06-reference/research

cloudflare clef clairic vs typesafe jev

2026-10-24·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)
competitordecision-modelscloudflaretypesafepricing

Cloudflare commoditized Jev's licence and interface, not its price or its quality: Clef leads Jev on public benchmarks and falls 5.3 points behind on held-out private tests

The question

"How do Cloudflare's Clef (classification) and Clairic (regression/range) compare to TypeSafe's Jev on pricing, latency, and positioning?"

Asked because Ray Data Co (RDCO) is actively adopting Jev for typed-set judgments, so a hyperscaler moving onto the same primitive is a live dependency question rather than a curiosity. Clef is the classification model and the actual Jev rival; Clairic is regression and extraction. The founder corrected that distinction on 2026-10-19.

Source-quality rules used in this brief

The three axes do not have equal evidence quality, so four labels appear throughout:

Two measurement traps caught this brief during review and are worth stating up front, because the press has fallen into both. First, the Jev Decision Index publishes two different scores for every model, a public-benchmark score and a private held-out score, and they rank Clef and Jev in opposite orders. Quoting either alone misleads. Second, that index's # column is a tie-band, not a rank: its own tooltip reads "tied: within 0.9 points of the first model in its group." Jev's band label of =10 means 41st of 229, not tenth.

What we already know (from the vault)

What the web says

Head-to-head: pricing

All of these bill per million input tokens with output free, which is unusually lucky for comparison. The hazards are elsewhere.

Jev (TypeSafe) Clef Clef-flash Clef-omni Clairic OpenAI Decisions
$ / M input tokens $0.042 (vendor) $0.240 (vendor) $0.038 (vendor) $0.150 (vendor) $0.100 (vendor) $0.100 (vendor)
Ratio to Jev (derived) 1.00x 5.71x 0.90x 3.57x 2.38x 2.38x
Output tokens free, "too cheap to meter" free free free free, "generates no tokens" free; no cache charges
Native billing unit dollars per token 21,818 neurons/M 3,455 neurons/M 13,636 neurons/M 9,091 neurons/M dollars per token
Free allowance none published (U) 10,000 neurons/day, including on the free Workers plan same same same (U)
Weights closed, hosted API only Apache 2.0, Hugging Face Apache 2.0 Apache 2.0 not released (U) closed
Self-host floor n/a 85 GB video memory (vendor) 41 GB (vendor) (U) n/a n/a
Maturity "early access"; no general-availability language no maturity label at all no label no label explicitly "in preview" generally available 2026-10-22

Sources: typesafe.ai homepage and docs.typesafe.ai/models.md; developers.cloudflare.com/workers-ai/platform/pricing/ (states last updated 2026-10-20); Cloudflare's rate of $0.011 per 1,000 Neurons; OpenAI's own community announcements of public beta (2026-10-06) and general availability (2026-10-22). The video-memory floors are Cloudflare product manager Michelle Chen paraphrased by The Register, qualified "assuming single concurrency and a 64k context window."

What had to be assumed to normalize, stated plainly:

  1. Cloudflare's dollar figures are its own neuron conversion. It bills in "neurons" and publishes both; I used its published dollars rather than recomputing.
  2. The free-allowance-to-token conversion is my arithmetic, not Cloudflare's (derived). At 10,000 neurons/day: roughly 458K Clef input tokens/day, 2.9M Clef-flash, 733K Clef-omni, 1.1M Clairic.
  3. Clairic bills schema tokens as input, and they dominate. Cloudflare's own worked example reads state_tokens: 61, schema_tokens: 472, input_tokens: 655. Schema is 72% of billed input, and a further 122 tokens are neither state nor schema and go unexplained. TypeSafe publishes no equivalent breakdown (U). Per-token cost is comparable across vendors; per-decision cost is not.
  4. Nobody sells "a million judgments" (U). No per-request or per-judgment stock-keeping unit exists at any of these vendors.
  5. Open weights are not free to run. Clef's 85 GB video-memory floor puts self-hosting above a single consumer or prosumer graphics card, so Apache 2.0 removes licence lock-in without removing cost.

