Independent knowledge work
Near frontierGDPval-AA v2 Elo
Artificial Analysis same-harness result. GLM 5.3 Flash sits in the leading cluster behind the displayed Opus variants.
SPIRITT
GLM 5.3 FlashGLM 5.3 Flash makes agent economics hard to ignore
Previewed as Ox Alpha, Z.ai's 320B MoE activates only 18B parameters per token, adds native vision and a 1M context, and scores 57 on Artificial Analysis at roughly $0.09 per evaluated task. SPIRITT serves this Chinese-developed model through U.S.-operated infrastructure, independently of Z.ai's own cloud, with text, images, tools, and $0.15/M input plus $0.50/M output reference pricing.
GLM 5.3 Flash across X
The official Ox Alpha reveal, independent cost and quality data, coding-agent adoption, architecture notes, and one direct frontier-model caveat.
Flash is the price and compute thesis, not a speed trophy
A newly trained multimodal base cuts active compute while keeping serious coding and agent capability.
What Z.ai changed
GLM 5.3 Flash is a 320B-total, 18B-active MoE with a newly trained base, 45 layers, hybrid sparse and linear attention, Manifold-Constrained Hyper-Connections, and a 30T-token multimodal pretraining corpus. It is released under MIT.
SPIRITT serves this Chinese-developed model through U.S.-operated infrastructure, independently of Z.ai's own cloud, with text and image input, function calling, reasoning, vision, a 1,048,576-token context, and reference pricing of $0.15/M input, $0.03/M cached input, and $0.50/M output.
Artificial Analysis scores it at 57 with roughly 49 output tokens per second and $0.09 per Intelligence task. The value is exceptional; the output stream is not unusually fast for its class.
320B total / 18B activeMIT weights1M contextText + imageHybrid attention$0.15 / $0.50 per M
AA Intelligence Index v4.1.1
Artificial Analysis model snapshots, August 29, 2026. GLM 5.3 Flash reaches 57 at a much lower cost than the leading premium models, but it does not lead overall.
The value case, with the frontier gap still visible
Artificial Analysis supplies the independent quality, cost, speed, and knowledge view. Z.ai's table supplies specific coding and tool-use hypotheses.
Vendor terminal agents
Terminal-Bench 2.1
Z.ai launch table. GLM is competitive, but the displayed Terra, Gemini, and Opus runs remain ahead.
Vendor software engineering
DeepSWE v1.1
Z.ai launch table. GLM makes a large within-family jump and beats Opus 4.8 in this setup, while Terra and Gemini lead.
Vendor automation
Near frontierAutomationBench v1.0.6
Z.ai launch table. GLM ranks second in the displayed cohort and improves 22.6 points over GLM-5.2.
Vendor tool use
WinnerToolathlon Verified
Z.ai obtained results through the official evaluation service and reports pass@1 averaged across three runs.
Independent economics
WinnerCost per AA Intelligence task
Artificial Analysis, August 29, 2026. Lower is better; GLM 5.3 Flash is the clear value leader in this displayed quality cohort.
Independent throughput
Output tokens / second
Artificial Analysis rolling model snapshots. The word Flash describes efficiency and cost here, not the fastest output stream.
Independent knowledge limit
AA-Omniscience accuracy
Artificial Analysis. GLM's 28% accuracy is below larger frontier peers even though its hallucination rate is more controlled than some open models.
Reference route price
WinnerOutput price ($ / M tokens)
Reference list prices for the exact hosted routes. Lower is better.
Independent source: Artificial Analysis GLM-5.3-Flash live model page and same-harness evaluation views observed August 29, 2026. Vendor source: Z.ai's August 26 launch report and model-card footnotes. SPIRITT route facts were checked against current hosting documentation and the live model route. Flash offers exceptional price-performance, but its speed, knowledge accuracy, and large self-hosting footprint remain real tradeoffs.
How It Works
From a multimodal brief to GLM 5.3 Flash seeing, coding, using tools, and refining the result inside one workspace
01
Open the full working context
Bring the repository, screenshots, documents, spreadsheets, or research packet into a SPIRITT workspace. The SPIRITT route can use text and images across a 1M context.
02
Pick GLM 5.3 Flash
Choose GLM 5.3 Flash in the model picker. Use its low-cost reasoning and function calling for iterative work, while keeping the expected output and evidence criteria explicit.
03
Let vision close the coding loop
Have it build, render, inspect, and refine. Keep tests and screenshots as the oracle, and route unusually difficult or high-risk decisions to stronger review rather than relying on price alone.
Spend frontier-model money only where the task earns it
GLM 5.3 Flash brings a 1M multimodal route, tools, files, terminal, browser, and memory into one low-cost agent workspace.
Questions
GLM 5.3 Flash FAQ
01What is GLM 5.3 Flash?+
GLM 5.3 Flash is Z.ai's August 2026 open-weight multimodal model with 320B total parameters, 18B active parameters, hybrid sparse and linear attention, a 1M context, and MIT licensing.
02How is GLM 5.3 Flash different from GLM 5.3 and GLM 5.2?+
Flash starts from a new 320B/18B multimodal base built for efficient serving. Full GLM 5.3 is a larger text-focused flagship. Z.ai reports large Flash gains over GLM 5.2 on DeepSWE and AutomationBench at roughly one-tenth the list price.
03Where does SPIRITT host GLM 5.3 Flash?+
SPIRITT serves this Chinese-developed model through U.S.-operated infrastructure, independently of Z.ai's own cloud. This describes the hosting operator and does not promise U.S.-only data residency.
04What is GLM 5.3 Flash best for?+
Cost-sensitive coding agents, multimodal frontend and document work, tool-heavy automation, and long-context workflows where price per accepted task matters more than maximum token streaming speed.
05What are the main limitations of GLM 5.3 Flash?+
Artificial Analysis measures about 49 output tokens per second and 28% knowledge accuracy on AA-Omniscience. The model is somewhat verbose, and the open checkpoint is still hundreds of gigabytes before runtime and context memory.
06What are the reference token economics for GLM 5.3 Flash?+
The hosted route's reference pricing is $0.15 per million input tokens, $0.03 per million cached input tokens, and $0.50 per million output tokens. SPIRITT plan pricing and usage accounting are separate product terms.
07Can I connect a Z.ai subscription from SPIRITT Subscriptions?+
Not currently. Choose GLM 5.3 Flash through SPIRITT Models. The Subscriptions tab currently connects ChatGPT, Claude, and Grok accounts, not Z.ai accounts.
08What is a good example of using GLM 5.3 Flash well?+
Ask it to turn a product brief and reference screenshots into a working responsive page: infer the design system, implement the code, launch it, compare desktop and mobile screenshots, fix visual differences, and return the final build evidence.
09Where is GLM 5.3 Flash best to test for full agentic capabilities?+
SPIRITT Workspaces. Pick GLM 5.3 Flash in the model picker and run real work in a fully equipped cloud environment: tools, browser, files, terminal, memory, and durable sessions. That is where the model can act as an agent, not just chat.
Buy from builders who use what they sellBuilt using