Skip to content
BetaBenchAlert v0.1 is in Beta.Numbers are real, but pages and rules can still change.See what changed

Claude Code

Haiku 4.5

Inside Claude Code we ask for claude-haiku-4-5 every hour. Everything below comes from those runs on our lab machine.

Claude Codeclaude-haiku-4-5Think setting highLast round just nowOther models

Write speed, last 24 hours

28.3answer tok/s

8.67s typical total wait

Likely middle: 6.75s9.31s

Visible answer tokens per second across the whole command. Hidden thinking does not make this number larger. Higher is faster. These are the middle values of 24 completed checks out of 24 scheduled in the last 24 hours.

Time to first word
4.53s
How long before the first visible word shows up. Lower is better.
Typical total wait
8.67s
How long the whole command took, start to finish, including the app booting. Lower is faster.
Last ping
3.19s
The newest tiny “reply ok” check. It measures wake-up time, not writing speed.

What stands out

Four facts pulled straight from the saved runs.

Board position
#6 of 10
Just ahead: Fable 5 at 7.90s total wait. Just behind: GPT-5.6 Terra at 10.4s total wait.
Answered when scheduled
24/24
Completed answers out of all scheduled checks in the last 24 hours.
Served as
claude-haiku-4-5
The model name the app reported back. If it differs from the pin, we say so.
Round-to-round swing
2.5×
How much the fastest round beat the slowest one. A big swing means a noisy day.

Recent rounds

Newest first. Fails stay in the list with their reason.

WhenResultWriting speedTotal wait
just nowokclean30.2 tok/s8.11s
1h agookclean36.8 tok/s6.62s
2h agookclean26.0 tok/s9.44s
2h agookclean25.3 tok/s9.69s
4h agookclean28.7 tok/s8.54s
5h agookclean39.3 tok/s6.21s
6h agookclean24.5 tok/s9.94s
7h agookclean23.6 tok/s10.4s

All rounds are in Logs.

Read next

Compare with GPT-5.6 Luna

Its closest rival in Codex — same window, same units, with the leader marked on every row.

Compare with Sonnet 5

The pin next to it in Claude Code. See what the tier jump buys inside one app.

More models

How to read this

Every number here comes from our lab machine running the paid app, one try per hourly round. The live check uses test kit v0.2; recovered history appears only after its raw file passes the same visible-speed rules. Our measurements, not the vendor’s.