A model string one character off bills cached tokens at four times the rate

Both spellings are real entries, so both calls succeed and both responses look fine ^[9]. Model identifiers are opaque strings with no type system behind them, an editor will not autocomplete one unless you built a constant, and a test suite will not fail on a string that is valid for a model you did not want ^[8]. The gateway concentrates the problem, because it folds several vendors' naming conventions into one namespace and each of the dangerous pairs is one character apart ^[10].

The gap to price is in the cache-read column. On the claude pair it is a factor of four ^[17]. The post gives input and output prices per 1M tokens and does not spell out the unit on the cache-read column ^[3]; if cache reads bill in the same per-million unit, the 0.3 difference works out to 300 dollars per billion cached-read tokens ^[18]. The grok pair's gap is about 67 percent ^[19]. glm-5.2 and glm-5.3 differ in every field, cache reads included, 0.169 against 0.22, roughly 30 percent apart ^[6][20].

The script the post publishes fetches the pricebook, groups names by similarity, compares the price fields of each near-duplicate pair, warns on any divergence, and treats the pricebook as the source of truth ^[12]. It uses the standard library plus difflib's SequenceMatcher, with SIMILARITY_THRESHOLD set to 0.85, and the post labels it "Illustrative example. Not production code." ^[13]. Consider where a fixed ratio threshold bites hardest: "grok-4.5" is eight characters, so one differing character is 12.5 percent of the string, against 6.25 percent for the sixteen-character "gemini-3.7-flash" ^[22]. Turn the threshold up to cut false positives and the short names fall out of the comparison first.

The gemini pair is the one a price comparison cannot flag. gemini-3.7-flash and gemini-3.8-flash match on all three fields shown, and the post calls that its own trap: price alone will not tell them apart ^[7]. A warning conditioned on divergent prices stays quiet there, so pinning that identifier takes the repo-side constant, not the pricebook diff ^[8].

The post is blunt about the alternative. The usual advice is "read the pricing page carefully", and, as the post wrote, "That's not a control. It's a hope." ^[14]. It is also careful about what the numbers are: a lower cache-read price means cached input tokens are billed at a lower rate, and the source file contains price fields only, so nothing in it speaks to latency, throughput or quality ^[15]. Vendors ship point releases into a shared namespace whichever way their own convention runs, with a hyphen, a dot or a word ^[11]. Eight of the pricebook's 33 entries sit in the four pairs the post calls out ^[21].