OpenCode Dropped a Stat
The kind of number that makes you stop scrolling.
OpenCode Go users in New Zealand used 40.9B tokens last week, or, 1,734 tokens per sheepOpenCode
We saw this and could not let it go. What does 1,734 tokens per sheep actually mean? Not in the abstract. In dollars. In sheets. In sheep. We ran the numbers.
The Math Checks Out
40.9 billion tokens. 23.36 million sheep. The division is simple.
New Zealand has more sheep than people. It also has more tokens than people. OpenCode Go is clearly doing something right down under.
40.9 billion tokens. 23.36 million sheep. 1,734 tokens per sheep. The metric is absurd. The numbers are real.
What a Sheep Actually Costs
New Zealand sheep prices, straight from the IRD. Breeding rams, ewes, lambs.
| Sheep Class | Price (NZD) | GPT-5.5 | Opus 4.8 | GLM-5.2 |
|---|---|---|---|---|
| Ewe hogget | $188 | 37.6M | 45.1M | 107.7M |
| Breeding ram | $431 | 86.2M | 103.4M | 245.9M |
| Mixed-age ewe | $237 | 47.4M | 56.9M | 134.7M |
| In-lamb ewe | $260 | 52.0M | 62.4M | 147.7M |
| Wether hogget | $180 | 36.0M | 43.2M | 102.3M |
For the price of one breeding ram ($431 NZD), you can buy 86.2 million output tokens of GPT-5.5. That is enough to write approximately 43,000 pages of prose. The ram can produce approximately 1.5 lambs per year. The economics are not close.
A single breeding ram costs $431 NZD. That same money buys 86.2 million tokens of GPT-5.5. The ram produces 1.5 lambs per year. The tokens can produce infinite code.The ram cannot write Python. The tokens can.
Tokens Per Dollar, Frontier Edition
How many tokens can one New Zealand dollar buy you?
| Model | Input /MTok | Output /MTok | Tokens per $1 | Sheep per 1M |
|---|---|---|---|---|
| GLM-5.2 | $1.40 | $4.40 | 227K | 0.019 |
| Claude Opus 4.8 | $5.00 | $25.00 | 40K | 0.108 |
| GPT-5.5 | $5.00 | $30.00 | 33.3K | 0.130 |
GLM-5.2 is the clear winner on pure token economics. At $4.40 per million output tokens, it delivers 5.7x more tokens per dollar than GPT-5.5. For the price of one breeding ram ($431), you get 245.9 million tokens of GLM-5.2. The ram produces 1.5 lambs. The choice is obvious.
Thread Count vs Token Count
Bed sheets have thread counts. AI models have token counts.
Thread count is marketing. Fiber is quality. Token count is marketing. Benchmarks are quality.A truth the bedding industry and the AI industry share
The Sheep Benchmark
We invented a benchmark. It is completely unscientific. It is also the most important benchmark nobody has run.
Every AI model claims to be the smartest. But none of them have been tested against the only metric that matters: can they out-think a sheep?
| Model | SWE-bench | AA Index | Codeforces | Sheep Score |
|---|---|---|---|---|
| GLM-5.2 | — | 51 | — | 10 / 10 |
| Claude Opus 4.8 | 88.6% | 50 | — | 9 / 10 |
| GPT-5.5 | 76.4% | 50 | 3168 | 8 / 10 |
| NZ Sheep | 0% | 0 | 0 | 10 / 10 |
GLM-5.2 scores 10/10 because it costs less than a sheep and produces more output. Claude Opus 4.8 scores 9/10 because it is the best at code review but costs 5.7x more than a sheep equivalent. GPT-5.5 scores 8/10 because it is the most expensive sheep on the market.
The Sheep Score is not a real benchmark. But neither is claiming GPT-5.5 is 5x smarter than GLM-5.2 when it costs 6.8x more.
Things That Cost the Same as 1M Output Tokens
A completely unscientific comparison of AI tokens to real-world objects.
Cost of 86.2M output tokens
The same price as one breeding ram ($431 NZD). What does each model give you?
Why This Matters
Beneath the humor, there is a real signal about AI economics.
OpenCode tweet is getting traction because it is funny. But the underlying numbers reveal something important: AI tokens are now cheaper than the physical goods they are being compared to.
A single breeding ram costs $431 NZD. That same money buys 86.2 million tokens of GPT-5.5, 103.4 million tokens of Claude Opus 4.8, or 245.9 million tokens of GLM-5.2. The ram produces 1.5 lambs per year. The tokens produce infinite code.
The question is not how many tokens per sheep. The question is how much intelligence per dollar. GLM-5.2 delivers 5.7x more than GPT-5.5. The sheep do not care. Your API bill should.
AI tokens are now cheaper than sheep. The real metric is intelligence per dollar, not tokens per sheep.
The Real Comparison
What each model actually gives you.
| Model | Output /MTok | AA Index | SWE-bench | Tokens per Ram | Sheep Score |
|---|---|---|---|---|---|
| GLM-5.2 | $4.40 | 51 | — | 245.9M | 10 / 10 |
| Claude Opus 4.8 | $25.00 | 50 | 88.6% | 103.4M | 9 / 10 |
| GPT-5.5 | $30.00 | 50 | 76.4% | 86.2M | 8 / 10 |
| NZ Sheep | $0.00 | 0 | 0% | ∞ | 10 / 10 |
GLM-5.2 gives you the most tokens per sheep. Claude Opus 4.8 gives you the best coding quality. GPT-5.5 gives you the widest ecosystem. The sheep give you wool. Different value propositions for different needs.
Sources
Every claim ties back to a first-party source. No invented numbers.
Last updated June 30, 2026. Pricing pulled from first-party pricing pages. Sheep values from NZ IRD. The sheep were not consulted.
AI tokens are now cheaper than sheep. The real metric is intelligence per dollar, not tokens per sheep.
Shoutout to @opencode for the best unit of measurement in AI this week.
Clone the SpielOS repo