GLM 5.3 Flash vs Claude Opus vs GPT 5.6: What can this open source model do

Zhipu's GLM-5.3-Flash arrived on August 26, 2026 as the model the community had been tracking under the codename Ox Alpha. It topped OpenRouter free-model usage charts, ended a long run by DeepSeek on the OpenCode leaderboard, and then shipped as open weights under the MIT license. For teams deciding whether an open model can replace a closed flagship, the question is concrete: how does this 320-billion-parameter mixture-of-experts compare with Anthropic's Claude Opus and OpenAI's GPT-5.6 on the workloads that actually drive cost?

The short answer is that GLM-5.3-Flash closes the gap on coding and agentic work at a fraction of the API price and with self-hosting as an option, while the closed flagships retain an edge on frontier reasoning depth, long-horizon reliability, and ecosystem polish.

What GLM-5.3-Flash is

GLM-5.3-Flash is the first natively multimodal model in the GLM-5 series from Z.ai (formerly Zhipu AI). It is a mixture-of-experts with 320 billion total parameters and 18 billion active per token, trained on a 30 trillion-token multimodal corpus. Two architecture changes define it compared with GLM-5.2: a hybrid sparse and linear attention design that cuts long-context serving cost while preserving retrieval accuracy, and Manifold-Constrained Hyper-Connections that improve scaling efficiency. The model serves a 1-million-token context window and up to 128K output tokens.

The release is open under the MIT license. Weights are available on Hugging Face and ModelScope in FP8 (default) and BF16 formats. That license permits commercial use, modification, and redistribution without a separate agreement. The model also runs through Z.ai's API and Z.ai Coding Plan, where Flash consumes roughly one third of the quota points of the full GLM-5.3.

GLM-5.3 itself is a different story: it reuses the same 744B MoE base as GLM-5.2 and gains entirely through post-training reinforcement. GLM-5.3-Flash starts from a newly trained base and is designed for efficiency at lower cost.

Benchmarks: where Flash lands

No single leaderboard settles a model choice. The useful pattern is to look at coding, agentic tool use, and reasoning together, and to keep the harness and effort level in mind. The table below uses numbers published by Z.ai for GLM-5.3-Flash and GLM-5.2, with competitor figures from vendor system cards and from Artificial Analysis independent runs. All figures are pass@1 unless noted, at high or max effort where the source specifies it. Claude Opus 5 is the current flagship (July 24, 2026); Opus 4.8 is kept as a reference for continuity.

Benchmark	GLM-5.3-Flash	GLM-5.2 (prior open baseline)	Claude Opus 5 (latest)	Claude Opus 4.8 (reference)	GPT-5.6 Sol (max)	GPT-5.6 Terra (max)
SWE-bench Verified (curated bug fixes, 5-trial mean)	-	-	96.0 (97.0 on vals.ai)	88.6	-	-
SWE-bench Pro (real GitHub issues, harder split)	-	-	79.2	69.2	64.6	-
Terminal-Bench 2.1 (CLI agent)	84.3	81.0	89.1 (Artificial Analysis)	85.0*	88.8 (91.9 Ultra)	87.4
DeepSWE v1.1 (multi-file engineering)	63.4	46.2	68.8	58.0	72.7	69.6
Frontier-Bench v0.1 (agentic coding, 74 tasks)	-	-	43.3 (over double Opus 4.8 at 21.1)	21.1	34.4	-
ARC-AGI-3 (novel reasoning)	-	-	30.2	1.5	7.78	-
Toolathlon Verified (tool use, 3-run avg)	78.4	59.9	-	76.2	74.9	-
AutomationBench (workflow execution)	48.8	26.2	26.0	41.0*	18.1	37.2
Agents' Last Exam (ALE-CLI slice)	26.3	20.4	-	27.0	28.6	-
HLE with tools (full set, GPT-5.6-luna as judge)	55.3	54.7	-	57.9	64.5	-
GDPval-AA v2 (knowledge work Elo, Artificial Analysis)	1773	1504	1861	1588	1736	1571
AA Intelligence Index (composite, 9 evals)	-	-	61 (leader, ahead of Fable 5 at 60)	56	59	55
CyberGym (vuln discovery)	84.5	77.2	-	78.1	83.6	-
OSWorld 2.0 (computer use)	-	-	70.6	-	62.6	50.2
Z.ai Code Bench (internal, Max effort)**	outperform GLM-5.2 at every effort; near Opus 4.8 at Max (29.0 vs 29.5)	23.4 @ 96K tokens	-	29.5 @ 120K tokens (Opus 4.8)	-	-

* Z.ai table value; Artificial Analysis lists Opus 5 AutomationBench at 26.0 on Zapier's harness, which differs from Z.ai's 41.0 for Opus 4.8. Keep harness in mind. ** Internal benchmark, run on Claude Code 2.1.207; useful for within-family trend, not cross-lab ranking.

