A budget AI model matched its flagship counterpart's accuracy across all 27 test cases but burned through 38,677 tokens compared to the flagship's 14,801 on difficult coding work, according to a benchmark published by The New Stack on August 26. The test compared Z.AI's GLM-5.3-Flash, priced at $0.075 per million input tokens, against the GLM-5.3 flagship released earlier that month at $1.188 per million input tokens, nearly 16 times more expensive. Despite identical accuracy, the models diverged sharply in how they reached their answers, with the cheaper version taking more than twice as long on complex tasks.

On a Python date-parsing function with strict edge cases, both models delivered code that passed all 12 hidden tests, correctly rejecting February 30, interpreting "12/31/99" as 2099, and returning None for dates missing a year. GLM-5.3-Flash required 455.8 seconds and generated 38,677 tokens of reasoning before producing the final 60-line function, while GLM-5.3 completed the same task in 174.7 seconds using 14,801 tokens. Flash's bill came to $0.019 versus $0.065 for the flagship, but if Flash charged flagship rates, its 38,677-token session would have cost $0.16, two and a half times the flagship's expense. On a scheduling puzzle placing five people into five meeting slots under seven interlocking rules, both models found the one valid schedule, with GLM-5.3 finishing in 18.9 seconds using 1,804 output tokens and Flash completing it in 33.5 seconds with 1,003 tokens, costing $0.0003 for Flash against $0.008 for the flagship. In a vendor email extraction test hiding a trap—an 8% discount applying to a revised $73,600 quote for 92 seats rather than the original $68,000—both avoided the error and returned all 10 fields correctly with the accurate final price of $67,712, with Flash winning on speed at 7.7 seconds versus 14.8 and on cost despite using 435 tokens against the flagship's 327.

The report states that Flash's speed isn't guaranteed and depends on work difficulty. According to The New Stack, on the coding task Flash spent 7.6 minutes and produced 38,677 output tokens to reach an answer the flagship reached in under three minutes with 14,801, while on the mid-level puzzle the two came in seconds apart, and on the easy extraction job Flash proved faster. The author writes that Z.AI's efficiency claim should come with a caveat: "Use Flash for your easiest tasks, because while it can handle difficult work, it is far less efficient than the flagship."

The report explains that difficult jobs come down to whether users would rather spend time or money. Flash delivered the same answers at about a quarter of the total cost but took more than twice as long to complete the coding task, revealing that its low per-token price masks the fact it works harder to get the same answer. The author notes that Flash burns far more tokens to do hard work, so its advantage rests on current per-token prices, and providers can change those at any time. The choice between models depends on the task at hand: Flash was faster on extraction but slower on the scheduling puzzle, though both runs finished within seconds. The report concludes that the budget model's low price is covering for the fact that it works harder to get the same answer, with its 38,677-token thinking session highlighting inefficiency that pricing structure currently obscates. Organizations running inference at scale will need to model token consumption patterns alongside sticker price when choosing between speed-optimized and cost-optimized endpoints. The risk of relying on promotional pricing becomes acute when workload patterns shift toward complexity.