Did DeepSeek V4 Flash just dethrone GLM 5.2?

For most of 2026, GLM-5.2 has been the undisputed king of open-weight coding models. Zhipu AI's 753-billion-parameter MoE flagship topped the open-source leaderboard on SWE-bench Pro, Terminal-Bench 2.1, and the Artificial Analysis Intelligence Index, and it landed within striking distance of Claude Opus 4.8 on agentic benchmarks. Then, on July 31, DeepSeek dropped the 0731 update to V4 Flash — and the gap suddenly looked a lot narrower.

DeepSeek V4 Flash is the smaller, cheaper sibling in the V4 family: 284B total parameters (13B active) against GLM-5.2's 753B (40B active). It costs one-tenth the price of GLM-5.2 on the API ($0.14/$0.28 per million input/output tokens vs. $1.40/$4.40). And on several key coding benchmarks, the 0731 update doesn't just close the gap — it pulls ahead.

So did DeepSeek V4 Flash just dethrone GLM-5.2? The short answer is: it depends on the benchmark. This article walks through every major score, the pricing math, and where each model still holds a clear edge.

The 0731 update changed everything

The original DeepSeek V4 Flash preview (April 24, 2026) was a capable but clearly second-tier model: 61.8 on Terminal-Bench 2.1, 7.3 on DeepSWE, 37.0 on DSBench-FullStack. It was fast and cheap, but nobody confused it with a frontier model. The July 31 update (DeepSeek-V4-Flash-0731) is a different beast entirely. DeepSeek claims it "outperforms DeepSeek-V4-Pro (Preview) despite its far smaller activated parameter count, and is broadly competitive with the strongest proprietary models available."

The numbers back that up. On Terminal-Bench 2.1, Flash jumped from 61.8 to 82.7 — a 34% improvement that puts it ahead of GLM-5.2's 81.0. On DeepSWE, it went from 7.3 to 54.4, nearly matching GLM-5.2's 46.2. On DSBench-FullStack, it climbed from 37.0 to 68.7, beating GLM-5.2's 61.8. The 0731 update effectively turned a budget model into a contender.

Where Flash wins: coding benchmarks

The chart above tells the story. DeepSeek V4 Flash leads on SWE-bench Verified (79.0% vs. 62.1%), Terminal-Bench 2.1 (82.7 vs. 81.0), and DSBench-FullStack (68.7 vs. 61.8). These are real-world coding benchmarks: SWE-bench Verified tests the model's ability to resolve GitHub issues by editing code, Terminal-Bench evaluates autonomous terminal-based development, and DSBench measures full-stack application building. On all three, Flash's 0731 update either matches or exceeds GLM-5.2.

The margin on SWE-bench Verified is particularly striking: 79.0% vs. 62.1% is a 17-point gap. GLM-5.2's SWE-bench Pro score (62.1) is actually higher than its Verified score, which is unusual — most models score lower on Pro because the tasks are harder. This suggests GLM-5.2 may have been optimized for the Pro split, while Flash generalizes better across the full Verified set.

Where GLM-5.2 still leads: reasoning and agentic work

GLM-5.2 isn't about to disappear. On the Artificial Analysis Intelligence Index, GLM-5.2 scores 51 vs. Flash's 50 (max effort) — a narrow but consistent lead on broad intelligence. On GPQA Diamond (graduate-level science reasoning), GLM-5.2 scores 91.2% vs. Flash's 88.1%. On AIME 2026 (math competition problems), GLM-5.2 hits 99.2% vs. Flash's 95.8%. And on MCP Atlas (tool-use and agentic capability), GLM-5.2 leads 76.8% to 69.0%.

The pattern is clear: when the task requires deep reasoning, multi-step tool use, or sustained agentic trajectories, GLM-5.2's larger active parameter count (40B vs. 13B) gives it an edge. Flash is remarkably efficient for its size, but it can't fully compensate for having 3× fewer active parameters on the hardest reasoning problems.

The price gap is enormous

This is where the comparison gets interesting. Here are the official API prices side by side:

* Estimated per-task cost for a typical coding agent interaction (~8,000 input + ~1,000 output tokens, 50% cache hit rate).

Flash is 10× cheaper on input tokens and 15.7× cheaper on output tokens. The cached input gap is even wider: Flash's cache-hit rate of $0.0028/M is nearly 93× cheaper than GLM-5.2's $0.26/M. For a team running high-volume agentic workloads, the cost difference is transformative — Flash can handle roughly 10,000 coding agent tasks for the same price as 1,000 on GLM-5.2.

Full comparison table

The verdict: dethroned or not?

If "dethroned" means "GLM-5.2 is no longer the best open-weight coding model," the answer is a qualified yes — on specific coding benchmarks. DeepSeek V4 Flash (0731) matches or beats GLM-5.2 on Terminal-Bench 2.1, SWE-bench Verified, DSBench-FullStack, and DeepSWE, all at one-tenth the price. For a team that primarily needs autonomous coding and terminal-based development, Flash is the better choice today.

If "dethroned" means "GLM-5.2 is obsolete," the answer is no. GLM-5.2 still leads on broad intelligence (Intelligence Index 51 vs. 50), graduate-level reasoning (GPQA Diamond 91.2% vs. 88.1%), math (AIME 99.2% vs. 95.8%), and agentic tool use (MCP Atlas 76.8% vs. 69.0%). For workloads that demand deep reasoning or long-horizon agentic trajectories, GLM-5.2 remains the stronger model.

The real story is that the open-weight landscape now has two genuinely competitive coding models at very different price points. GLM-5.2 is the premium option for the hardest reasoning tasks. DeepSeek V4 Flash is the high-value option for high-volume coding work. Both are MIT-licensed, both support 1M-token contexts, and both are within striking distance of closed frontier models. The "dethroning" narrative oversimplifies a market that's getting more interesting, not less.

The practical takeaway: if you're building a coding agent today, start with DeepSeek V4 Flash for the cost savings, and escalate to GLM-5.2 (or a closed frontier model) when the task requires deeper reasoning. That two-tier strategy is exactly the kind of workflow that's making open-weight models irresistible to the Fortune 500.