Reasoning effort levels: GPT-6.1 Sol, Claude, Grok 4.7, DeepSeek

There is no shared effort scale. GPT-6.1 Sol and both Claude 5.5 models offer five levels, Grok 4.7 offers four, and DeepSeek offers low, high and max with other names mapped onto them. On Sume, only the OpenAI and Grok rows show an Effort menu, and it has three choices: Low, Medium and High.

Vendor facts below are from GPT-6.1 Sol, GPT-6 Sol, Anthropic's effort page, Grok 4.7 and DeepSeek's Thinking Mode, all read on 2026-10-02.

What levels does each vendor offer?

The names look the same and are not. Anthropic says Sonnet 5.5's levels are recalibrated, so a level does not produce the same amount of thinking as the same level on Sonnet 5, and it recommends a fresh sweep on your own evals.

Does effort change more than thinking?

On Claude, yes. Anthropic's page says effort affects all tokens in the response: text, tool calls and function arguments, and thinking. Lower effort means fewer and terser tool calls; higher effort may mean more calls and a plan explained before acting. For a video agent that is a direct cost lever, because each tool call is a turn that resends the thread.

Anthropic also says Opus 5.5 defaults to medium while Sonnet 5.5 defaults to high, so omitting the parameter does not give you the same behavior on the two.

What does Sume expose?

In the Sume picker, a row carries parameter definitions only where each one reaches a real knob. The OpenAI rows, GPT-6.1 Sol, GPT-6 Sol and the rest of that family, offer Effort (Low, Medium, High), a Fast service-tier switch and a Thinking switch. The Grok row offers Effort and Thinking, but no Fast. The Claude and DeepSeek rows, which run through OpenRouter, ship no parameters at all, because the repo says the per-vendor reasoning contract behind that hop is unverified and an Effort menu would be a guess.

The product default is medium. The repo notes it is not max because at max a benchmark smoke test spent the whole pre-text window reasoning, with the first assistant token near 51 seconds. xhigh and max are accepted as wire values but are not offered as picker choices.

For Grok 4.7 the repo maps Sume's max to xAI's xhigh, and maps minimal and none to low, because reasoning cannot be disabled there.

What about the API?

The Formats run body documents model as an Agents catalog id and lists no effort field. If you need a specific effort on a model from the API, the docs give you no field to send, so assume the product default and verify on the receipt rather than inferring.

Agent Completions take only sume-agent as model. Choosing a model on that endpoint is not available.

How do you choose a level for a video agent?

Effort is a cost dial as much as a quality dial, so set it per step rather than once for the whole agent.

• Start at the default and raise it only for the planning step, where a better shot list saves a render.
• Do not raise it for routine tool calls such as polling a job.
• Compare usage.debited_usd_micros at two settings on the same input. Effort changes tokens, and tokens are the bill.
• On rows with no effort menu, change the model before you change anything else, since there is no knob.