The findings that matter more than the rates:

TypeSafe's own hedge is worth keeping: its docs say output tokens are "currently free of charge," and the launch post concedes the company "can't prove it isn't subsidized."

Head-to-head: latency

Nobody has measured these models under one controlled setup, and both independent sources say so themselves. Percentile shorthand below: p50 is the median, p95 the 95th percentile.

Source Who Clef Clef-flash Jev Rig
Cloudflare's 43-eval suite vendor-measured 209.3ms median / 238.6 p95 38.8 / 122.4 524.1 / 536.0 not stated (U)
Jev Decision Index 0.4.1 independent 102ms 53ms 196ms (1.14); 524ms (1.13) local models on 1x RTX PRO 6000; Jev over HTTP
Arize AI independent 0.50s p50 not tested 0.08s p50 (1.13) both hosted and called over the network

TypeSafe's own figure, for completeness: "End-to-end response time is 70ms-500ms" (vendor) - a range, with no percentile and no average, self-measured and "generally run from our laptops on the West Coast."

Five cautions, each load-bearing:

  1. The two independent sources tested different Jev versions, so their numbers are not interchangeable. The Decision Index runs jev-1.14.0; Arize ran Jev 1.13. The Index's own 1.13 figure is 524ms against 1.14's 196ms, so version choice moves the answer by a factor of nearly three.
  2. Arize is the one source where both models were hosted endpoints called over the network, and there Jev is about 6x faster (0.08s against 0.50s p50 on RAGTruth). Arize's laptop caveat - "Local latencies are from one laptop and aren't comparable with the cloud numbers" - applies only to Kev 9B, Jeeves 9B, Strands Decider 2B and Laya. Clef 27B ran on Workers AI in what Arize calls the cloud conference. Arize does not itself claim this is uniquely apples-to-apples, so that reading is mine.
  3. The Decision Index runs Clef on a datacenter graphics card and Jev over the internet, and says so: "Jev's latency is a hosted HTTP round trip and is not a controlled comparison with local GPU runs." Its 102ms-versus-196ms ordering therefore measures hosting as much as modelling.
  4. Cloudflare's own two Clef tables do not reconcile. The 2026-10-09 post gives a pre-optimization Clef median of 262ms / 438ms p95; the 2026-10-01 launch post claimed 209.3 / 238.6 for the same unoptimized model. Cloudflare never addresses it. Neither table is the Clef latency.
  5. Clairic's headline speed claim is checkable and partly fails. Cloudflare's three-step insurance-claim demo reports 19.3s cascaded against 61.7s for an all-LLM pipeline, "about 3.2 times faster," "49% less" cost. The cascade arm sums correctly (0.46 + 0.22 + 18.6 = 19.28s). The baseline does not: 16.1 + 21.9 + 16.4 = 54.4s against the 61.7s stated, a 7.3s gap - 11.8% of the stated total - that Cloudflare never accounts for. On the summed baseline the speedup is 2.8x rather than 3.2x. Further: the baseline is GLM-5.3 rather than a frontier model; the 49% saving is driven by output tokens in the reply step (1,329 against 2,544), not by Clairic's input price; and 49% is cost while 3.2x is time, so the widely-copied "49% cheaper / 3.2x faster" pairing conflates two measurements.

Net read: there is no defensible claim that either model is faster than the other. Each independent source favours a different one, and in each case the rig explains the result. On the only hosted-to-hosted comparison available, Jev wins. TypeSafe's self-measured 70-500ms band is consistent with both independent figures for it (196ms and 80ms), so the vault's prior shorthand of "sub-half-second" was reasonable and needs no correction.

The accuracy result that decides the question

The question named three axes, but whether Cloudflare is commoditizing the primitive turns on a fourth, so it gets its own section. The headline finding is a split by methodology, and it is more interesting than either side's marketing.