What the row pattern shows:

• Coding. GLM-5.3-Flash moves from 46.2 to 63.4 on DeepSWE v1.1, a 17.2-point gain over its predecessor, and reaches 84.3 on Terminal-Bench 2.1. That sits below Opus 5 at 89.1 on the same Terminal-Bench harness tracked by Artificial Analysis and below GPT-5.6 Sol at 88.8 (91.9 in Ultra mode). On SWE-bench, the current Opus 5 leads the field at 96.0 Verified and 79.2 Pro, while GPT-5.6 Sol posts 64.6 on Pro. Flash has no published SWE-bench Verified or Pro number, so direct comparison on that harness remains open.
• Agentic tool use. Toolathlon Verified at 78.4 and AutomationBench at 48.8 are the clearest wins for Flash on Z.ai's harness. Both numbers top the Opus 4.8 and GPT-5.6 Terra figures Z.ai listed (76.2 and 74.9 on Toolathlon). Flash went from 26.2 to 48.8 on AutomationBench on that harness, an 86 percent relative gain. On the independent Zapier AutomationBench, Opus 5 records 26.0 and GPT-5.6 Sol records 18.1, which shows how harness choice shifts absolute values.
• Reasoning. Opus 5 separates on novel reasoning: 30.2 on ARC-AGI-3, about four times GPT-5.6 Sol at 7.78 and far above Opus 4.8 at 1.5. Flash has no published ARC-AGI-3 number. HLE with tools sits at 55.3 for Flash, below GPT-5.6 Sol at 64.5, on Z.ai's reporting with GPT-5.6-luna as judge.
• Knowledge work. GDPval-AA v2 Elo at 1773 places Flash above Opus 4.8 at 1588 and above GPT-5.6 Sol at 1736 on Z.ai's chart, but below Opus 5 at 1861 on Artificial Analysis's independent run. The composite AA Intelligence Index places Opus 5 at 61, ahead of Fable 5 at 60 and GPT-5.6 Sol at 59.
• Computer use and frontier coding. Opus 5 leads Frontier-Bench v0.1 at 43.3, over double Opus 4.8 at 21.1 and ahead of GPT-5.6 Sol at 34.4, and leads OSWorld 2.0 at 70.6 versus 62.6 for Sol.
• Cyber. CyberGym at 84.5 is Flash's headline as the best in Z.ai's table, ahead of Opus 4.8 at 78.1 and GPT-5.6 Sol at 83.6 on that specific harness. The fuller GLM-5.3 (non-Flash) extends further on exploitation depth, where Flash was not listed as leader.

GLM-5.3-Flash approaches the closed flagships on coding and agentic benchmarks at the levels Z.ai measured, outperforms its own predecessor by large margins on long-horizon tasks, and trades places with GPT-5.6 depending on the harness. The newest Opus 5 now sets the high mark on SWE-bench Verified and Pro, on ARC-AGI-3, on Frontier-Bench v0.1, and on the AA Intelligence Index, while GPT-5.6 Sol retains its edge on DeepSWE v1.1 and on Terminal-Bench 2.1 in Ultra mode.

Price and availability

Price is where Flash is distinct. The other two are closed and API-only. Flash is both API-cheap and weights-open.

Model	List price per 1M tokens (input / output)	Promotional price	Context window	Max output	License	Availability
GLM-5.3-Flash (Z.ai)	$0.15 / $0.50	$0.075 / $0.25 until Sep 9, 2026 24:00 UTC+8	1M	128K	MIT (weights)	Hugging Face, ModelScope, Z.ai API, Z.ai Coding Plan
GLM-5.3 (full)	$1.40 / $4.40	-	1M	128K	MIT (weights, planned ~2 weeks after launch)	Z.ai API, Coding Plan
Claude Opus 4.8	$5 / $25	-	1M (Opus 5 and 4.8 at 1M; 4.5 at 200K)	128K (300K via Batch beta header)	Proprietary	Claude API, Bedrock, Vertex AI, Foundry
Claude Opus 5 (current flagship, July 24, 2026)	$5 / $25	-	1M	128K	Proprietary	Claude API, Bedrock, Vertex AI, Foundry; retirement not sooner than July 24, 2027
GPT-5.6 Sol	$5 / $30 (API docs list $4 / $20 promotional until Nov 21, 2026)	-	1.05M	128K	Proprietary	API (July 9, 2026 GA), ChatGPT (paid plans), Codex
GPT-5.6 Terra	$2.50 / $15 (docs list $2 / $12 promotional)	-	1.05M	128K	Proprietary	API, ChatGPT, Codex
GPT-5.6 Luna	$1 / $6 (docs list $0.20 / $1.20 promotional after July 30)	-	1.05M	128K	Proprietary	API, ChatGPT, Codex

