GLM 5.3 vs Opus 5 vs GPT Sol 5.6: Have open source models finally caught up?

Every few months a new open-weight release forces the same question: can a model you can self-host stand next to the closed flagships on the work that actually ships? GLM-5.3, released by Z.ai on August 14, 2026, is the strongest test of that question so far. Open-weight models have caught up where engineering teams feel it most, on long-horizon agentic coding and automation. The closed models still hold a clear lead on novel reasoning and knowledge work. GLM-5.3's weights were not published at launch; Z.ai released the model through its API and held the checkpoint for a safety review.

This article compares GLM-5.3 against the two closed models it is measured against: Anthropic's Claude Opus 5 (July 24, 2026) and OpenAI's GPT-5.6 Sol (July 9, 2026). It uses published benchmark numbers from vendor system cards, independent trackers, and the CursorBench 3.2 view that reports all three in one place.

The three models at a glance

	GLM-5.3	Claude Opus 5	GPT-5.6 Sol
Maker	Z.ai	Anthropic	OpenAI
Released	Aug 14, 2026	Jul 24, 2026	Jul 9, 2026
Parameters	743B (MoE)	Not published	Not published
Context window	1M tokens	1M tokens	1.05M tokens
API price (in/out per 1M)	Not published for 5.3; GLM-5.2 was $1.40 / $4.40	$5 / $25	$5 / $30
Openness	Open-weight promised late Aug 2026	Proprietary	Proprietary

The price row matters for the conclusion. GLM-5.3 has no published per-token rate; Z.ai's pricing table still ends at GLM-5.2, so any 5.3 cost estimate is a guess. Opus 5 and GPT-5.6 Sol both list $5 per million input tokens, with Opus 5 at $25 output and GPT-5.6 Sol at $30 output.

Coding and agentic engineering: where open caught up

Agentic coding is the benchmark family that predicts whether a model can run a real engineering agent: fixing bugs across a repository, configuring environments, running tests, and recovering from failures. On these, GLM-5.3 lands in the same tier as GPT-5.6 Sol and within reach of Opus 5.

Benchmark	What it measures	GLM-5.3	Opus 5	GPT-5.6 Sol
Terminal-Bench 2.1	CLI agent tasks	88.2%	89.1%	91.9% (Ultra)
DeepSWE v1.1	Long-horizon engineering	66.9%	68.8%	72.7%
SWE-bench Pro	Real GitHub issues	Not published (5.2: 62.1%)	79.2%	64.6%
Frontier-Bench v0.1	Agentic coding	Not published (FrontierSWE 78.1)	43.3%	34.4%
AutomationBench	End-to-end automation	48.2%	26.0%	18.1%

GLM-5.3 improves on every GLM-5.2 coding row, lifting Terminal-Bench 2.1 from 81.0% to 88.2% and DeepSWE v1.1 from 46.2% to 66.9%. On AutomationBench it posts 48.2%, the highest score in Z.ai's own chart across closed and open models. That is the clearest signal that open-weight coding has reached the frontier on process automation.

The gaps remain real. GPT-5.6 Sol leads GLM-5.3 on DeepSWE v1.1 and on Terminal-Bench 2.1 in Ultra mode, where parallel sub-agents push its score to 91.9%. Opus 5 leads on SWE-bench Pro by a wide margin, 79.2% against a model that has not published a comparable SWE-bench Pro number. On Z.ai's own chart, Claude Fable 5 and GPT-5.6 Sol still beat GLM-5.3 on Terminal-Bench 3.0, DeepSWE, and HLE with tools. Kimi K3, an open-weight model already available, edges GLM-5.3 on a few agentic rows. GLM-5.3 leads the open field; it does not lead the entire field.

Novel reasoning and knowledge work: where closed still leads

Reasoning benchmarks measure whether a model solves problems it has not seen in training. ARC-AGI-3 is the hard public version of this test. Opus 5 scores 30.2%, over three times GPT-5.6 Sol's 7.78%, and nearly four times its own predecessor. GLM-5.3 has published no ARC-AGI-3 number.

Opus 5 also posted a perfect 42 out of 42 on IMO 2026 without tools, above the historical gold-medal threshold of 29, and 90.8% on ArXivMath June 2026. GPT-5.6 Sol leads on the older, saturated ARC-AGI-2 at 92.5%, a benchmark with little signal for current reasoning next to ARC-AGI-3.

On human-graded knowledge work, GDPval-AA v2 Elo places the three close together: Opus 5 at 1,861, GLM-5.3 at 1,769 (up from 1,508 on GLM-5.2), and GPT-5.6 Sol at 1,736. GLM-5.3 clears GPT-5.6 Sol here, which is notable for an open model, but Opus 5 still leads the group.