Jev Decision Index 0.4.1 (independent), updated 2026-10-22. Run by one person (multimodalart on Hugging Face), footer reads "Not affiliated with TypeSafe AI," frozen suite of 110,201 requests across 37 benchmarks and 229 entries, local models on one RTX PRO 6000. It publishes two scores:

Public-benchmark score Private "Full score" Calibration error Latency
Cloudflare clef (28B) 61.7 (above both Jev versions) 54.9 4.0 102ms
Jev 1.14 60.7 60.2 8.0 196ms
Jev 1.13 58.0 59.2 7.8 524ms
clef-omni (30B-A3B) 56.5 45.0 4.0 191ms
clef-flash (9.7B) 56.1 44.4 2.5 53ms

The Full score is "0.60 x same-skill private + 0.40 x new domains," and the index excludes public benchmarks from it for a stated reason: "The 37 open tasks shown on this site. Anyone can see them, and train on them, so they no longer count toward the Full score."

So Cloudflare's leadership claim is not refuted, it is located. Clef genuinely leads Jev on the public benchmarks Cloudflare cited, and that lead reproduces independently. It then reverses by 5.3 points on tests nobody can train against. That pattern is the signature of benchmark contamination or overfitting to the public suite, and it is exactly what the index was restructured to expose. It is also the single most decision-relevant fact in this brief, because RDCO's workload is private data, not a public benchmark.

The split also explains the apparent disagreement among independent sources without needing a second theory:

Two things the brief should not overclaim:

Clairic has no independent evaluation of any kind (U). It is on neither leaderboard, reasonably, since both score decision models. Every Clairic performance number in circulation is Cloudflare's, and secondary coverage repeats them, with PPC Land's own headline conceding the saving was measured "in its own test." Cloudflare's one checkable mini-eval is small and candid (vendor-measured): 21 days of public Transportation Security Administration checkpoint counts, 3.3% mean error, 17 of 21 actuals inside the 80% interval, with Cloudflare noting three of the four misses fell on Columbus Day weekend because the model "doesn't account for events the series doesn't show." That is n=21 on one series.

Head-to-head: positioning

Jev Clef Clairic
What the primitive is for (vendor) "unstructured state in, typed probabilistic decisions out"; a "frontier-intelligence function call" "reads an input state and a set of typed questions, then returns a probability for every allowed answer... no free-form output to parse and no reasoning tokens to wait for" "extract, compute, estimate, and forecast values from unstructured inputs"
Output shape Choice / Noul / Score; Noul returns probability-of-yes with no separate confidence probability per allowed answer; 1-64 questions per request typed spans, numbers, dates, durations, lists; p_null plus confidence; estimates return median and an 80% interval
Buyer "developers"; no stated ideal customer profile (U) the agent builder; the pitch is removing human-in-the-loop developers already making structured-output calls to frontier LLMs
Place in stack hosted API callable from anywhere; no edge, on-premises or private-cloud offer (U) inside the Worker, on the request path; Cloudflare's own header is "Built for the hot path" same; env.AI.run(), REST, or AI Gateway
Modality text; images gated behind a TypeSafe-Danger-Zone-Images header on 1.14 only vision encoder inherited from the Qwen backbone; clef-omni adds audio and video text plus up to 4 images
Context 64k per request / 32k per question 65,536 (Clef); 24,576 (flash) 65,536

Cloudflare's framing is frontal assault on one axis and category extension on the other, and it is explicit about both.

Cloudflare credits TypeSafe with creating the category, by name and repeatedly. The Clef post opens: "Jev introduces a new decision model concept into the world of AI." The Clairic post: "When System One models were announced by TypeSafe a month ago, it took the industry by storm... We were inspired by the concept of System One models." Cloudflare's own landscape table lists the Decide column as "Jev, Clef," placing itself inside TypeSafe's box rather than claiming a new one.

