Contact emails
[email protected], [email protected]

Explainer
https://github.com/webmachinelearning/prompt-api/blob/main/README.md#thinking-mode


Specification
https://webmachinelearning.github.io/prompt-api


Summary
Enables developers to control the reasoning effort and receive thoughts from 
the built-in AI Prompt API (window.LanguageModel). Developers opt in by passing 
options when creating session and/or prompting the model (eg, { effort: "high", 
includeThoughts: true }). High-level semantic effort tiers ("none", "low", 
"medium", "high"), map to underlying configs (eg token budgets). Ephemeral 
reasoning tokens are automatically stripped from the persistent session 
context, and only surfaced to API clients when requested via `includeThoughts: 
true`, for flexible response ergonomics. This feature is an enhancement of the 
Prompt API (https://chromestatus.com/feature/5134603979063296).


Blink component
Blink>AI>Prompt


Web Feature ID
languagemodel


Motivation
Current generation on-device language models (such as Gemma 4) are commonly 
trained as reasoners. By utilizing an internal scratchpad of intermediate 
thinking tokens prior to producing the final response, these models can achieve 
significantly higher accuracy on math, code generation, logical deduction, and 
multi-step planning tasks including those involving tool calling. Today, 
however, the Prompt API offers no API surface to configure or observe thinking 
mode, preventing web developers from leveraging a core capability of the model 
already present on the user's device. Developer opt-in is critical, since 
reasoning chains can increase token generation by 5× to 20× compared to visible 
answers, directly impacting latency and battery consumption on 
resource-constrained client devices. Major LLM APIs (eg Gemini, OpenAI, and 
Anthropic) have converged on exposing explicit reasoning control surfaces. 
Adding thinking options to the Prompt API unblocks reasoning-heavy web 
applications while giving developers fine-grained control over when to pay the 
compute cost of thinking (effort) and whether to stream the intermediate 
reasoning steps to their UI (includeThoughts).


Initial public proposal
https://github.com/webmachinelearning/prompt-api/issues/209


Goals for experimentation
None


Requires code in //chrome?
True


Tracking bug
https://crbug.com/546151578


Estimated milestones

No milestones specified



Link to entry on the Chrome Platform Status
https://chromestatus.com/feature/5098001721851904?gate=5117395982221312


This intent message was generated by Chrome Platform Status.

-- 
You received this message because you are subscribed to the Google Groups 
"blink-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to [email protected].
To view this discussion visit 
https://groups.google.com/a/chromium.org/d/msgid/blink-dev/6abc4dd6.7ebe5f11.106463.0009.GAE%40google.com.

Reply via email to