GLM-5.3-Flash vs Claude Opus 5 vs GPT-5.6 Sol

Z.ai’s GLM-5.3-Flash turned out to be Ox Alpha, the anonymous model developers had been praising on OpenRouter all week — and the obvious next question is how it stacks up against the models people actually pay for. The short answer: it does not beat Anthropic’s or OpenAI’s flagship, and it is not close on raw intelligence. It beats their mid-tier models, which is where most production traffic actually runs, and it does so at a price that is not in the same universe.

The flagships, head to head

Spec	GLM-5.3-Flash	Claude Opus 5	GPT-5.6 Sol
Intelligence Index	57.0	63.0	58.9
Input / MTok	$0.15	$5	$5
Output / MTok	$0.50	$25	$30
Cached input / MTok	$0.03	$0.50	$0.50
Context window	1M	1M	1.05M
Weights	MIT, open	Closed	Closed

Those Intelligence Index figures come from Artificial Analysis, which runs the same nine-evaluation battery across every model it tracks, so they are the one number here that is genuinely apples to apples. Claude Opus 5 leads the whole board at 63.0. GPT-5.6 Sol sits at 58.9. GLM-5.3-Flash lands at 57.0 — six points behind Anthropic, under two behind OpenAI, and ahead of Claude Opus 4.8 (57.3 is effectively a tie) and GPT-5.5 (56.3), both of which were flagships within the last six months.

Six points is not nothing. On the hardest agentic work — long tool-use chains, multi-hour tasks, the kind of thing that fails silently rather than loudly — that gap shows up as retries and abandoned runs. Anyone reading “within two points of GPT-5.6 Sol” as “the same model for a fiftieth of the price” is going to be disappointed.

The comparison Z.ai would rather you make

The flagship table is not where this model is aimed. Put it against the tiers that carry real production volume and the picture inverts completely.

Model	Intelligence Index	Input / MTok	Output / MTok
GLM-5.3-Flash	57.0	$0.15	$0.50
Claude Sonnet 5	55.3	$2	$10
GPT-5.6 Terra	55.0	$2	$12
GPT-5.6 Luna	51.2	$0.20	$1.20

GLM-5.3-Flash scores higher than Claude Sonnet 5 while costing about 13× less on input and 20× less on output. It scores higher than GPT-5.6 Terra at roughly the same ratio. And against GPT-5.6 Luna — OpenAI’s cheapest tier, the one built specifically to win on price — it is both cheaper on every axis and nearly six points more capable. That is the actual competitive claim, and it is a strong one.

Agentic coding: read the footnotes

Terminal-Bench v2.1 is the benchmark everyone quotes for terminal-driven agent work — 89 curated tasks from the Laude Institute and Stanford researchers. On Artificial Analysis’s run of it, GPT-5.6 Sol at extra-high effort scores 89.5%, Claude Opus 5 at max effort scores 89.1%, and Grok 4.6 scores 88.4%. Z.ai reports 84.3% for GLM-5.3-Flash.

That last number deserves a caveat, and it is the reason this section is not a simple ranking. Z.ai’s 84.3% is self-reported, measured on the company’s own harness. Artificial Analysis has not yet published an independent Terminal-Bench figure for GLM-5.3-Flash, so it is not sitting on the same leaderboard under the same conditions as the 89.5% and 89.1%. Agent benchmarks are notoriously sensitive to scaffolding, retry policy and effort settings; a five-point spread across two different harnesses is within the range those choices alone can produce. Treat it as “in the neighbourhood”, not as a verified fourth place.

One more framing detail worth noticing: Z.ai’s own comparison table benchmarks GLM-5.3-Flash against Claude Opus 4.8, not the current Opus 5. Against 4.8 it comes within seven-tenths of a point on Terminal-Bench 2.1 (84.3 vs 85.0) and within half a point on Z.ai’s internal code benchmark. Both are true. Both are comparisons against a model Anthropic has already superseded.

What you give up

Three things, and none of them show up in a price table. The first is speed: Artificial Analysis clocks GLM-5.3-Flash at 50.2 output tokens per second against a 65.3 median for its size class, with 1.47s to first token. For a model with “Flash” in the name it is slower than its peers, and in an interactive agent loop that is felt more than a benchmark point is.

The second is verification. The frontier numbers here have been run by a neutral third party; the GLM coding numbers largely have not. That is probably a matter of time — Artificial Analysis will get to it — but until then the asymmetry is real.

The third is the operational fine print. The weights are MIT-licensed and on Hugging Face, which is a genuine advantage: you can self-host and nothing leaves your network. But that means roughly 306 GiB in FP8 and Hopper-generation NVIDIA silicon or newer. If you use the hosted API instead, you are sending code to a Chinese provider, and during the anonymous preview the data-retention terms differed by route. For a lot of enterprise buyers that single fact settles the question before any benchmark does.

So which one

If the work is hard agentic engineering and a failed run costs more than the tokens — pick Opus 5. Its six-point lead is real and it is the reason the price holds. If you are already in OpenAI’s ecosystem and want the top of it, GPT-5.6 Sol is within touching distance of Opus 5 on Terminal-Bench for the same input price. If you are running high-volume, latency-tolerant work — bulk refactors, test generation, log triage, anything where you would otherwise reach for Sonnet 5, Terra or Luna — GLM-5.3-Flash is now the strongest thing in that price bracket by a clear margin, provided you can send the data.

The interesting shift is not at the top of the table. Anthropic and OpenAI still own the frontier, and nothing here changes that. What has changed is that the cheap tier is no longer where you go to accept a worse model — and both labs price their mid-tier as though it still is.