Haiku 5.5
Latest Nerf Watch result
On 2026-10-28, Haiku 5.5 scored 92.46 on Nerf Watch's suite 2026.10.1, 1.52 points above its 7-day baseline.
- Latest run
- 92.46 on · suite 2026.10.1
- Previous run
- 100.00 on · suite 2026.10.1
- Change from previous run
- −7.54
- 7-day baseline
- 90.94 · 7 baseline runs
- Change from baseline
- +1.52
- Trend
- Steady
A score below the model's own baseline, the mean of its recent runs. One day's dip can be noise, and a lower score is a lower score on this suite, not proof that a provider changed a model.
Score history
As a table
| Day, UTC | Score | Suite |
|---|---|---|
| 92.46 | 2026.10.1 | |
| 100.00 | 2026.10.1 | |
| 93.75 | 2026.10.1 | |
| 75.54 | 2026.10.1 | |
| 91.91 | 2026.10.1 | |
| 96.86 | 2026.10.1 | |
| 96.54 | 2026.10.1 | |
| 82.01 | 2026.10.1 | |
| 79.41 | 2026.10.1 |
Speed
Response time
On , Haiku 5.5's daily response-time probe recorded a median completion time of 1.6 s across three successful responses at high reasoning effort and Standard/default service.
The daily local probe is scheduled at 00:10 UTC. It uses high reasoning effort and three successful fresh single-turn responses to one fixed small task. A day without three successful responses has no point. Lower is faster; dates are UTC.
Drag across the chart to zoom in.Pinch to zoom; drag with two fingers to move.
Output speed
On , Haiku 5.5's Nerf Watch run recorded 230 output tokens per second.
Tokens per second from the daily Nerf Watch run, scheduled at 01:00 UTC: total output tokens divided by total completion time across its four fixed tasks. Effort is low, except DeepSeek V4 Pro uses high, its lowest available variant. Service is Standard/default, or Ultrafast for the two Ultrafast lines. Output includes reported reasoning or thinking tokens once; OpenCode’s separately reported reasoning is added to output. Timing includes waiting within the measured turn and excludes setup and process exit. Higher is faster; dates are UTC.
Drag across the chart to zoom in.Pinch to zoom; drag with two fingers to move.
Facts from the research notes
API price
No exact-model record in these research notes.
Context window
Published context window recorded in Research note 003, as of 2026-10-13: 1,000,000. Record dated 2026-10-07. Served · API. Total; every current Claude tier had 1M; prompts over 100,000 cost more.
Context-window documentation · Research note 003
Parameters
No exact-model record in these research notes.