Which new text models can a Sume Format run name as model?

Short answer

A Format run lets you name the orchestrator with the model field, and Sume's agent model registry on origin/main lists three of this week's four text launches: Claude Haiku 5.5, GPT-6 Astra, and a GLM row called GLM 5.3 Flash. It does not list Mistral Large 4 (preview), and it does not list a GLM 5.3 Fast row. An Agent Completion is different: its model field accepts only sume-agent.

The Format call docs say the field is an Agents catalog id, that an id outside the catalog gets 400 invalid_request, and that the receipt shows the id that ran. The public catalog endpoint lists media and tool models only, so the registry, not that endpoint, is where text models live.

What the registry says today

The table is limited to what the repo shows and what the vendor page says. Sume's registry row names are not all the same as the vendor's API ids, so check the picker spelling before you send an id.

What to send

For a Format run, put the id in the body next to the spend cap. The sample below uses only fields the Format call page documents. Replace the handle and slug with your own Format.

curl -sS -X POST "https://api.sume.com/v1/formats/$HANDLE/$SLUG/runs" \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: haiku-trial-001" \
  -d '{"model":"claude-haiku-5.5","generation_spend_cap_usd":5}'

Where it stops

The model field selects only the orchestrator. The Format's tools still choose the image, video and audio models, so a cheaper orchestrator does not change what a rendered second costs. If you want Mistral Large 4 or GLM 5.3 Fast to drive your own workflow, run it in your code and call Sume's REST endpoints for the media. Sume does not route to either model.

One more check before you rely on this: a model can be in the registry but gated by deployment. The registry marks Haiku 5.5 and GLM 5.3 Flash with the openrouter gate, and Astra with always. Send one cheap run and read the model on the receipt to confirm which id ran.

Checks before you rely on a listing

A registry listing is evidence about what Sume's code knows, not a promise about your workspace on a given day. Three checks are cheap. First, read the model on the first receipt you get back, because the Format call docs say the receipt shows the id that ran. Second, read usage.debited_usd_micros after the run and compare it to the rate card for that model. Third, keep the spelling you sent in your logs, since the registry accepts several spellings for one row (Haiku 5.5 answers to claude-haiku-5-5, claude-haiku-5.5, haiku-5.5 and Haiku 5.5) and a later cleanup could narrow them.

None of this week's launches changes the Format's tools. The model field selects only the orchestrator, so a Format that renders with Seedance 2 renders with Seedance 2 whichever text model planned the shots.

• Vendor ids on the vendor pages read 2026-10-08: Anthropic lists claude-haiku-5-5; OpenAI lists the Astra page under gpt-6-astra; Mistral lists mistral-large-4.
• Sume does not claim Mistral Large 4 or GLM 5.3 Fast anywhere in its registry, docs or public catalog.