Is DeepSeek V4-Pro still cheaper than GLM-5.3 after the August 2026 price increase?

Off-peak, yes — V4-Pro costs about $0.0297 per 30K-input/5K-output task versus $0.0640 for GLM-5.3, roughly half. At peak hours (01:00-04:00 and 06:00-10:00 UTC) V4-Pro costs $0.0594 and the gap nearly closes. GLM-5.3 charges a flat $1.40/$4.40 per million tokens with no clock to manage, which is worth something on its own.

Which Chinese AI model is best for coding agents in August 2026?

GLM-5.3, released August 14, 2026, is purpose-built for coding with mandatory thinking mode and a Coding Plan from $18/month, making it the best fit for sustained agent work. DeepSeek V4-Pro wins on raw price off-peak and offers flexible reasoning effort plus native OpenAI Responses API and Anthropic-format endpoints. Kimi K3 costs the most at $3/$15 but ships genuinely open weights.

Do any of these models have open weights?

Kimi K3 released open weights on July 27, 2026 and is the only one of the three you can fully self-host today. GLM-5.3's open-weight release was staged and not available at its August 14, 2026 launch. DeepSeek V4-Pro is API-only. If self-hosting or data residency is a hard requirement, Kimi K3 is the answer despite costing 5x V4-Pro off-peak.

Quick Answer

V4-Pro vs GLM-5.3 vs Kimi K3: Cheap Frontier 2026

Published: August 17, 2026

The Short Answer

DeepSeek V4-Pro still wins on price off-peak. GLM-5.3 wins on coding. Kimi K3 wins if you need the weights.

Model	Input / Output (per MTok)	30K-in / 5K-out	Open weights?
DeepSeek V4-Pro (off-peak)	$0.66 / $1.98	$0.0297	No
DeepSeek V4-Pro (peak)	$1.32 / $3.96	$0.0594	No
GLM-5.3	$1.40 / $4.40	$0.0640	Staged, not at launch
Kimi K3	$3 / $15	$0.1650	Yes (July 27, 2026)

Verified August 17, 2026. DeepSeek’s peak/off-peak schedule took effect 16:00 UTC August 16, 2026.

The August 16 repricing cut DeepSeek’s lead from roughly 4x to 2x off-peak, and to near-parity at peak. For the first time in this category, price alone no longer settles the question.

What Changed

DeepSeek V4-Pro was a flat $0.435 / $0.87. It is now $0.66 / $1.98 off-peak and $1.32 / $3.96 at peak — an output increase of 127% to 355%. Cache-hit input rose from $0.003625 to $0.044 at peak, a 12x jump and the largest single change in the schedule.

Peak hours are 01:00-04:00 and 06:00-10:00 UTC. The remaining 17 hours are half price.

Where Each Wins

DeepSeek V4-Pro (GA August 13, 2026) is still the cheapest frontier-class option if you can schedule around the clock. It carries a 1M context and a 384K max output — the largest output ceiling of the three by a wide margin. Flexible reasoning effort (low / high / max) lets you buy thinking only where it pays, which is a genuine cost lever rather than a marketing bullet. Native OpenAI Responses API support with one-click Codex setup, plus an Anthropic-format endpoint, make it the easiest to drop into existing tooling. Concurrency is capped at 500.

GLM-5.3 (Z.ai, released August 14, 2026) is the coding specialist. It shares the 743B base with GLM-5.2 but is post-trained for agentic coding, and thinking mode is mandatory — you cannot turn it off, which is a deliberate quality-over-cost stance. Cached input is $0.26. The Coding Plan starts at $18/month, and off-peak usage (outside 14:00-18:00 UTC+8 on weekdays) consumes 50% fewer points. If your workload is a coding agent running all day, the subscription math beats per-token billing on all three.

Kimi K3 is the expensive one at $3/$15 flat — 5x V4-Pro off-peak — and it is the only one you can actually run yourself. Open weights landed July 27, 2026. Flat pricing with no peak/off-peak and no thinking-mode surcharge makes it the most predictable of the three. If your constraint is data residency, air-gapped deployment, or long-term independence from a vendor’s pricing decisions, the premium is the point.

The Clock Problem

DeepSeek’s advantage is now conditional on when you run. Peak in local time:

US Eastern: 21:00-00:00, 02:00-06:00 — the workday is entirely off-peak.
Central Europe: 03:00-06:00, 08:00-12:00 — mornings peak, afternoons free.
China (UTC+8): 09:00-12:00, 14:00-18:00 — the whole business day.

GLM-5.3 has its own inverted schedule: its off-peak is outside 14:00-18:00 UTC+8 weekdays, which is a much narrower peak window than DeepSeek’s.

For a North American team, DeepSeek’s price increase is largely theoretical and it remains the clear cost winner. For an Asian team on business hours, V4-Pro peak ($0.0594) versus GLM-5.3 flat ($0.0640) is a 7% difference — at which point you should pick on capability, not price.

The Cache Consideration

If you run a large stable system prompt, cache-hit rates dominate:

Model	Cache-hit input	vs base
DeepSeek V4-Pro (off-peak)	$0.022	~97% off
DeepSeek V4-Pro (peak)	$0.044	~97% off
V4-Pro (before Aug 16)	$0.003625	~99% off
GLM-5.3	$0.26	~81% off

DeepSeek still has the deepest cache discount by a wide margin — but its absolute cache cost rose 6-12x. Teams whose architecture was built on near-free cache hits took the largest increase in this comparison, even though the headline percentages looked worse elsewhere.

The Decision Framework

Cheapest frontier tokens, Western working hours → DeepSeek V4-Pro. Off-peak pricing is still roughly half of GLM-5.3.
Sustained coding agent, want a flat subscription → GLM-5.3 Coding Plan at $18/month. Per-token billing loses to subscriptions on continuous agent work.
Need the weights — residency, air-gap, or independence → Kimi K3. It is the only real option, and the 5x premium buys optionality.
Asia-Pacific business hours → GLM-5.3. You’d pay DeepSeek’s peak rate most of the day for a 7% saving.
Very long outputs → DeepSeek V4-Pro, at 384K max output.
Want one bill you can forecast without a scheduler → Kimi K3. Flat, no clock, no thinking surcharge.

Last verified: August 17, 2026. Prices from official vendor pricing pages.