A few practical notes behind those numbers:

• At list, Flash is one tenth the input and output price of GPT-5.6 Terra, and roughly 33 times cheaper on input than Opus 5 and GPT-5.6 Sol. During the promotional window that started at launch, the gap widens further to about 67 times on input compared with the closed flagships.
• All three families now include a 1M context window at no extra per-token premium on the standard API tiers. Claude Opus 4.8 added premium pricing only for prompts exceeding 200K on early 1M beta; Opus 5 and GPT-5.6 list the 1M window at standard rates, with long-prompt surcharges only for GPT-5.6 requests above 272K input tokens (2x input, 1.5x output for the whole request).
• Cache behavior differs. Opus 5 charges $6.25 per 1M for 5-minute cache writes and $0.50 for cache reads, with 50 percent off via Batch. GPT-5.6 charges 1.25x input for cache writes and 90 percent off for cache reads. Z.ai lists cached input storage as free for a limited time for GLM-5.3-Flash.
• GLM-5.3-Flash was available for local deployment on day one through SGLang, vLLM (via a vendor Docker before upstream merge), and TokenSpeed, with KTransformers guidance for lower-cost setups. GLM-5.3 full weights were not yet posted at Flash launch and were scheduled for about two weeks later after safety hardening.

What Flash can do in practice

In production terms, Flash is built for coding agents and tool-heavy workflows where an open model with low per-token cost matters.

• Coding agents. The combination of Terminal-Bench 2.1, DeepSWE, and NL2Repo shows a model that handles multi-file edits, long command sequences, and repo-scale generation without the cost pressure of a $5 input flagship. Teams that run thousands of agent turns per day see the effect directly in cost per task. Opus 5 at 79.2 on SWE-bench Pro and 68.8 on DeepSWE remains the quality high mark for bug-fix work, while Flash offers volume economics.
• Tool calling and automation. Toolathlon and AutomationBench are the strongest evidence for Flash as a workflow executor on Z.ai's harness. Those are exactly the tasks where an in-house agent makes many tool calls, and where a 48.8 on AutomationBench versus 26.2 for the predecessor signals a qualitative shift from demo to usable.
• Long context with control. The hybrid linear and sparse attention keeps 1M-context serving cost down compared with dense attention, and the Z.ai stack reports parity on retrieval probes while serving on both NVIDIA and domestic accelerator clusters.
• Vision. As the first native multimodal in the GLM-5 line, Flash posts 89.4 on CharXiv Reasoning with tools, 78.0 on Chartography with tools, and 62.4 on OfficeQA Pro on Z.ai's tables, above the Opus 4.8 figures Z.ai listed on those specific vision-with-tools probes.

Where it does not lead: on the most competitive closed-model benchmarks that stress long-horizon planning across very large repos, Opus 5 and GPT-5.6 Sol retain higher scores on SWE-bench Pro (79.2 and 64.6 versus no published Flash number), on ARC-AGI-3 (30.2 for Opus 5 versus 7.78 for Sol versus no published Flash number), and on Frontier-Bench v0.1 (43.3 for Opus 5 versus 34.4 for Sol). Flash also trails its larger sibling GLM-5.3 on Terminal-Bench 3.0 and on deeper exploitation benchmarks such as ExploitBench and ExploitGym, where full-size models with heavier post-training still dominate.

Requirements to run GLM-5.3-Flash locally

Flash is open and self-hostable, but it is not a laptop model. The 320B total size sets a data-center floor. Do not plan for single-GPU consumer deployment.

Supported frameworks

• SGLang - see cookbook entry for GLM-5.3-Flash.
• vLLM - vendor Docker vllm/vllm-openai:glm53-flash with TP=8 and tool and reasoning parsers, upstream support in flight via pull requests 53906 and 36507.
• TokenSpeed - LightSeek recipe for GLM-5.3-Flash.
• KTransformers - tutorial for lower-memory setups with offload.

