Two GLM 5.3 cards sit one row apart on OrcaRouter, and the gap between their rates is the largest inside any single model family in this catalogue: $0.15 input and $0.50 output for the Flash card against $1.26 and $3.96 for the full one, per million tokens [OURS]. That is 8.4 times the input price and 7.92 times the output price. The glm 5.3 flash price query that lands on the cheaper card is the second-biggest price query in this batch, at 198 impressions over 28 days.
The two rate rows, line by line
| Line | Flash | Full | Ratio |
| Input | $0.15 | $1.26 | 8.4× |
| Output | $0.50 | $3.96 | 7.92× |
| Cached read | $0.03 | $0.234 | 7.8× |
| Context | 1,000,000 | 1,000,000 | — |
| Max output | 128,000 | 128,000 | — |
The two cards are the same size in the two places size usually varies. A 1,048,576-token context on both — the card rounds to 1,000,000 — and a 128,000-token output ceiling on both [OURS]. Whatever the 8.4× is paying for, it is not envelope.
Cache reads scale with the same ratio, so caching does not close the gap. A cached workload on the full card costs 7.8 times what the same workload costs on the flash card, before a single output token is generated.
What the benchmark tables say
Both cards carry independent evaluator rows, which makes them directly comparable — a rarity in this batch, where half the rows on most cards are vendor-published.
| Row | Flash | Full | Difference |
| AA Intelligence | 41.8 | 44.8 | +3.0 |
| AA Coding | 71.5 | 74.8 | +3.3 |
| GPQA Diamond | 91.2 | 91.7 | +0.5 |
| Humanity’s Last Exam | 39.9 | 42.3 | +2.4 |
| terminalbench_v2_1 | 84.27 | 83.9 | −0.4 |
| terminalbench_v4_0 | 32.83 | 41.92 | +9.1 |
Two facts stand out.
On the saturated benchmarks, flash is effectively the same model. Half a point on GPQA and three points on the intelligence index is inside the noise band for most buyers. Nobody choosing between $0.15 and $1.26 is doing it to gain 0.5 on GPQA.
On terminalbench_v2_1, flash scores higher. 84.27 against 83.9 is a rounding error in the other direction, and it is worth naming because it breaks the assumption that the expensive card dominates. It does not dominate every row.
The one row where the full card clearly wins is terminalbench_v4_0: 32.83 against 41.92, a 9.1-point gap on the hardest generation of that benchmark. That is the honest case for the full model — not general intelligence, but long-horizon terminal work.
Where the money actually goes
Take a workload of 50 million input and 10 million output tokens a month.
| Flash | Full | |
| Input | $7.50 | $63.00 |
| Output | $5.00 | $39.60 |
| Total | $12.50 | $102.60 |
The flash card is $90.10 cheaper per month on that volume [OURS arithmetic on the published rates]. Put the other way: the full card costs eight times as much for a model that is three points better on the intelligence index and nine points better on one hard benchmark.
Whether nine points on v4.0 is worth $90 a month depends entirely on whether the workload is v4.0-shaped. For chat, summarisation, extraction and short code generation — the tasks the saturated rows describe — the flash card is the same product at an eighth of the price.
The counter-intuitive pair of numbers
Two telemetry readings cut against the price ordering.
The full card is faster. Median time to first token 3,625 ms against 5,522 ms for flash, and 89.9 output tokens per second against 61.7 [OURS]. The cheaper card is the slower card, which is the usual pattern: more traffic, more queueing.
The full card has more traffic than you would expect. 4,511,881,701 tokens over seven days, against 9,291,982,928 for flash [OURS]. Flash carries about twice the volume, not eight times. A card costing 8.4 times more, on half the traffic, is not a card people are avoiding.
That combination says the two cards are not substitutes in practice. Teams route the bulk to flash and the hard cases to the full model, and the hard cases are frequent enough to keep a 4.5-billion-token week on a card that costs eight times as much.
The takeaway
GLM 5.3 Flash costs $0.15 input, $0.50 output and $0.03 cached per million tokens; the full GLM 5.3 costs $1.26, $3.96 and $0.234 [OURS]. Same context window, same output ceiling, 8.4 times the price.
The benchmark tables show three points of intelligence, half a point of GPQA, and one nine-point gap on terminalbench_v4_0 — where flash also loses the speed contest, 61.7 tok/s against 89.9. The 8.4× buys long-horizon terminal work, not general capability.
Sourcing note: both cards’ rates, context windows, output ceilings, quality scores, benchmark values with their `source` fields, and seven-day p50, output speed and token totals are read from the OrcaRouter public model cards for `z-ai/glm-5.3-flash` and `z-ai/glm-5.3` on 2026-10-23 [OURS]. The six benchmark rows compared here are all tagged artificialanalysis.ai, so the two cards were measured on the same harness [INDEPENDENT]. The impression figures come from Search Console for the 28 days to 2026-10-21 [OURS]. The price ratios, the monthly cost table and the row-by-row differences are our own arithmetic. No competing gateway or inference marketplace is named. All values checked 2026-10-23.
Click here for more details .