Breaking

Keel 0.2.0 uses Jev to choose coding steps while Claude Opus 5.5 does the reasoning

September 24, 2026 at 18:46 EDT

AI Agents
Software Dev & Coding
Open Source

A public macOS coding workspace now places TypeSafe's Jev model ahead of Claude Opus 5.5: Jev may choose only among options the application has already prepared, and Opus is left with the open-ended reasoning. The developer who published Keel 0.2.0 says that split cut cost and time in their own workflow by about 80%, but the repository does not present that figure as a measured benchmark.

24 Sep 2026 · Keel

Jev chooses. Opus 5.5 reasons.

A macOS coding harness lets a typed model pick a prepared route or tool focus, then leaves open-ended work to Claude Opus 5.5. The builder reports about 80% less cost and time. The release does not treat that as a measured gain.

REPORTED WORKLOAD

An 80% cut leaves about one fifth

100%

Earlier workflow

20%

After the reported cut

Five blocks against one: the same unit, share of the builder's earlier cost and time. This is a personal report, not a coding benchmark.

WHAT THE APP ACTUALLY RUNS

Prepare, pick, recheck, then act

1 · Prepare
Host lists real options

→

2 · Pick
Jev or Laya, or abstain

→

3 · Recheck
Still eligible and current?

→

4 · Act
Run the choice or fall back

~80%

Reported cost and time cut

Focus IDs: inspect, implement, verify, answer

Modes: Laya, Jev, Normal

15+

macOS for the packaged Laya model

What holds up

Opus stays on open reasoning. Jev or Laya may only return a prepared ID, and the host can reject a stale or unauthorized choice.

What is not shown

No measured coding-quality gain, no automatic training, and no control of external agent tool loops. The release check made no paid live Jev call.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.