, DeepSeek's post-hike cache-hit price of 0.05 RMB/M is still the lowest input unit price on the board. Hence the reversal in the title: coders shouldn't switch just yet - check your own cache-hit rate first, then run the numbers. Two caveats: your jobs must be schedulable into off-peak windows, and that 1.5 RMB off-peak input price is inferred - verify against the official page before committing.
Scenario two: zero cache is GLM's sweet spot. Ordinary chat, RAG retrieval, one-shot long-document summarization - the cache almost never hits, so whoever's sticker price is lower wins. GLM's input of 0.8 and output of 2.8 are 53% and 62% of V4-Flash's off-peak tiers (~1.5 / 4.5) respectively; stacked with GLM's current half-price launch window, that drops further to 27% and 31%. In this band, "cutting in at half price" is literal.
Scenario three: peak hours, where the spread stretches past 3x. During the daytime 9-12 and 14-18 windows, the V4 line doubles outright (V4-Flash output 9 RMB, V4-Pro 27 RMB), while GLM has no peak/off-peak spread. For workloads that can't be shifted - live customer service, daytime batch jobs - GLM's peak output of 2.8 RMB is about 31% of V4-Flash's 9 RMB, and about 1/10 of V4-Pro's 27 RMB: the media's "one-tenth" is most literal in this scenario. Conversely, if all your jobs can be scheduled into nighttime off-peak hours, GLM's advantage shrinks to its minimum.
3. Capability Check: Beyond Cheap, Is It Smart Enough?
Three coordinates beyond price (all third-party or media bases, 2026-08-27 snapshot):
- AA composite intelligence index: GLM-5.3-Flash at 57, level with Claude Opus 4.8; DeepSeek V4-Pro at 53. GLM-5.3-Flash has the most extreme "score per price" in the table - but AA is a composite index; the ranking on your own tasks (a specific class of code, a specific vertical) may differ, so run your own eval set before committing.
- Kimi K3 is a capability premium, not value for money. 2.8 trillion total parameters (beating DeepSeek V4's 1.6T and ERNIE 5.0's 2.4T - the world's first open 3T-class model), KDA hybrid linear attention plus attention residual, native vision understanding, and a 1M context; output at 100 RMB/M is more than triple the previous generation. What you're buying is the capability ceiling - a 1679 on Frontend Code Arena to top Claude Fable 5, independently designing and validating a chip in 48 hours - not token unit price.
- Context and open weights: GLM-5.3-Flash offers a 1M context, 128K max output, and MIT open weights; Kimi K3 also has 1M context and opened its weights on 2026-07-27. For teams planning private deployment or a cost backstop (self-hosted migration if API prices climb), these two are the only open-weight options in the table.
4. Behind the Ledger: Why Some Raise Prices While Others Cut
This price war isn't sentiment - it's financial statements (all media-relayed bases, 2026-08-27 snapshot):
| Company | Key figures |
|---|---|
| DeepSeek | ~475M RMB revenue in the first 7 months (10x all of 2025); net loss of 715M RMB; API gross margin 82.9%, overall 44.6%; ~11B RMB in AI infrastructure investment (23x revenue) (The Information, citing informed sources) |
| MiniMax | H1 revenue of $117M (+283%); open-platform revenue of $73.929M (+703%); overall gross margin 17.9% (12.1% a year earlier); July token consumption 20x January's (Yan Junjie, earnings call) |
| Zhipu | API call pricing cumulatively up 83% through Q1, while call volume grew 400% |
| Intl. reference | OpenAI Q1 gross margin 39%; Anthropic expects to lift gross margin from 40% to 63% this year |
This table explains two things. First, the hikes are an industry move, not an outlier: seven cloud and model vendors have raised prices since the start of the year (per BlockBeats/Lvdong tally) - with inference costs high and infrastructure investments routinely in the billions, gross-margin repair is a collective issue, and OpenAI and Anthropic are working on the same problem. Second, the reason Zhipu dares to cut prices while everyone else raises them is that it has already validated "reprice without losing volume": pricing up a cumulative 83% while call volume grew 400% - whereas DeepSeek's 350% hike halved its OpenCode call volume is the counter-example. Who's right is undecided, but one thing is certain: a price table in a pricing-power contest can change at any time.
One more demand-side coordinate (Kun Xiaoqi Human, Weibo, 2026-07-27 - a community tally, for reference only): weekly token consumption has DeepSeek V4 Flash at 5.44T (+14% week over week), Xiaomi MiMo-v2.5 at 4.46T (+5%), MiniMax M3 at 4.24T (+15%), and Zhipu GLM-5.2 at 2.9T (+25%). DeepSeek is still number one by consumption, but Zhipu is growing fastest - GLM-5.3-Flash's pricing is clearly aimed at share.
5. Choose by Scenario: Five Conclusions
- No-cache, input/output-heavy chat / RAG / one-shot long documents: pick GLM-5.3-Flash. Even at the off-peak comparison it's nearly half of V4-Flash, lower still during the half-price window, and the 1M context plus 128K max output hold up. The full integration walkthrough is in our GLM-5.3-Flash Integration SOP.
- Coding agents, batch jobs, and other high cache-hit workloads: run the cache math before switching. Above 90% hit rate and schedulable off-peak, V4-Flash is still ~32% cheaper (inferred basis); low hit rate or daytime-only jobs favor GLM's 0.23 RMB cache price with no peak/off-peak spread. For general money-saving plays, revisit our API Cost Optimization SOP.
