Jev Alternatives: Clef, Perplexity and Strands Decider

§ 01What a decision model does#

A decision model takes a state (a ticket, a document, an agent’s trace) and a set of questions declared in advance as choices, scores or yes-or-no, and returns a probability for every allowed answer. AWS’s post puts the trade plainly: decision models “are faster and more capable at a given size, always produce an answer from the selected options, and can run with very low latency.” The cost is just as plain: one parallel pass “makes it significantly worse at solving complex problems than reasoning models”, and none of them write text.

§ 02The five side by side#

All four new models sit on Qwen bases, per their model cards and posts: Clef and Perplexity’s model on Qwen3.8-27B, Clef-flash on Qwen3.5-9B, Strands Decider on Qwen3.5-2B. Context windows differ too: Jev takes 64k tokens per request, Clef 65,536, and Perplexity’s request limit is under 262,144.

§ 03Cloudflare’s benchmarks against Jev#

Cloudflare published the only head-to-head table so far, run on its own evaluation set. It calls the Clef models “smarter, faster, and fully Jev-API compatible”, which means existing Jev code can point at them.

Jev still wins two rows in Cloudflare’s own table, When2Call and BRIGHT, and Clef-flash drops sharply on CLINC150. The latency gap is partly hosting: Clef runs on Cloudflare’s own network.

§ 04Perplexity and AWS#

Perplexity’s pitch is price and images. At $0.04 per million input tokens it undercuts Jev slightly, and its state can carry images; the docs count about 1,000 input tokens per megapixel. Perplexity’s X post cites an 85.71% score across benchmarks; its docs publish no benchmark table, so that number is unverified here.

AWS went the other way: small, local and fully open. Strands Decider 2B answers in a median 115 ms on an RTX 3090, per its README, and scores 167 of 231 (0.723) on the public JevBench set. AWS published the training data and every iteration it tried, which makes it the one to read if you want to build your own.

§ 05What is verified and what is not#

• Prices, sizes and licenses are read from each vendor’s docs and model cards on October 1, 2026.
• The benchmark table is Cloudflare’s. It chose the tasks and ran every model; no independent comparison of all five exists yet.
• Perplexity’s 85.71% appears only in its X post.
• Jev’s architecture is still undisclosed; the three new models published theirs.

§ 06Why it matters#

Jev proved the demand; October 1 removed its moat. A buyer now has a cheaper hosted option, two open models with video input, and a 2B model that runs on a laptop, all speaking roughly the same API. For anyone running agents, the decision layer (route this, retry that, stop here) just became a commodity. The value moves to the work the agent does after it decides. For how agent harnesses compare, see our harness ranking.

§ 07What we are watching for#

• An independent benchmark across all five on the same tasks.
• TypeSafe’s response: a price cut, open weights, or image input.
• Whether Vercel’s AI Gateway adds the new models next to Jev.

§ 08Update log#

• October 1, 2026: page opened, from the three launches and TypeSafe’s docs.

§ 09Sources#

• Cloudflare, Introducing Clef, October 1, 2026; Clef and Clef-flash docs.
• Perplexity, Decisions API quickstart; pplx-decider-v1-27b on Hugging Face.
• Strands Agents, Introducing Strands Decider 2B, October 1, 2026; GitHub repository.
• TypeSafe AI, model docs.