Two 2026 open-weights flagships, both with 1M-token context. Here is the vendor-published benchmark data, the official per-token pricing, the weight availability — and where each one actually wins.
On the two vendors' own benchmark tables, DeepSeek-V4-Pro-0813 and GLM-5.3-Prime are close enough that neither wins outright. The real differences are not on the leaderboard but in three things: price (DeepSeek off-peak is roughly 1/4.3 of Prime for the same job), maximum single output (384K vs about 128K), and deployment shape (both families publish weights, but Prime itself is an API serving tier).
Every row below is taken from an official doc or model card. Anything not confirmed there is marked as unconfirmed rather than guessed.
| Spec | DeepSeek-V4-Pro-0813 | GLM-5.3-Prime |
|---|---|---|
| Vendor / version | DeepSeek (deepseek-v4-pro, weights dated 0813) | |
| Context window | 1M tokens | |
| Max output | 384K tokens | |
| Input modality | ||
| Thinking mode | Switchable non-thinking / thinking (thinking is the default) | |
| Tools / JSON output | ✓ / ✓ (also supports the Anthropic API format) | |
| Open weights | ✓ Hugging Face, MIT licence | |
| Throughput / concurrency | About 1.5–2× GLM-5.3 output throughput, around 77 tok/s in third-party tests | |
| Time-of-day pricing | Yes: off-peak is half of peak |
Both vendors list the other in their own comparison tables, so these numbers come from both model cards and can be cross-checked.
| Benchmark | DeepSeek-V4-Pro-0813 | GLM-5.3 (Prime inherits) | Source |
|---|---|---|---|
| Terminal Bench 2.1 | 87.9 | 88.2 | Both model cards |
| DeepSWE | 62.7 | 66.9 | Both model cards |
| NL2Repo | 61.5 | 58.0 | Both model cards |
| CyberGym | 83.3 | 84.5 | Both model cards |
| Toolathlon Verified | 74.1 | 73.0 | Both model cards |
| Agents' Last Exam | 25.7 | 28.5 | Both model cards |
| HLE (with tools) | 60.0 | 62.5 | Both model cards |
| AutomationBench | 48.2 (v1.0.6) | Different versions — not directly comparable | |
| GDPval-AA v2 | 1590 | 1769 | Z.ai model card (incl. its DeepSeek run) |
These are vendor-run results, not same-condition third-party reproductions: the agents' harnesses, timeouts, context limits and rollout averaging all differ, and AutomationBench is not even the same version. GLM-5.3-Prime's benchmark numbers are inherited from GLM-5.3 because Prime is a serving tier, not a new model. Treat this as directional, not as an absolute ranking.
List prices per million tokens. DeepSeek splits peak and off-peak; GLM-5.3-Prime does not.
| Line item (per 1M tokens) | DeepSeek V4-Pro | DeepSeek V4.1-Flash | GLM-5.3-Prime |
|---|---|---|---|
| Input · cache miss | $0.66 off-peak / $1.32 peak | $0.15 off-peak / $0.30 peak | $2.80 |
| Input · cache hit | $0.022 off-peak / $0.044 peak | $0.003 off-peak / $0.006 peak | $0.56 |
| Output | $1.98 off-peak / $3.96 peak | $0.60 off-peak / $1.20 peak | $8.80 |
| Time-of-day rule | Peak = Mon–Fri 01:00–04:00 and 06:00–10:00 UTC (excl. Chinese public holidays); everything else, including weekends, is half-price off-peak. | Flat rate, no split | |
For the same job, DeepSeek V4-Pro off-peak is about 1/4.3 of Prime and even at peak costs roughly half; V4.1-Flash is another order of magnitude cheaper. Cache-heavy workloads widen the gap further (25× on the cache-hit rate).
GLM-5.3-Prime is not a new model: it is a high-speed serving tier of GLM-5.3. Z.ai and OpenRouter both describe it as inheriting GLM-5.3's full capabilities while inference acceleration lifts output throughput to roughly 1.5–2×, with third-party measurements around 77 tokens/s. Its selling point is streaming feel — realtime coding chat, streamed code generation, and long-horizon agents that already wanted GLM-5.3 quality but found the base ID too slow.
DeepSeek has no equivalent accelerated tier; it buys the same outcome a different way, splitting the day into peak and off-peak windows at half price and publishing concurrency limits (500 for Pro, 2500 for Flash). So these are two different optimisation paths: GLM optimises tokens per second, DeepSeek optimises dollars per million tokens.
DeepSeek-V4-Pro-0813 weights are published on Hugging Face under an MIT licence, alongside DeepSeek-V4.1-Flash (2026-09-10, vision-capable). If you want to self-host on permissive terms, this is the simpler line.
If the cheaper Flash tier is what you want, note it was updated on 2026-10-17 to V4.1-Flash-1017 (MIT weights, 1M context, image input, Terminal-Bench 4.0 at 42.3). Dedicated page:
zai-org/GLM-5.3 (including Flash and BF16 variants) is downloadable on Hugging Face under Z.ai's own glm-5.3 licence. But GLM-5.3-Prime is an API serving tier with no separate weights — if you self-host, you get GLM-5.3 / Flash weights, not Prime's throughput.
Go DeepSeek. Half-price off-peak plus a cache-hit rate of $0.022/MTok against $0.56 compounds into multiples on long-running batch work.
Go GLM-5.3-Prime. Its whole reason to exist is faster token output, which pays off most where a human is watching the stream.
Go DeepSeek V4. Its 384K single-output ceiling is roughly three times Prime's ~128K, so whole-repo generation or book-length rewrites do not hit the wall.
Neither Pro tier takes images. On the DeepSeek side use V4.1-Flash (vision-capable), or move to a natively multimodal model such as Claude Sonnet 5.5.
Prime is a high-speed serving tier of GLM-5.3, not a new model: same capability band, roughly 1.5–2× output throughput from inference acceleration, measured at about 77 tokens/s by third parties. Its benchmark numbers are the GLM-5.3 numbers — Terminal Bench 2.1 at 88.2, for example. The model ID is z-ai/glm-5.3-prime.
On the 2026 tables the two vendors published, they are very close: Terminal Bench 2.1 is 87.9 vs 88.2, CyberGym 83.3 vs 84.5, Toolathlon Verified 74.1 vs 73.0, NL2Repo 61.5 vs 58.0. Which one leads depends on the benchmark, and both are vendor-run under different harnesses, so this is not an absolute ranking. The real differences are price, maximum output length and deployment shape.
GLM-5.3 weights are on Hugging Face (zai-org/GLM-5.3, with Flash and BF16 variants) under Z.ai's own glm-5.3 licence, but GLM-5.3-Prime is an API serving tier with no separate weights. DeepSeek-V4-Pro-0813 weights are also published, under an MIT licence, alongside V4.1-Flash in the same family.
Both are listed at roughly 1M tokens of context in the official docs. Maximum single output differs, though: DeepSeek V4 allows up to 384K output tokens versus about 128K for GLM-5.3-Prime. If your workload is one long generation, that gap matters more than the window itself.
List prices for 1M input (cache miss) plus 250K output: DeepSeek V4-Pro is about $1.16 off-peak and $2.31 at peak, V4.1-Flash about $0.30 off-peak, and GLM-5.3-Prime $5.00. Off-peak, DeepSeek V4-Pro runs at about 1/4.3 of Prime; where caching applies, the cache-hit rate differs by 25×.
Page data was read from the primary sources below on 2026-10-13. Prices and specs change — check the vendor pages before you commit.