Medium offered the best balance of performance, cost and generation time across our Haiku 5.5 tests: a voxel pagoda, a 3D seaplane and a kart racer.
It scored 987 overall for an estimated $0.04628 across the three generations. High got to 1,007, but took 59 minutes in total against Medium’s 21 minutes 10 seconds. At these prices, the waiting is a bigger consideration than the cost.
- Capabilities Index
- 987Voxel 1,024 · 3D 952 · GameDev 986
- Estimated cost
- 4.6¢$0.04628 across all three generations
- Generation time
- 21m 10sSum of the three recorded durations
Haiku 5.5 arrived on 7 October, bringing adjustable effort to Anthropic’s smallest model. We tested all three settings in Claude Code. The results aren’t evenly good: the pagodas are impressive for the price, the seaplanes are weaker, and Low’s game doesn’t run.
The pagoda is the one to open first
Even Low gets the basic scene together: a multi-tiered pagoda, cherry trees, a gate and paths around a small island. Medium adds a fuller garden. High goes further with layered terrain and dense planting around the building.
The framing is frustrating. Parts of the structure are cut off, and the dark façades make some of the detail hard to see. We’ve kept the original camera views here, including those problems. Still, getting this much into a scene for less than two cents is a good result.
Low falls down on the game
The Low submission loads a missing Three.js dependency, then throws before the race can start. It gets our fixed 800 task-failure score. We didn’t repair it and score the repaired version.
Medium and High both produce playable racers. Our checks reached a completed race and a working restart using the original game logic. High scores a little better, but they’re close: 986 for Medium, 997 for High.
There are still bugs. High’s mute control stops the music but leaves engine and effects audio playing. Some item checks use inventory supplied by the test harness, so those passes don’t tell us that every pickup works naturally during a race. These are playable prototypes, with more testing and fixing needed.
What the extra time bought
| Effort | Voxel | 3D | GameDev | Overall |
|---|---|---|---|---|
| Low | 1,008 | 905 | 800 | 904 |
| Medium | 1,024 | 952 | 986 | 987 |
| High | 1,044 | 980 | 997 | 1,007 |
Most of the overall improvement comes between Low and Medium, helped by getting a working game. High improves all three category scores again, but by smaller amounts.
The seaplane scores go from 905 to 952 to 980. Low’s aircraft is almost black in its original rendering. Higher effort helps, although even High remains below the stronger 3D entries. You can compare the captures in the Seaplane archive.
High costs about 8.2¢ for the suite, against Medium’s 4.6¢. Its 20-point overall gain took an extra 37 minutes 50 seconds. High produced the most detailed pagoda and the highest scores, while Medium delivered working outputs on all three tasks in substantially less time.
Sources & references
- Claude Haiku 5.5: configurations and evidence ↗
The registered Low, Medium and High outputs in Claude Code. Article figures use the 10 October 2026 data snapshot.
- Lumina Capabilities Index ↗
Frozen article snapshot: capabilities-index-v4-309a588efc1a. Linked live charts may change as the dataset grows.
- Scoring and resource methodology ↗
Three equally weighted categories, final rounding, failure scoring and estimated API-equivalent resource costs.
- Anthropic’s Haiku 5.5 announcement ↗
7 October 2026. Context for the model’s release and adjustable effort settings.