Computer use, browsing, and security

Computer use and tool orchestration are where Opus 5 shows its broadest lead. It scores 70.6% on OSWorld 2.0 and 85.8% on MCP Atlas, against GPT-5.6 Sol's 62.6% and 75.3%. GLM-5.3 has not published OSWorld results, but it posts 28.5% on Agents' Last Exam, a long-horizon agentic test.

On agentic browsing, GPT-5.6 Sol leads with 92.2% on BrowseComp against Opus 5's 90.8%. GLM-5.3 has not published a BrowseComp result.

Security is GLM-5.3's standout category. It scores 84.5% on CyberGym, the top row in Z.ai's launch chart, and 54.4% on ExploitBench. Its weights were staged behind a safety review for this reason. The model leads defensive discovery while trailing on exploit development under time budget, which matches the capability skew its positioning claims.

The open-source asterisk

The title asks whether open source has caught up. That depends on what "open" means in August 2026. GLM-5.3 is API-available today. Its weights were not published at launch, and Z.ai says they will follow roughly two weeks after release, after a safety evaluation. AI Release Tracker classifies the model as proprietary until the checkpoint ships. GLM-5.2 carried an MIT license; GLM-5.3's license has not been published.

A promised checkpoint is not a downloadable one. Teams that need to self-host for data residency, cost control, or vendor independence cannot do so yet, and the open-weights community's quantization and audit cycle has not started.

When the weights land, GLM-5.3 brings a genuine deployment advantage. Its IndexShare architecture reuses the same indexer across sparse-attention layers, cutting per-token compute by 2.9 times at a 1M context, and its speculative-decoding acceptance length is up to 20% higher than GLM-5.2. Z.ai cites a reference throughput near 115 tokens per second. Throughput depends on your serving stack, hardware, and quantization, so the number that matters is the one you measure on your own infrastructure.

Cost and value

Pricing separates the models sharply once GLM-5.3's eventual rate lands near GLM-5.2's $1.40 / $4.40. Opus 5 at $5 / $25 and GPT-5.6 Sol at $5 / $30 cost roughly five to seven times as much per output token. For agents that chain hundreds of tool calls and burn 50,000 to 120,000 output tokens per task, that gap compounds.

A closed model posts the highest single score. GLM-5.3 wins on cost at volume through score per token and score per second. Its Code Bench efficiency result, 31.4% at roughly 50,000 output tokens against a closed model's 29.5% at 120,000, points the same way. For high-volume agent work the cost math favors it. For the hardest single tasks, Opus 5 still returns the best verified result.

How to choose

Pick the model from the workload, then confirm with a shared evaluation on your own tasks. Use Opus 5 when novel reasoning, knowledge work, computer use, or the highest SWE-bench Pro quality carries the decision, and budget the token volume. Use GPT-5.6 Sol when the OpenAI Responses API, hosted tools, and browsing are central to the architecture. Use GLM-5.3 for long-horizon agentic coding and automation once its weights have shipped and you have verified the license and throughput on your hardware.

Run each candidate against one bounded task with a defined trigger, allowed actions, approval boundary, expected final state, and escalation cases. Measure verified completion, wrong reads and writes, recovery after a failed tool call, human review time, latency, and cost per accepted result. Keep model selection behind a stable workflow interface so you can change providers without rewriting product authorization or customer experience.

At Ginger Labs we build embedded agents inside SaaS and web products that progress defined multi-step work using a customer's schemas, records, and permissions. The customer keeps ownership of its API, data model, domain rules, user permissions, tenant boundaries, and definition of a correct result. A model comparison like this one stays useful only when the agent's interface hides the model behind your own authorization and evaluation surface.

Conclusion

Open-weight models have caught up on the engineering work that dominates agentic product roadmaps. GLM-5.3 matches GPT-5.6 Sol and nears Opus 5 on Terminal-Bench and DeepSWE, leads the field on AutomationBench and CyberGym, and beats GPT-5.6 Sol on knowledge-work Elo. They have not caught up across the board: Opus 5 still leads on ARC-AGI-3, IMO 2026, SWE-bench Pro, and computer use, and GLM-5.3's open status is still a pending checkpoint, because the weights have not shipped. Weigh the workload, wait for the weights, and verify on your own tasks before you commit.

Sources

• GLM-5.3 benchmark and release coverage, AI Release Tracker, Aug 14, 2026.
• GLM-5.3 specs, benchmarks, and API, Kingy AI, Aug 14, 2026.
• Claude Opus 5 benchmark record, BenchLM, citing Anthropic's Opus 5 system card.
• Introducing Claude Opus 5, Anthropic, Jul 24, 2026.
• GPT-5.6 launch evaluation table, OpenAI, Jul 9, 2026.
• GPT-5.6 Sol model profile, BenchLM, Aug 12, 2026.
• Claude Opus 5 vs GPT-5.6 Sol benchmark comparison, CodingFleet, Jul 25, 2026.
• CursorBench 3.2, Cursor.