Cloudflare's Jev Rival Got Cheaper, Faster and Multimodal This Week. On TypeSafe's Own Tests, It Still Loses.

Cloudflare picked a fight with TypeSafe AI's Jev on October 1, open-weighting a pair of models called Clef and Clef-flash built to do one narrow job: answer bounded questions -- yes/no, multiple choice, a ranking -- without generating a word of text. Eight days later, on October 9, it escalated. A new multimodal Clef-omni now reads text, images, audio and video in a single call; Clef-flash's price dropped below Jev's; and Clef itself got roughly twice as fast to serve, with no retraining involved. What Cloudflare's own announcement doesn't say is that its newest, most capable model scores worse than the one it's replacing on the exact business tasks TypeSafe built Jev to handle. That reversal matters because decision models are quietly becoming the part of an agent stack trusted with consequential calls -- approve the refund, escalate the ticket, flag the transaction -- precisely because they don't generate free text that can hallucinate a wrong answer in prose. A buyer picking between the two now has to weigh a cheaper, faster, broader model against one that, on the seller's own numbers, got worse at the job.

TypeSafe AI created this category. Jev, which launched in mid-September, never generates text at all -- it takes a typed schema of allowed answers and returns calibrated probabilities for each one, useful for the parts of an agent's workflow a chat model handles unreliably: routing a support ticket, approving a transaction, flagging a security alert as real or noise. TypeSafe's $870 million Series A, reported the same day as Cloudflare's second move, priced that approach at a $7.5 billion valuation. Cloudflare's answer isn't just structurally identical -- it's a declared drop-in: Clef's API matches Jev's, so switching is a model-ID change, nothing else. Both companies are betting on the same idea: that a large share of what gets marketed as 'AI agent reasoning' is actually a handful of bounded decisions repeated thousands of times a day, and that running those decisions through a model built to score typed answers is cheaper and more auditable than routing them through a general chat model and hoping its free-text output parses cleanly.

That last row is the real story. In the October 1 post, the original Clef -- a text-only, 27-billion-parameter model -- beat Jev on three of TypeSafe's own four workflow evaluations: invoice processing, customer-service exact actions, and security-incident classification. It lost only on "agent trace observability." Clef-omni, the larger, multimodal model built on a 30-billion-parameter backbone that Cloudflare is now promoting as the flagship, reverses that scorecard against the identical Jev baseline figures: it loses on invoice processing and customer service, ties on security incidents, and still loses on agent trace observability -- winning none of the four outright. Clef-omni does beat Jev cleanly on four public benchmarks Cloudflare also cites -- BFCL, API-Bank, a home-appliances dataset, and BANKING77 -- but those aren't the workflows TypeSafe uses to sell Jev to customers.

The speed and price changes are mostly infrastructure, not new weights. Cloudflare says the roughly 2x median speedup on hosted Clef -- 616 milliseconds down to 305 at a 3,400-token input, 262 down to 152 at 800 tokens -- is entirely a serving change: a patch to the SGLang inference engine (pull request #42721) due to land in SGLang 0.5.22, after which anyone self-hosting Clef's open weights gets the same gain for free. The price cut came with a quieter trade-off: Clef-flash's hosted context window shrank from 64,000 tokens to 24,000, because Cloudflare says only 0.24% of real requests ever used more than that -- Jev's own context window, by contrast, caps the combined state and question at 32,000 tokens even though it accepts up to 64,000 total. Clef-omni launched at $0.15 per million tokens, between Clef-flash's new price and Clef's; across Cloudflare's own 43-benchmark suite, median hosted latency lands at 38.8 milliseconds for Clef-flash, 209.3 for Clef, and 524.1 for Jev. (Self-hosting has its own floor. Cloudflare says Clef-flash needs a GPU with at least 41GB of VRAM and the full Clef needs 85GB, both at a 64k context window and single concurrency -- a meaningfully bigger ask than renting the hosted endpoint.) None of that changes who owns the only hosted endpoint, either -- Cloudflare still runs the sole hosted version of either Clef variant, the same way TypeSafe runs Jev's.

Clef has the ability to disrupt the way we use agents.

That is Cloudflare's own description of the stakes, made about a market TypeSafe is roughly a month into building and just raised $870 million to defend. Price and full API compatibility are real competitive pressure regardless of whose benchmark numbers hold up -- a developer can swap endpoints today and feel nothing break. Whether Clef-omni's business-task scores improve once someone outside Cloudflare runs the same four tests is the actual open question, and right now the only numbers anyone has are the ones the company publishing them chose to publish. That isn't unique to Cloudflare -- TypeSafe's own Jev Decision Index scores for Jev are just as self-reported -- but it does mean the comparison table above is the most anyone outside either company can currently build, and it should be read as exactly that: two vendors grading their own homework against each other, with the one independent voice in this story, The Register, flagging that nobody has checked it yet.