- Need the capability ceiling (frontend code, complex reasoning, vision): Kimi K3 is the capability-premium tier - 100 RMB/M output is not a value pick; GLM-5.3-Flash's AA 57 is the best score-per-price on the board. If you're choosing a coding subscription instead of raw APIs, see our AI Coding Plan Comparison.
- Going global or benchmarking against international vendors: Claude Opus 4.8 remains the capability anchor (AA 57), but its converted price is roughly 40x GLM's. In most Chinese-market scenarios, that international budget can be saved for something else.
- The one conclusion that applies to every scenario: don't hard-code today's price table into an annual contract. The industry has entered a pricing-power contest - seven vendors raising prices and Zhipu cutting them are happening simultaneously, and a price snapshot's shelf life may be a few weeks. Keeping a switchable path in your architecture is worth more than betting on any one vendor's sticker price.
FAQ
Q1: GLM-5.3-Flash output at 2.8 RMB versus DeepSeek V4-Flash off-peak at 4.5 RMB - should I just switch to GLM? A1: No - it depends on the scenario. For zero-cache chat / RAG / long documents, GLM's input and output are 53% / 62% of V4-Flash's off-peak prices (27% / 31% during the half-price window), genuinely cheaper. But for coding-agent workloads with cache-hit rates above 90%, counting 900K of every 1M input tokens as cache hits puts GLM at 0.287 RMB/M input versus V4-Flash off-peak at 0.195 RMB - V4-Flash is about 32% cheaper (inferred basis). Check your own cache-hit rate first, then run the ledger.
Q2: How are peak hours defined? Can my workload shift off-peak? A2: Per the Lanjing News basis, 9am-12pm and 2pm-6pm daily are peak windows, and calls within them bill at peak rates. Batch processing, offline evals, and nighttime agent runs naturally fit the off-peak tier; live customer service and daytime real-time interaction can't shift and must be compared at peak rates - where GLM, with no peak/off-peak spread, holds its biggest advantage.
Q3: Kimi K3 at 100 RMB/M output - what's the point? A3: The capability ceiling, not value for money. 2.8 trillion total parameters (the world's first open 3T-class model), KDA hybrid linear attention plus attention residual, native vision, 1M context, a 1679 on Frontend Code Arena to top Claude Fable 5, and a chip independently designed and validated in 48 hours. It plays in the capability-premium tier - not at the same table as cost-squeezed lite flagships like GLM-5.3-Flash.
Q4: Why is Claude Opus 4.8's price "converted" rather than official? A4: Because this article didn't collect its official list price; it uses Zhipu's official framing that GLM-5.3-Flash is priced at roughly 1/40 of it, reverse-derived to about 32 RMB/M input and 112 RMB/M output. That's a converted basis, not an official sticker price - useful for order of magnitude only; when pricing in USD, verify against Anthropic's official page.
Q5: Will this price table change? Should I lock in an annual contract now? A5: It will, and possibly quickly. Seven cloud and model vendors have raised prices since the start of the year (BlockBeats tally) while Zhipu is cutting for share; Zhipu itself saw call volume grow 400% while raising prices a cumulative 83% through Q1, and DeepSeek's 350% hike halved its OpenCode call volume. The industry is in a pricing-power contest: keep your architecture switchable (multiple API channels plus an open-weights backstop) and re-quote quarterly rather than writing today's snapshot into an annual contract.
References
- Lanjing News (2026-08-27 15:17): "Nine Days After DeepSeek Raised Prices, Zhipu Comes In at One-Tenth the Price" - peak/off-peak pricing, V4-Flash/V4-Pro prices, the seven-vendor hike tally, and the DeepSeek/MiniMax/Zhipu financial bases: https://baijiahao.baidu.com/s?id=1874660001676882489
- Cailian Press / Kechuang Board Daily (2026-08-27): "100,000 Domestic AI Cards Carry the Bull Market: Zhipu Open-Sources GLM-5.3-Flash" - total/active params, AA 57, the 100K-GPU cluster: https://baijiahao.baidu.com/s?id=1874642322871069438
- Tencent Cloud developer community (2026-07-17): "2.8 Trillion Parameters: Kimi K3 Officially Released" - K3 params, KDA, pricing and capability data: https://cloud.tencent.com/developer/news/4268105
- Zhipu AI open documentation, GLM-5.3-Flash model card - official pricing, context and output limits, MIT weights: https://docs.bigmodel.cn/cn/guide/models/vlm/glm-5.3-flash
- Artificial Analysis composite intelligence index (as relayed by Cailian Press / Lanjing) - GLM-5.3-Flash at 57, V4-Pro at 53, Opus 4.8 at 57
- LatePost, The Information, BlockBeats/Lvdong, and Kun Xiaoqi Human on Weibo (2026-07-27) - OpenCode call volume, DeepSeek's financials, the hike tally, and the weekly-tokens community stat, all as relayed in Lanjing News' 2026-08-27 report
- Related reading: GLM-5.3-Flash Open-Source Launch Hotspot (same batch), GLM-5.3-Flash Integration SOP (same batch), AI Coding Plan Comparison, LLM Cost-Performance Comparison, LLM Price War Hotspot, DeepSeek V4 Flash Hands-On Benchmark, Kimi vs Qwen vs GLM Long-Context Review, API Cost Optimization SOP
This is a representative comparison assembled from public reporting and official pricing; prices are a 2026-08-27 snapshot (DeepSeek's off-peak input price is inferred and Claude Opus 4.8's is a converted basis, both tagged), not our own stress test and not investment advice; prices and peak/off-peak rules may change at any time - verify against official pages before purchasing.