When we say Shogo’s unlimited plans deliver 12x the work per dollar, that number isn’t a marketing round-up — it comes straight from a head-to-head benchmark where both stacks completed the same tasks. This is the math.
What we measured
We ran the full Aider Polyglot coding benchmark — 83 real Exercism tasks spanning 49 JavaScript and 34 Python exercises — on two stacks:
- Shogo + Hoshi 1.0 (our own model), and
- OpenClaw + Claude Haiku 4.5.
Crucially, the scores are matched: Hoshi solved 83/83 and Haiku solved 81/83. Because capability is essentially equal, the only interesting variable left is cost — which makes this an apples-to-apples comparison of work per dollar, not a cheaper-but-worse trade-off.
The rate card
The gap starts with the underlying token prices. Per million tokens:
| Hoshi 1.0 | Claude Haiku 4.5 | |
|---|---|---|
| Input | $0.14 | $1.00 |
| Cached input | $0.01 | $0.10 |
| Output | $0.30 | $5.00 |
Hoshi 1.0 is roughly 7–10x cheaper on input and about 18x cheaper on output — and output is the dominant cost axis for code generation, where the model is writing far more than it reads.
Cost per task
Applied across all 83 tasks, those rates produce a wide spread in real spend:
| Shogo + Hoshi 1.0 | OpenClaw + Haiku 4.5 | |
|---|---|---|
| Total cost (83 tasks) | $0.41 | ~$5.2 |
| Cost per task | $0.005 | ~$0.06 |
A half-cent per task versus about six cents. That’s the ~12x figure: $0.06 ÷ $0.005 ≈ 12.
Methodology and honesty notes
We want this number to hold up, so here’s exactly how it’s derived — including where it’s exact and where it’s an estimate:
- Hoshi’s cost is exact. It’s summed from per-task token logs (input, cached input, and output tokens) recorded during the run, multiplied by the rate card above. Hoshi’s total breaks down as input $0.12 + cache $0.11 + output $0.18 = $0.41.
- Haiku’s cost is an estimate. The OpenClaw harness logged only pass/fail and duration — not token usage. So we estimate Haiku’s spend by applying its list rates to Hoshi’s measured token workload on the identical 83 tasks. It’s a proxy, not a metered bill.
- The realistic range is 12–30x. The ~$5.2 figure assumes Haiku gets the same aggressive prompt-caching Hoshi did. With little or no cache reuse, the cached tokens bill at full input rate and the estimate climbs toward ~$15 (≈30x). We headline the conservative end — 12x — on purpose.
In short: the rate-card multiples (7–18x) are exact list prices, and the per-task multiple lands at a defensible ~12x for identical results.
Why this powers unlimited plans
This is the reason Shogo can offer unlimited agent usage at the same moment the big providers are pulling it back. When each unit of work costs roughly a twelfth as much, the same usage window stretches about 12x further — so “unlimited within generous fair-use windows” is a sustainable default for us rather than a loss-leader we have to ration or ban.
It’s the same story we told when we launched the model: Hoshi 1.0 beats OpenClaw + Haiku 4.5 on every eval — and does it an order of magnitude cheaper. The unlimited plans are simply that economics, passed to you.
About Shogo
Shogo is the platform where AI agents build the software you need instead of replying with text. See what unlimited gets you on a paid plan, or deploy your first agent in minutes on the free tier.