Then it does three things an ordinary fast-follower does not:

  1. It made the swap cheap. "These models are smarter, faster, and fully Jev-API compatible," and "Clef follows the System One API, so you can switch an existing Jev integration to Clef by changing the endpoint and model." Cloudflare adopted TypeSafe's interface and built the migration path on TypeSafe's behalf. In practice it is three changes rather than two, since a Cloudflare account and token are also needed, and Arize's successful hosted calls are third-party evidence the interface works.
  2. It benchmarks itself on TypeSafe's own scoreboard, claiming to beat Jev "in 3 of 4 areas" on "Typesafe's own workflow evals." As above, that holds on public tasks and not on private ones.
  3. It gave the weights away under Apache 2.0, against a closed hosted-only incumbent. With one asterisk: Cloudflare's product manager confirmed to The Register that the training data is not public, so "open source" here means open weights rather than an open model.

Clairic is not aimed at Jev. Cloudflare's carve is explicit: "Clairic handles the first [a number]... Clef handles the second [a choice from a fixed set]... LLMs handle the third." And: "Until now, every System One model has been a decision model: Jev and Clef... Clairic is the first model in a new class." It is a land-grab in territory Jev does not occupy, and the two chain: "Extraction models turn raw input into typed values, decision models turn values into decisions, and LLMs turn decisions into words." Clairic's weights are not released, so the commoditization logic that applies to Clef does not extend to it today.

TypeSafe has not responded publicly and has not cut price. Zero mentions of Cloudflare, Clef, or Clairic across its homepage, launch post, 1.14 post, enterprise page, and docs. Jev remains $0.042.

The category filled in about five weeks, which is the real context. Jev launched 2026-09-15; OpenAI announced Decisions at DevDay around 2026-09-29 (secondary-sourced, U) and reached general availability 2026-10-22; Amazon's Strands Decider 2B and Cloudflare's Clef both shipped 2026-10-01; Liquid AI's d1-ultra and Cloudflare's Clairic both landed 2026-10-19 (dates for Strands, d1-ultra and the DevDay announcement were not traced to primary sources this pass, U). Perplexity, Microsoft, Fastino and others sit on the boards.

Convergences and contradictions

Synthesis for RDCO

Cloudflare is commoditizing the primitive, but it has commoditized the licence and the interface rather than the price or the quality, and the distinction decides what RDCO should do. The licence is unambiguous: Apache 2.0 weights on Hugging Face against a closed hosted-only incumbent. The interface is close behind, by Cloudflare's claim and with third-party corroboration that an outsider can call Clef through a Jev-shaped API. Price is where the headline breaks down. The Clef that competes with Jev on accuracy costs 5.71x Jev at list, and Arize measured roughly that ratio in practice at $0.223 against $0.045 per thousand decisions. The model that does undercut Jev, Clef-flash at $0.038, scores 44.4 against Jev's 60.2 on held-out tests. That is a cheaper, weaker product rather than a commoditized price, and the only regime where Cloudflare is genuinely cheaper at equal quality is inside its free daily allowance.

Quality is where the research actually paid off, and the finding is not the one the press is carrying. Cloudflare claimed to lead TypeSafe's own leaderboard. That claim reproduces on the index's public-benchmark view, where Clef posts 61.7 against Jev's 60.7, and it reverses on the private held-out Full score, where Clef trails 54.9 to 60.2. The index excludes public tasks from its headline for a stated reason: anyone can see them and train on them. So the honest reading is that Clef's advantage lives on the benchmarks it could have been tuned against and disappears on the ones it could not. Arize's independent parity result, run on public benchmarks, fits that pattern rather than contradicting it. For a buyer whose workload is private data, the private column is the one that matters.

Cloudflare's strategic intent is plainly to make typed judgment a cheap platform feature, though the margin story is my inference rather than its statement. Its own words are narrow: "The interest in Jev shows the need for a fast, small, specific, classifier model, and we chose this to be our niche to start experimenting with RL environments." The reading that the money is meant to land in Workers, AI Gateway, and the reinforcement-learning fine-tuning service is inference, and it stays unconfirmed until that service has a published price (U). What is observable is that TypeSafe is structurally exposed if the inference is right: its enterprise page offers no platform, no fine-tuning, no eval tooling, no dedicated capacity, no on-premises option and no service-level agreement. The product is one model behind one endpoint. The larger threat, though, is not Cloudflare at all. It is that the Decision Index is a board of 228 open reproductions of Jev, 40 of which beat it, with an unfamiliar name at the top.

