Introducing Vantor Halo 2 — faster tool calls and a 400k context window

Today we are releasing Halo 2 to every API account.

What changes

Halo 2 is a drop-in replacement. Change the model name and you are done:

from vantor import Vantor

client = Vantor()
reply = client.responses.create(
    model="halo-2",
    input="Summarise this changelog in three bullet points.",
)
print(reply.output_text)

Deprecations

halo-1 will stop accepting requests in 90 days. We will email every account still sending it traffic at 60, 30 and 7 days.

Questions about the migration? Reply here and the team will pick them up.

The tool-call latency drop is very real. Our dispatch agent went from 2.9s to 1.4s per turn without changing anything else.

Is structured output still guaranteed on Halo 2 when strict: true? I had one malformed JSON in about 4,000 calls over the weekend.

@kenji it should never happen under strict mode — can you send the request id in a DM? We would like to look at that one.

Price drop and faster. Who do I thank?

Confirming the 400k window works with streaming. Fed it a whole repo (310k tokens) and got the first token back in about 4 seconds.