OpenRouter, Cloudflare, and Netlify start serving Jev

On September 18, OpenRouter announced beta support for Jev. The listing adopts TypeSafe’s stated pricing and latency: $0.042 per million input tokens, free output, response times between 70 and 500 ms, and claimed speedups over comparable language model workflows. The endpoint differs from standard chat completions, routing instead through POST https://openrouter.ai/api/alpha/decisions under the model identifier ~typesafe/jev-latest, which resolves to Jev 1.13. OpenRouter also offered $20 in platform credits to ten projects tagging @jjacky.

A day earlier, Cloudflare developer and AI lead Rita Kozlov shared that Jev was available on Cloudflare AI Gateway following public coordination with TypeSafe. Cloudflare serves typesafe/jev through Workers AI (env.AI.run) with a listed 32,000-token context window, while billing rates are managed through the Cloudflare dashboard.

Netlify announced similar availability in a changelog update on September 17. Developers can access Jev through Netlify AI Gateway by installing @typesafe-ai/sdk, initializing a TypeSafeClient within Netlify Functions, and billing usage directly to platform credits. The accompanying sample demonstrates routing incoming contact submissions between sales, support, and spam.

With Vercel having added typesafe-ai/jev on September 16, developers can now access the model through Vercel, Cloudflare, Netlify, or OpenRouter without waiting for direct access from TypeSafe.

Provider listings reflect slight differences in categorization and limits. OpenRouter labels its endpoint as alpha and the model as beta. Cloudflare lists Jev under text generation. While Netlify and OpenRouter cite a 32,000-token context window, TypeSafe’s documentation specifies 64,000 tokens for the full request and 32,000 tokens for state combined with the longest question.

On September 24 at 16:16 UTC, Diogo Almeida posted a screenshot of OpenRouter’s context-length chart. The filter is prompts and completions of 1K to 10K tokens, and the count is absolute requests. The ranking under the chart lists Jev 1.13 by typesafe first, at 242 million requests and 15.1%. DeepSeek V4 Flash 0731 is second at 160 million and 9.9%. DeepSeek V4 Flash 0423 is third at 120 million and 7.5%. Gemini 2.5 Flash Lite is 69.5 million and 4.3%. GPT-5.6 Luna is 67.3 million and 4.2%. The date labels on the same image run from Mar 30 to Sep 14. The caption says Jev is the top model in that band. It does not print a separate total for the week of the September 15 launch. Fraser Price replied that the chart is request count, which matches the subtitle “Requests by prompt & completion length.” Alex Atallah, OpenRouter’s CEO, replied with a heart. We did not pull a rankings export, and we did not call the Decisions API.

On September 25 at 22:21 UTC, OpenRouter posted typesafe/jev-router. The router uses Jev to pick a model and a reasoning effort before each turn. The thread says Jev reads the conversation text and scores difficulty, precision, whether a bigger model or more effort would help, whether a cheaper model is enough, and whether the task changed. Attachments are not sent to Jev. The call is under zero data retention. The router keeps a working model for the rest of the session and can change effort without switching. It switches when the expected gain is larger than the cost, including the cached chat it would lose. If the Jev call times out or the output is invalid, the request fails instead of falling back to another router.

The same thread says that on four agent benchmarks Jev Router solved 237 of 423 tasks and Auto Router solved 130 of 423. The post calls that 82% more tasks. A later post says that on five agent benchmarks Jev Router had a faster median time to first token than every other router tested. It does not print those times. The model id for chat completions is typesafe/jev-router. The model page we opened documents that chat call. It does not reprint 237 or 130. TypeSafe quoted the announcement at 22:29 UTC.

On September 26 Theo posted that he spent $1,000 benchmarking Jev Router. He wrote that DeepSWE performance was roughly the same as GPT-6 Astra on low, that it cost slightly more, and that it took almost 5 times longer. The post includes a chart. We did not read a task-level table off that image, and we did not call the router.

On September 28 Christoph Görn logged sixteen of his own prompts through typesafe/jev-router on OpenRouter. The write-up is on görn.name. The answered_by field names openai/gpt-6-luna five times, deepseek/deepseek-v4.1-flash three times, and moonshotai/kimi-k3 once. Seven of the sixteen calls errored, which he reads as replies that did not contain a JSON object. Nine successful calls averaged 2.99 seconds, from 1.26 to 4.41. He had set a 3 second timeout on each socket operation, not on the whole request. Sizing cost about $0.0032. The outcomes he counted: tiny 3, everyday 1, large 1, self 4, error 7. He writes that he expected TypeSafe’s Jev and that the log shows general-purpose models instead. We did not pull his log file. The same post’s local arm, Jared Palmer’s Kev-4B, is in the Kev story.

The model page opened October 8, and Arena’s October 7 session count, are on the Jev Router page.

OpenRouter posted at 00:16 UTC on October 15 that Jev 1.14 is now on OpenRouter. Opened later that UTC day, typesafe/jev-1.14 lists $0.042 per million input tokens, output free, and a 262K context. The TypeSafe row on that page prints P50 0.15 s and uptime 100.00%. The latest alias page already linked Jev 1.14. GET https://openrouter.ai/api/v1/models did not list typesafe/jev-1.14. GET of that id returned HTTP 404. Reopened at 11:04 UTC, the HTML page still returned HTTP 200 and GET of that id still returned HTTP 404. This desk did not call OpenRouter. TypeSafe’s own jev-latest, Date header 11:04 UTC, still returned jev-1.13.0. The desk numbers sit on jev-1.14.0.

Taka Sasaki posted at 04:26 UTC on October 19 that Jev 1.14 is on Cloudflare AI. Opened that UTC morning, typesafe/jev-1.14.0 returned HTTP 200. Context window 32,000 tokens. Input $0.042 per million tokens, cached input $0.00, output $0.00. The page lists Noul, Choice, Score, and Box. Image inputs and Box questions are marked beta. A sample noul on a black-rectangle PNG prints 0.95. Cloudflare’s older typesafe/jev card stays 32,000 tokens. OpenRouter’s typesafe/jev-1.14 listing stays 262K. We did not call Workers AI.