All four require Hopper or newer NVIDIA GPUs or AMD Instinct gfx950 via ROCm for verified paths. GB10/DGX Spark (sm_121) is not on the vendor verified list.

Memory and hardware

Weights dominate the budget before KV cache:

• FP8 checkpoint (default, zai-org/GLM-5.3-Flash): about 306 GiB for weights alone, 386 GiB is the figure cited on the vLLM recipe including overhead scope, and BF16 doubles weight memory to roughly 598 GiB (about 772 GiB total cited on the recipe page).
• NVFP4 quant (community weight-only, LibertAIDAI/GLM-5.3-Flash-NVFP4): quantizes only the routed-expert FFN tensors (about 97 percent of parameters) from BF16, leaving attention, vision tower, routers, and norms in BF16. That reduces the checkpoint from 598.5 GiB to about 181 GiB, leaving headroom for KV cache. This quant is not vendor-verified and support for the glm5_next architecture is still merging upstream.
• KV cache on top: scales with context length. Budget KV separately. The vLLM and TechCompare estimates put the full 1M-context cache well above 100 GiB, which is why full 1M context on 8x80GB cards is tight.

Deployment	GPUs	Precision	What fits	Command hint
Single-node, 8x H200 (141G)	8x H200	FP8	Full weights plus KV for moderate context	vllm serve zai-org/GLM-5.3-Flash --tensor-parallel-size 8
Single-node, 8x H100 (80G)	8x H100	FP8	Same, with less KV headroom	Same, with lower gpu-memory-utilization
Single-node, 4x GB200 (192G)	4x GB200	FP8 with --kv-cache-dtype fp8 and MTP 5	Reduced interconnect hops	vllm ... --tensor-parallel-size 4 --kv-cache-dtype fp8 --speculative-config ...mtp...
Prefill/decode disaggregation, 8x GB200 split 4+4	8x GB200	FP8	Higher throughput via NIXL KV transfer	Two vllm serve pools with NixlConnector (Hopper must use BF16 KV)
2x DGX Spark (GB10), 121 GiB unified each	2x GB10	NVFP4 weight-only (~181 GiB)	Experimental only, not vendor-verified, requires ~200 GiB cache path and MAX_MODEL_LEN tuning	Community docker-compose recipe at TP=2, after upstream PRs merge

Disk and system memory:

• Download: about 306 GiB for FP8, about 600 GiB for BF16, and more for any cached GGUF variants.
• System RAM: 256 GiB per node is the guidance for NPU deployments; for GPU, size RAM for tokenizer workers and KV offload headroom.
• Storage: NVMe SSD, 500 GiB free minimum, 1 TB recommended.

Single-consumer reality check

Community discussion after launch was blunt for a reason: Flash is roughly 10 times larger than GLM-4.7-Flash (30B-A3B) and about three times larger than the prior Air class (106B-A12B). It will not run on a single RTX 4090 or a 128 GiB Mac Studio, even at Q4_K_M, without severe offload that makes token speed impractical. Single-GPU deployment at BF16 or FP8 is not feasible. Quantized paths like NVFP4 on two GB10s are the closest to consumer, and they remain experimental and unsupported.

When local makes sense

Local deployment pays when data must stay on premises, when burst capacity matters, or when per-token cost at high volume would exceed amortized hardware. For ad-hoc or low-volume use, the Z.ai API at $0.15 per 1M input is the lower-friction path, with the same 1M window and standard tool and reasoning parsers.

How the three compare on the decision

Criterion	GLM-5.3-Flash	Claude Opus 5 (latest, July 24, 2026)	GPT-5.6 (Sol as flagship)
Best use	Open, self-hostable coding and agentic workflows at scale; cost-sensitive 1M-context work	Frontier reasoning with adaptive thinking, SWE-bench Pro leader, knowledge-work leader, 1M context at $5/$25	Frontier coding with highest Terminal-Bench Ultra at 91.9, DeepSWE leader at 72.7, broadest family with ultra subagent mode
Price per task	Lowest: about 3 percent of closed-flagship input price at list	$5 input / $25 output, 50 percent off via Batch, 90 percent off cache reads	$5 input / $30 output Sol at list ($4/$20 promo), Terra at half, Luna at one fifth
Openness	MIT weights, modifiable, commercially usable	Proprietary weights	Proprietary weights
Multimodal	Native image input (tested on CharXiv, Chartography, BabyVision, MVBench, MMVU)	Vision input supported	Image input supported
Speed and efficiency	Hybrid linear and sparse attention for cheaper long context; domestic accelerator support shown	AA Intelligence Index 61 leader; Fast mode 2.5x at $10/$50	Throughput leader via Cerebras up to 750 tok/s; Ultra subagents for parallel tasks
Limits	No published SWE-bench Pro or ARC-AGI-3; trails Opus 5 on Frontier-Bench and SWE-bench; not single-GPU local	No self-hosting; hallucination rate 50 percent on AA-Omniscience, up 14 points vs Opus 4.8	No self-hosting; premium output price at $30 vs $25 for Opus; trails Opus 5 on 9 of 12 shared benchmarks