For RDCO the pricing axis is irrelevant and the dependency axis is the real one. RDCO's Jev volume is a handful of judgments per /ideas run, which sits inside Cloudflare's free daily allowance even on its most expensive classification model, so in absolute terms either choice is a rounding error and cost is not a reason to move. The reasons worth weighing are that Clef's weights are Apache 2.0, which removes single-vendor risk from a primitive the founder made a standing default; that Cloudflare is already RDCO's platform, so Clef is reachable from Workers RDCO already runs and through the AI Gateway rail already configured for Scribble Works; and that Jev is formally "early access" with console-gated rate limits and no service-level agreement. Against all of that sits the finding above: on held-out private tests, which is the regime RDCO operates in, Jev is the better model by five points.

The practical conclusion is that RDCO does not have to choose, and should not pre-commit. Because Cloudflare adopted TypeSafe's interface, jev_judge.py needs two constants changed plus a Cloudflare token, against a 301-line standard-library-only script. That is a flag, not a migration, and it is the test [[2026-09-27-alphasignal-system-one-models-jev-laya-clm8b]] called for and could not run. Until RDCO measures on its own labelled sample, Jev stays the default under [[feedback_jev_for_typed_set_problems]]: it is wired, its price is immaterial at this volume, and it leads on the only independent evidence that is contamination-resistant. Two side findings are worth carrying. First, for the harness work in [[2026-10-22-automode-classifier-channel-text-injection-surface]], Cloudflare argues that feeding a decision model typed values instead of raw text shrinks the prompt-injection surface, because "Clef decides from a five-line summary of typed values instead of a free-form email, which can also be susceptible to prompt injection." That reasoning applies to RDCO's own classifier-gate design whoever wins. Second, the independent accuracy leader is Fastino GLiDE 1.2 at 71.5 and 93ms, eleven points above Jev, and RDCO has never looked at it.

Why this is in the vault

[[2026-10-20-technically-whats-a-decision-model]] explicitly flagged that its source predated Clef and Clairic and could not assess them, and [[feedback_jev_for_typed_set_problems]] makes Jev a standing default for every RDCO skill with a fixed-set decision. This brief supplies the vendor-dependency check that default was missing, and gives /ideas and any future Jev-wired skill a costed decision: parameterize the provider, keep Jev as the default on the strength of the private-test result, and read any Clef-versus-Jev benchmark claim by asking first whether it was scored on public or held-out data.

Test plan (build task, not research)

Pre-registered so the result cannot be retrofitted:

  1. Add a --provider {jev,clef,clef-flash,glide} flag to jev_judge.py, changing only API_URL, MODEL and the credential lookup. Confirm Cloudflare's interface-compatibility claim survives contact with RDCO's actual payload, which uses state plus a questions map. That claim is currently vendor-stated and untested here.
  2. Build the labelled sample first: idea-type assignments and near-duplicate calls from the existing idea bank that the founder or Ray has already judged, with judgments withheld from every model. This sample is private by construction, which is the point.
  3. Score agreement against the held-out labels, plus wall-clock latency measured from RDCO's own Mac mini, since no vendor pair has been measured on one rig and the two independent sources disagree.
  4. Report calibration, not just accuracy: whether high-confidence answers are in fact more often right. The two independent sources contradict each other on calibration, so this is the measurement with the most to gain.
  5. Acceptance: Clef replaces Jev only if it wins on agreement at equal or better measured latency on RDCO's own sample. No published benchmark counts toward this, including the Decision Index.

Open follow-ups

Related

Sources

TypeSafe (vendor-primary)

Cloudflare (vendor-primary)

Independent measurement

Press and third-party

Claims carried without a primary source, flagged (U) in text The TypeSafe funding amount and valuation; the Microsoft-Decision-1 price; the OpenAI DevDay announcement date; the Amazon Strands Decider 2B and Liquid d1-ultra launch dates; the "commoditise in 16 days" framing.

Vault paths cited