Agent Completion or Format run: where can you pick Haiku 5.5?

The difference in one line

If you want to choose Claude Haiku 5.5 as the orchestrator of a Sume run, use a Format run. If you call POST /v1/agent/completions, the only value model accepts is sume-agent, and any other value is a 400 invalid_request. Omitting it gives you the same agent.

This matters this week because Haiku 5.5, GLM 5.3 and Mistral Large 4 all arrived within two days, and people will try to paste their ids into the nearest endpoint.

Side by side

Both surfaces run the same agent and return the same receipt shape, per the Agent Completions page. The difference is where the instruction lives and which fields exist.

What to do with a Haiku idea

If the task changes on every call and you cannot save a Format, you cannot select Haiku 5.5 through Agent Completions. You can still save a small Format whose instructions take an input object, then run it with model set. Whether that fits depends on whether the task really changes shape or only its data.

If your cost concern is the orchestrator's tokens, compare it with the spend cap first: the Format docs say usage.billable_amount_usd_micros does not include the agent's own LLM turn, while usage.debited_usd_micros does.

Error you will see

Sending model: "claude-haiku-5-5" to Agent Completions returns 400 invalid_request; the error table on the Agent Completions page lists a model other than sume-agent as a cause. Nothing runs and nothing is charged for a rejected create.

Making the choice

Pick by who knows the task ahead of time. If the task is a repeatable recipe, such as a vertical teaser from a product brief where only the product and the copy change, a Format is the natural home: the Format stores how, the run supplies input, and you can set model. If every call is a one-off prompt and you only care that an agent does it, Agent Completions is the lighter path, and you accept that Sume chooses the orchestrator.

There is a cost angle to the choice as well. Haiku 5.5 lists at $0.10 input and $0.50 output per million tokens on Anthropic's pricing page, read 2026-10-08. At 30,000 input and 2,000 output tokens, that turn is 30,000 x 0.10 / 1M = $0.003 plus 2,000 x 0.50 / 1M = $0.001, or $0.004. Because the media usually costs far more than the turn, most teams will pick the surface for its workflow fit and not for the orchestrator price.

• Choose Agent Completions for ad-hoc prompts with a required spend cap and no saved state.
• Choose a Format run when you need a model id, a stored recipe, previous_run_id continuation or bulk runs.