jev-latest now gives Jev 1.14; decision leaderboards and four new graded studies (24 Oct 2026)

This dated report records the evidence behind the update of 24 October 2026. TypeSafe moved its jev-latest alias from Jev 1.13 to Jev 1.14 between two daily checks, without a dated announcement; the pages that name a default model now say what an unpinned request gets. The benchmarks page gains a section on the four decision leaderboards that list Jev, with vendor rank claims checked against each leaderboard's own data, and four newly graded studies. No Jev call was made, no route was called, nothing was submitted to a leaderboard and no benchmark was run.

What the move means if you do not pin a version

• SDK defaults. TypeSafe's Python and TypeScript SDKs and the Vercel AI SDK provider send jev-latest unless you name a model, and their versions did not change (typesafe-sdk 0.7.5, @typesafe-ai/sdk 0.7.0, @ai-sdk/typesafe-ai 3.0.26; registry check 24 Oct 2026, 05:18 UTC; R23-S41). On TypeSafe's API those requests now get Jev 1.14 with no change to your code. Details: Python, TypeScript.
• Integrations that send jev-latest. Of the Claude Code and MCP integrations read in source, oh-my-claudecode, jkudish/jev-mcp, claude-model-guard and claude-jev-advisor send jev-latest, so on TypeSafe's API they now get 1.14; ohmyjev pins jev-1.13.0 and does not move. These pins are as published on Claude Code and MCP; the files were not re-read on 24 Oct.
• Thresholds. TypeSafe's models page: "If you have tuned confidence thresholds against one version, send that version's ID instead of the alias, and switch to the new version when you are ready." Each response's model field gives the versioned ID that answered (R23-S05). How to set a cutoff: confidence thresholds.
• Jev 1.13. jev-1.13.0 is still listed with its own ID and the same price ($0.042 per million input tokens, output free) and no deprecation date. TypeSafe's 14 Oct post says it "will be available for a while" and that TypeSafe intends to give "60 days warning" before deprecating it (R23-S05, R23-S09). Whether a request by that ID still works was not tested.
• Limits. The models page gives Jev 1.14 the same context limits and price as 1.13; its only linked accuracy page is still the one for Jev 1.13 (limitations).

Six routes: what each publishes for jev-latest

So the same alias name gave a different version on Opper than on TypeSafe's own API that morning. Send a versioned ID if the version matters. Full table: versions and aliases. Documented

Four decision leaderboards that list Jev

Each leaderboard was read at a pinned commit or with its access time, and Jev's position was counted from the operator's own data. Positions are listings at one time, not accuracy results; they sit outside the ledger, carry no grade and are not combined. Reported

• Jev Decision Index (community Hugging Face Space, "not affiliated with TypeSafe AI"; version 0.4.1 at commit 0b5ecaf, 22 Oct): the headline score rests on private tests that are "never published"; per-model scores are public for 37 benchmarks. Jev 1.14: 60.24, 41 of 229 as listed. Its own history shows the "Decision Index 0.2" that an earlier ledger row (B18) cites.
• Artificial Analysis Decision Index (version 1.0; read 05:16 UTC): Jev 1.14 57.07, 17 of 32 entries as listed; per-item results not found.
• JevMan (Opper's Pac-Man benchmark; Opper serves the listed models): jev 1.14 mean score 3597, 1 of 12 as listed in the 21 Oct file, next 3416; the page's tie rule was not applied.
• JevBench (Leanmcp; data of 16 Oct, not pinned to a commit): Jev 1.13 76.5, 3 of 18 as listed; no Jev 1.14 row.

Claims checked against the data: Base Labs' "#1 of 208" for mammoth-1 matches at 20 Oct (71.24, first of 209 entries) and not at 22 Oct (68.44, shown "=3"); elephant-1 "beating Jev 1.14" matches by 0.02 points, inside one tie band; Jeffan-31B's "#1" matched as a tie on 21 Oct and not from 21 Oct, 22:37 UTC; Baseten's Artificial Analysis figures (65.9 and 65.2) do not match the 24 Oct listing; the Artificial Analysis article's "Jev 1.13 leads ... with 58" does not match its live data; Eevo's JevMan claim holds against Jev 1.13, not 1.14; Decido's "published" Jev mean is its own average and cannot be checked. Every claim and its evidence: benchmarks.

Four newly graded studies

The ledger now has 45 graded rows: A 14, B 14, C 10, D 7. Graded comparisons with OpenAI's Decisions API: 6; with Microsoft-Decision-1: 2; with SubQ Snap: 2. Against small text-generating LLMs the count is unchanged (ahead in 6 of 10 graded studies). egarim/decision-1-vs-jev was read and is not a Jev comparison: it compares Microsoft-Decision-1 with a self-hosted reproduction named open-jev. Nine benchmark-named repositories created on 23–24 Oct are listed as leads, not opened. Pages that quote these grades were updated with the same grades and flags: alternatives, the decision-API comparison and reranking and RAG.

Access and notices

Checked 24 Oct 2026, 05:12 UTC: signups open, no free credit for new users, API operational ("All services are online"; no new incident since 14 Oct); price unchanged at $0.042 per million input tokens for both versions. Home shows three notices: no credit for new users, the jev-latest move (added; it comes off after seven days or sooner if superseded) and lookalike resellers. None was retired. Details: access status. Documented

What the daily scan covered

Window: 23 Oct 2026, 05:18 UTC to 24 Oct 2026, 05:11 UTC (23 h 53 min), with no gap or overlap with the previous window, plus a 72-hour look-back. Found: the alias move; Opper's alias lag; OpenAI's post on general availability (not opened); an opinion piece on Jev's hype cycle (not a catalogue item); a coding-agent review-loop article (lead); nine benchmark-named and about twelve other new repositories (names only). Lead checks: the eu/jev pull request to LiteLLM (#46875) is still open; IntelligenceLab/Jev-Vision models and sdmlai/nano-jev are Jev-named and not published by TypeSafe. Not searchable or not reachable: Reddit (HTTP 403), X, YouTube, TikTok and GitHub code search.