OpenClaw has added a separate typed Decision Model role that can classify, score and support routing decisions, but the shipped core does not automatically choose among chat models by prompt context; its own documentation says no automatic Decision consumers are connected yet, and no end-to-end cost-or-quality improvement from model routing has been demonstrated.
OpenClaw keeps automatic AI routing out of reach
Software update adds tools for quick judgments but stops short of picking cheaper models on its own
The story continues
TypeSafe AI takes aim at agent token costs
1 October 2026Software developers rush to plug a non-generative decision engine into coding agents to bypass costly text models.
In a nutshell
OpenClaw released a software update that lets automated digital assistants make quick, focused choices without relying on large conversational programs. While third-party makers claim these small models make running software much cheaper, OpenClaw only built the basic controls: the software does not automatically sort questions or switch between cheap and expensive tools on its own.
Highlights
- OpenClaw released version 2026.9.6 on September 23, adding a dedicated setup for quick choices.
- The update supports cloud-based tools like TypeSafe Jev as well as systems running locally on a computer.
- Internal tests of a 4B local configuration recorded a median response time of 154 ms and 78.4% accuracy.
- OpenClaw confirmed that its experimental assistance setting connects to no automatic features.
From the Editor’s Diary
Building the pipes to make quick software choices is straightforward, but making them route everyday tasks cheaply and reliably without human help remains an unproven step.
Who's involved
OpenClaw
an open-source platform that helps people build and run automated software assistants
goal → give its assistants a neutral way to make quick, structured choices without changing their main conversational model
TypeSafe AI
a company that builds Jev, a system designed to return quick answers rather than long conversational text
goal → establish fast decision tools as a cheaper alternative and make Jev work inside platforms like OpenClaw
Josh Lehman
an OpenClaw software maintainer who designed the new decision system
goal → make quick decisions a standard feature across OpenClaw rather than a tool tied to one company
OpenClaw plugin developers
outside programmers who write add-on tools for OpenClaw
goal → use the new setup to sort tasks, set safety limits, and recommend which model should handle a job
In short
OpenClaw has added a way for its software assistants to make quick, narrow judgments without relying entirely on wordy conversational systems. This matters because it gives automated helpers a dedicated setup to answer simple yes-or-no questions, rate risks, and sort requests into categories like billing or technical support.
The most likely next step is that independent programmers will build add-on tools that try to use these quick judgments to cut computing costs.
Whether that saves money is uncertain, because nobody has yet tested whether switching between models inside OpenClaw actually lowers expenses without hurting answer quality.
Previously in this story
TypeSafe hits 13% adoption record at Vercel
26 September 2026A new artificial intelligence model built only to make fast decisions set an adoption record on developer platform Vercel.
TypeSafe AI cuts software costs with Jev
19 September 2026A stripped-down model answers narrow questions to automate computer code without conversational chatter.Browser Use cuts web agent costs to $0.0039
17 September 2026A stripped-down decision engine searches flights in seven seconds without costly visual models.
How it unfolded
TypeSafe announces its decision tool
TypeSafe AI founder Diogo Almeida unveiled Jev on September 15 as a system designed to deliver fast software choices rather than written prose. The release prompted OpenClaw maintainers to reconsider the expense of using full conversational models for tiny internal judgments. On September 18, maintainer Josh Lehman proposed code to give OpenClaw its own dedicated decision setup.
OpenClaw creates a dedicated decision role
OpenClaw updated its software to give decisions their own place alongside conversational models. Lehman's design allowed users to choose a decision setup for all assistants or set different rules for individual helpers, producing clear scores, choices, or true-or-false answers. OpenClaw then added a connector so TypeSafe's Jev could fill that role. While outside observers noted the new structure, some began to speculate that it would soon direct questions to cheaper models automatically.
OpenClaw expands to local models and testing tools
OpenClaw broadened the system so it would not rely on a single vendor. It added support for local servers like Kev and desktop classifiers using the ONNX format. Maintainers also introduced a tool called decision_evaluate, which lets an assistant consult its chosen decision system when asked. A new experimental setting for automatic assistance was added as well, though the team noted that no automatic tools were connected to run it.
OpenClaw defines the line on automation
OpenClaw published guides explaining how the system works: a decision model inspects supplied details against written guidelines and returns a score or label. While an add-on could use that label to direct a task, the platform itself does not switch conversational models on its own. Outside programmers soon filled the gap by building independent add-ons like jev-claw, which evaluate task difficulty and suggest which model to use.
OpenClaw releases version 2026.9.6
OpenClaw released version 2026.9.6 on the evening of September 23. The official release notes made the decision role a standard feature, allowing operators to run cloud-based systems like Jev or local options on their own machines. The update confirmed the underlying plumbing, but it did not include any automatic system to sort user messages, nor did it offer data proving that sorting cuts costs.
Where things stand
OpenClaw 2026.9.6 includes a working framework for quick judgments. Anyone configuring an assistant can attach an external service like TypeSafe Jev or run local tools on their own computer, letting the assistant ask for scores or categorical choices when instructed.
The platform does not automatically read every incoming message to swap between cheaper and pricier models. OpenClaw states that its automatic assistance setting has no active tools attached. Claims of lower costs still come from vendor benchmarks and independent add-ons, while OpenClaw's own tests focus on speed and accuracy. Proving whether this saves money will require an automatic system tested against standard setups.