Clef
Released Oct 1, 2026 · knowledge cutoff unpublished
- Status
- Unknown
- Location
- United States
- Type
- System One
- Modality
- Multimodal
- Context window
- 66K
- Max output
- —
- Speed
- —
- Price ($/MTok in / out)
- $0.24 / —
- Cost per task
- —
- Open weights
- Yes
Strengths & weaknesses
Strengths
- Speaks the same System One API as Jev, so an existing Jev integration moves over by changing the endpoint and model name
- Reads images and video as well as text and JSON — up to four images go in ahead of the state, so a screenshot or photo can be part of the decision
- Scores every allowed option in one non-autoregressive pass with no intermediate text, so there is nothing to parse and no reasoning tokens to wait for
- Open weights plus a hosted Workers AI endpoint, and Cloudflare launched an RL fine-tuning platform for decision models alongside it
Weaknesses
- Weaker than Jev on knowledge-heavy multiple choice — Cloudflare's own table has Jev ahead on GPQA Diamond, MMLU-Pro, BBH and When2Call, so its wins are in routing, classification and grounding tasks
- A user on Cloudflare's community forum reports hosted calls appearing to stop counting state at about 2,048 tokens despite the advertised 65,536-token window — test long inputs before relying on the full context
- Calibration is Cloudflare's claim, not yet independently audited — confidence has to be measured on your own traffic before setting an automation threshold
Not an LLM: a "System One" decision model, Cloudflare's first trained in-house by the Workers AI team, post-trained from Qwen3.8-27B with a joint schema head that returns a probability for every allowed option of typed noul, choice and score questions in one forward pass, so none of this site's benchmarks apply and every cell is null by design (Cloudflare's 48.0 GPQA Diamond is a decision-format run, not comparable). Filed under the fast tier alongside Jev and Laya. Workers AI lists $0.24 per M input tokens (21,818 neurons) and publishes no output rate; press reports say output is unbilled, but no Cloudflare page seen states that, so outputPerMTok is null. Images are capped at 4 per request, 4 MiB and 16 MP each. Cloudflare reports a 209.3 ms median latency against Jev's 524.1 ms and says the pair lead 7 of 10 benchmarks in Jev's Decision Index 0.2.1 suite; all of these are its own figures. Calibration training pairs label-smoothed cross-entropy with a Brier loss. Released with Clef-flash on 2026-10-01 (blog.cloudflare.com/clef-decision-models). No Artificial Analysis coverage; knowledge cutoff unpublished.
Benchmarks0 of 10 reported
MMLU-Proretired
GPQA Diamond
SWE-bench Verified
SWE-bench Pro
Terminal-Bench 2.1
AIME (latest)retired
Humanity's Last Exam
LMArena Elo
GDPval-AA v2
ARC-AGI-2
For developers
API model strings
- cloudflare
@cf/cloudflare/clef
Licence
Apache License 2.0 · Permissive · commercial use permitted
Retirement
No retirement announced
Lineage
No recorded predecessor