Beam 1.0 vs Clef Flash: One Writes Code, One Cannot Write a Sentence

Put Beam 1.0 and Clef Flash side by side and the comparison looks absurd until you notice what it is actually about. Beam 1.0 is Reflection's 501.3-billion-parameter mixture-of-experts model, released with Apache 2.0 weights on 16 October 2026 and built for coding, reasoning and long-horizon agentic work; it reads text and writes text. Cloudflare's Clef Flash is a 9-billion-parameter model that has no text output path at all — not a degraded one, none. It takes a state plus a schema of typed questions and returns a probability for every allowed answer. Any page that ranks these two against each other on benchmark scores has misunderstood both. The useful question is which slot in a pipeline each one occupies, because in a well-built system they are not rivals; they are adjacent stages.

That adjacency is worth understanding precisely, because it explains a class of production failure that no single-model benchmark surfaces: the parser you bolted onto a chat model because it was the only thing you had, and the fifty-first answer out of a hundred that did not fit your schema.

What each model actually returns

Start with output shape, because everything else follows from it.

• Beam 1.0 — free-form text, with mandatory reasoning. Supported efforts are low, medium, high, xhigh and max, and max is the default, so the out-of-the-box configuration thinks the longest. Reasoning, tool calling and structured outputs are all supported, and structured outputs plus a JSON schema is the closest Beam gets to a constrained return.

• Clef Flash — a distribution over a declared answer set. Cloudflare's documentation describes a forward pass that scores each permitted option; there is no generation step and therefore no parse step. The vendor's own framing is that this is a decision model, and the second product in the release was a reinforcement-learning fine-tuning platform for building more of them.

The practical difference is what happens when the input is out of scope. A text model asked something outside its competence will produce confident prose, and your system has to notice. Clef Flash will produce probabilities that sum to one across the choices you gave it, which means an out-of-scope input shows up as a flat distribution rather than a plausible sentence — provided you built a threshold for that, which the vendor does not supply. Cloudflare's own published table shows that case as the model's weakest cell.

The evaluations do not overlap, and that is the finding

Clef Flash's scores come from the Jev Decision Index and a set of decision-specific suites, all run by Cloudflare, and several are run against a rival's own benchmark — Jev is published by TypeSafe, the vendor of the model Clef Flash is measured against. On those suites the numbers are strong: BFCL case-exact 98.8, API-Bank 93.1, a home-appliance simulation at 97.7, ContractNLI macro-F1 84.3, RouterBench 79.9. Clef Flash edges the larger 27B Clef on BFCL, on API-Bank and on the appliance suite, the last by fourteen points. None of it has been independently reproduced.

Beam's evaluation regime is different in kind. Reflection's model card reports agentic and software-engineering benchmarks — SWEBench Verified 81.6, Terminal Bench v2.1 80.1, DeepSWE v1.1 43.3, GPQA Diamond 90.6 on the BF16 checkpoint — all vendor-reported. The one outside number is Artificial Analysis, which lists Beam 1.0 at an Intelligence Index of 30, nineteenth of 120 models on index version v4.3.2, and records no price and no speed because it still sees "No publicly available API" for the model. There is nothing in Clef Flash's column that corresponds to an Intelligence Index, and nothing in Beam's column that corresponds to a case-exact router score. They are not on the same axis.

What you can compare is the cost structure, and here the two are at opposite extremes. Cloudflare hosts Clef Flash on Workers AI at $0.038 per million input tokens with output tokens billed at zero, because there are none — roughly a sixth of the $0.24 the same rate card charges for the larger 27B Clef, and just under the $0.042 we charge for a decision model that answers the same request shape. Beam 1.0 has no published per-million-token price at all — Reflection's documentation carries a full rate-limit page and no rate card. Self-hosting is the only priced path, and it starts at four N​VIDIA B300 GPUs.

Where each one belongs in a real system

The honest answer is that Clef Flash replaces a step you are probably doing badly, and Beam 1.0 replaces a step you are probably doing expensively.

Clef Flash's target is classification, routing, gating and triage inside an agent loop — anywhere the correct output is one of a known set of labels and the cost of a malformed answer is a downstream exception. Its published latency profile, roughly 38.8 ms median and 122.4 ms at p95, is the part that matters there, because it is measured in a regime where a text model's time-to-first-token is not.

Beam 1.0's target is the work itself: resolving GitHub issues, operating in a CLI, running multi-step tool workflows. The argument Reflection makes is token efficiency rather than peak capability — the model card claims comparable solution quality to leading open models at roughly three times fewer inference-time reasoning tokens, and the strongest support for that is Artificial Analysis recording 79M output tokens for Beam across the Intelligence Index against a 130M median, tenth of forty-six on that measure. On a long agent run where every turn replays the prior conversation, that is the line item that compounds.

Neither is on our catalogue, and we are not going to imply otherwise. Clef Flash runs on Cloudflare's own edge GPUs and the Apache-2.0 weights are yours to self-host; Beam 1.0's weights are on Hugging Face under Apache 2.0 with a first-party API that has no published price.

The reason to keep both reachable

A pipeline that uses a decision model at the gate and a reasoning model behind it is a strictly better architecture than one that uses a single chat model for both, and the cost of building it that way is the operational cost of a second dependency — a second contract, a second set of keys, a second failure mode. That is the part a router removes rather than the model part.

If the gate steps in your system are the ones Clef Flash describes, the nearest already-routed equivalent is TypeSafe's Jev 1.13 — the model Cloudflare chose as its benchmark counterpart, API-compatible with Clef Flash's request shape, and on our catalogue at $0.042 per million input tokens with output billed at zero and a 65,536-token context.

For the reasoning stage behind the gate, our catalogue carries the models Beam is sized against: GLM-5.2 at $1.40 and $4.40 per million tokens with a 1,000,000-token window, Kimi K3 at $3.00 and $15.00 with 1,048,576, and Qwen3.8-Max at $2.00 and $6.00. One OpenAI-compatible key covers all of them, provider list price passes through with nothing added on top, and automatic failover means the gate stage does not go dark because one provider had a bad minute.

You can also keep the two stages on different models without committing to either: a routing rule that sends schema-classification traffic to the decision endpoint and free-text work to a reasoning model is the routing DSL's ordinary case, and it means the decision stage's accuracy and the reasoning stage's cost can be tuned independently instead of trading off inside one model's parameter count.

Which one you are actually shopping for

If you arrived here because your agent occasionally emits a JSON blob your parser rejects, you do not want a 501-billion-parameter model. You want the decision model, and the question to ask is whether your answer set is genuinely enumerable and what your abstain policy is — because Clef Flash gives you probabilities, not a refusal.

If you arrived here because your coding agent's bill is dominated by reasoning tokens you cannot see, you do not want a 9-billion-parameter classifier. You want Beam 1.0, or one of the models it is measured against, and the question to ask is whether you can host it or whether you are waiting on a rate card that has not been published.

The one thing a head-to-head framing gets right is that both are bets on the same thesis from opposite ends: that the expensive part of an agent is not the answer, it is everything around it. Cloudflare's answer is to remove generation from the steps that never needed it. Reflection's answer is to make generation cheaper per token of thinking. Deployed together they cost less than either deployed alone, which is a better outcome than either vendor's benchmark table will tell you.

None of that is hypothetical: the documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.