Which should you pick

• Pick GLM-5.3-Flash when you need MIT-licensed weights, want to self-host or keep data on premises, or run high-volume coding agents where $0.15 per 1M input changes what is economical to automate. It handles 1M-context repo work, strong tool calling, and natively multimodal tasks without a proprietary dependency.
• Pick Claude Opus 5 when the task rewards SWE-bench Pro quality (79.2 versus 64.6 for Sol), novel reasoning on ARC-AGI-3 (30.2 versus 7.78 for Sol), or knowledge-work Elo at 1861, and where enterprise tooling on Bedrock or Vertex AI matters. It is the current AA Intelligence Index leader at 61.
• Pick GPT-5.6 Sol when you want the highest published scores on DeepSWE v1.1 at 72.7, Terminal-Bench 2.1 Ultra at 91.9, and HLE with tools at 64.5, plus the option to fan out via ultra subagents. Use Terra for GPT-5.5-class quality at about half the price, or Luna for the cheapest high-speed tier in the OpenAI family.

GLM-5.3-Flash trails the newest closed flagship on several frontier benchmarks. It resets the price and deployment tradeoff. For coding and agentic work where an open model that runs at one tenth the API cost of GLM-5.3 and approaches Opus-class performance on tool-heavy workloads is sufficient, it is the most capable MIT-licensed option Z.ai has shipped. When the workload is SWE-bench Pro bug fixes, ARC-AGI-3 novel reasoning, or Frontier-Bench agentic coding, Opus 5 remains the verified leader.

Sources

• GLM-5.3-Flash: Frontier Intelligence, Flash Cost, Z.ai, August 26, 2026. Accessed August 27, 2026.
• zai-org/GLM-5.3-Flash - Hugging Face, Z.ai. Accessed August 27, 2026.
• zai-org/GLM-5.3-Flash-BF16 - README - Introduction, Z.ai. Accessed August 27, 2026.
• GLM-5.3: Frontier Coding with Emergent Cyber Capabilities, Z.ai, August 14, 2026. Accessed August 27, 2026.
• Pricing - Z.ai Developer Docs, Z.ai. Accessed August 27, 2026.
• GLM-5.3-Flash - vLLM Recipe, vLLM. Accessed August 27, 2026.
• GLM-5.3-Flash - Unsloth Documentation, Unsloth, August 26, 2026. Accessed August 27, 2026.
• Introducing Claude Opus 5, Anthropic, July 24, 2026. Accessed August 27, 2026.
• Claude Opus 5 - Overview and What's new in Opus 5, Anthropic. Accessed August 27, 2026.
• Pricing - Claude Platform Docs, Anthropic. Accessed August 27, 2026.
• Claude Opus 5 Benchmarks: Every Number, Verified, Jesse Moraga, July 25, 2026. Accessed August 27, 2026. (Opus 5 system card: SWE-bench Verified 96.0, SWE-bench Pro 79.2, Frontier-Bench 43.3)
• Opus 5: Fable 5 level intelligence at a lower cost per task, Artificial Analysis, July 24, 2026. Accessed August 27, 2026. (AA Intelligence Index 61, GDPval-AA 1861, Terminal-Bench 89, hallucination rate)
• Claude Opus 5 vs GPT-5.6 Sol - CodingFleet, July 25, 2026. Accessed August 27, 2026. (12 shared benchmarks, 9-3 Opus 5)
• GPT-5.6: Frontier intelligence that scales with your ambition, OpenAI, July 9, 2026. Accessed August 27, 2026.
• Pricing - OpenAI API Docs, OpenAI. Accessed August 27, 2026.
• Advancing the price-performance frontier with GPT-5.6, OpenAI, July 30, 2026. Accessed August 27, 2026.
• GPT-5.6 benchmarks across Intelligence, Speed and Cost, Artificial Analysis, July 9, 2026. Accessed August 27, 2026.
• BarryDeen - glm53-flash-dgx-spark - NVFP4 notes, GitHub. Accessed August 27, 2026.