· via dev.to (home feed)
Output tokens are 17% of LLM volume but half the API bill, pricing analysis finds
A dev.to analysis prices Gemini 4 Neon, Claude Opus 5.5 and GPT-6.2 Sol on an identical workload and finds output tokens, a sixth of the volume, drive half the bill — with cached-input rates making honest comparisons harder.

A fixed workload, four price tags
An analysis published on dev.to compares the API pricing of Gemini 4 Neon, Claude Opus 5.5, Claude Opus 5.5 (fast) and GPT-6.2 Sol on an identical workload held constant across models: 10 million input tokens and 2 million output tokens per month, with no cache hits. The whole cost model fits in one formula, according to the author: monthly cost equals input tokens times the input rate plus output tokens times the output rate, with rates quoted per million tokens and cached input billed as a separate line.
Using rates read directly from the vendors' pricing pages on 18 and 20 November 2026, that workload costs:
- Gemini 4 Neon at $3.00 in / $15.00 out: $60 per month, $720 per year
- Claude Opus 5.5 at $4.00 in / $20.00 out: $80 per month, $960 per year
- Claude Opus 5.5 (fast) at $8.00 in / $40.00 out: $160 per month, $1,920 per year
- GPT-6.2 Sol at $2.00 in / $10.00 out: $40 per month, $480 per year
The headline illustration: Gemini 4 Neon and Claude Opus 5.5 differ by exactly $1 per million input tokens, and on this workload that $1 becomes $240 over a year.
Output tokens are a sixth of the volume, half the bill
The analysis' central finding is that output tokens dominate costs out of proportion to their share of traffic. On Neon, 10 million input tokens at $3.00 cost $30, and 2 million output tokens at $15.00 also cost $30 — meaning exactly half the bill comes from 16.7% of the tokens.
According to the dev.to post, this is not a quirk of one vendor: every model in the comparison prices output at exactly five times its input rate, and in every row the output line lands at 50% of the total. The ratio of output to input tokens is what actually decides the bill, the author argues, and it is usually the number nobody measures. Chat-shaped applications often have an output share even higher than the 1:5 ratio used here.
Cached input breaks the comparison
The second finding concerns prompt caching. Neon publishes a cached-input rate of $0.15 per million tokens, which is 5% of its $3.00 input rate. Anthropic's page, as read by the author, does not publish a cached-input rate for Opus 5.5, and the author declines to estimate one.
This matters as soon as prompts carry a long shared prefix — a system prompt, a tool schema, or a document that gets resent on every call. If 80% of input hits the cache, the input side of the Neon calculation falls from $30 to roughly $7.20. Two models that both publish cache rates can be compared on cache-heavy traffic; a pair where only one does cannot. The author specifically warns against calculators that quietly assume zero cache hits for one model and the published rate for the other.
The fast tier doubles everything
Claude Opus 5.5 (fast) is priced at $8.00 in / $40.00 out — cleanly double the standard tier. On the same 12-million-token workload that comes to $160 per month, or $1,920 per year, which is $100 per month more than Neon and 167% higher overall. The author's advice: if a latency-sensitive feature justifies the fast tier, that multiplier belongs in the design document rather than appearing later as an invoice surprise.
How to reproduce the numbers
The post lays out a repeatable method: pull last month's actual input and output token counts from the provider's usage dashboard rather than guessing the split; multiply each by the per-million rate from the vendor's pricing page and record the date it was read; split cached and uncached input if caching is in play; and re-run the calculation whenever a vendor changes a rate.
The author lists caveats: all yearly figures are simply monthly costs times twelve, and no volume discounts, committed-use pricing or free tiers were applied because none of the pricing pages read listed them for these models. Prices change with little notice, so the source pages should be checked before any number goes into a budget. The author maintains a public matrix of rates with the dates they were read at modelcost.org, and notes the draft was written with AI assistance but that every price and calculation was checked by hand.
Why it matters
Most pricing comparisons stop at the three headline numbers per million tokens, and developers naturally anchor on the input rate. This analysis shows why that misleads: the token mix — above all the output share and the cache hit rate — decides the bill far more than a $1 difference in input price. A team that budgets on input rates alone can be off by hundreds of dollars per year per workload, and a team switching to a fast tier can double costs without changing a single line of application code. Measuring your own output-to-input ratio before choosing a model is the cheap fix.
- #llm
- #api-pricing
- #gemini
- #claude
- #openai