Reflection Beam vs DeepSeek V4 Pro: A Third the Size, Twice the Activation Ratio

The tempting way to frame Reflection Beam against DeepSeek V4 Pro is cheap newcomer versus established flagship. That framing is wrong in a way worth spelling out before anything else, because the two models are built on opposite readings of the same idea. DeepSeek V4 Pro is a 1.6-trillion-parameter mixture-of-experts model that activates 49 billion parameters per token — about 3.1 per cent of itself. Reflection Beam, announced on 5 October 2026, is a 501-billion-parameter mixture-of-experts model that activates 23 billion per token — about 4.6 per cent of itself. Beam is a third the size and, per token, a *less* sparse design. Reflection's whole thesis is inference efficiency, and the model it is being compared against here is ahead of it on the sparsity metric that thesis rests on.

That is not a debunking. It is the setup, because the two models sit at very different points on price and availability, and the honest comparison is about which one you can actually build on. DeepSeek V4 Pro is on the catalogue at $0.66 per million input tokens and $1.98 per million output, with a 1,048,576-token context and 384,000 tokens of maximum output. Beam has no published price at all, is reachable only through a waitlist, and its weights are promised for later this month under Apache 2.0.

Same idea, different ratio

Sparsity is the shared premise. Both models keep almost all of their parameters asleep at any given token and pay only for the small slice that runs. What differs is how far each lab pushes it and what it buys.

DeepSeek V4 Pro runs 49 billion active parameters out of roughly 1.6 trillion total. That is an extremely aggressive ratio, and it is what lets DeepSeek sell a flagship-tier model at $0.66 and $1.98 — a rate that would be unthinkable for a dense model of that capability. The trade is that total parameter count is enormous: the model is large to store and to serve at all, which is why hosting it is a specialist's business rather than something most teams do themselves.

Reflection Beam runs 23 billion active out of 501 billion. Smaller in every dimension: a third the total parameters, less than half the active ones, and a less aggressive activation ratio. Reflection's argument is not that Beam is the sparsest model available — it deliberately does not claim that — but that 23 billion active parameters is a size you can serve on a comparatively small number of GPUs, which changes the deployment economics rather than the per-token arithmetic.

• Total parameters — DeepSeek V4 Pro roughly 1.6 trillion against Reflection Beam 501 billion.

• Active per token — 49 billion against 23 billion.

• Activation ratio — about 3.1 per cent for DeepSeek V4 Pro, about 4.6 per cent for Reflection Beam. Beam is the less sparse of the two.

• Modal input — DeepSeek V4 Pro is text-only; Reflection Beam is text-only. Neither takes images, audio or video, which is a rare case of two models in this class agreeing.

• Context and output — 1,048,576 tokens in and 384,000 out for DeepSeek V4 Pro, against 256,000 in and 128,000 out for Beam, whose context figure carries a beta footnote.

• Price — $0.66 and $1.98 per million tokens for DeepSeek V4 Pro, with cache reads at $0.022, against nothing published for Beam.

• Availability — DeepSeek V4 Pro has been on general release since 24 April 2026; Beam is a waitlisted beta announced on 5 October 2026 with weights promised this month.

The price, including the part nobody mentions

DeepSeek V4 Pro's $0.66 and $1.98 are the provider's list rate, passed through on our catalogue with nothing added. There is a wrinkle worth budgeting for, though, and it is easy to miss: the rate card carries two timed windows, 01:00 to 04:00 UTC and 06:00 to 10:00 UTC, during which the price is multiplied by two — $1.32 in and $3.96 out. Both fall inside the working day for European and Asian teams, so a pipeline running on European hours will spend part of its day at double rate without anything failing.

Cache reads at $0.022 per million tokens are the other half of the picture and the larger one. That is a thirty-to-one discount on the input rate, and for the long-document and long-conversation workloads DeepSeek V4 Pro is built for — 1,048,576 tokens of context, 384,000 tokens of output — the effective cost of a session that re-reads a stable prefix is far below the headline. Any comparison that multiplies volume by $0.66 without modelling cache reuse will badly overstate what this model costs to run.

Beam has no equivalent to any of that. No list rate, no cache tier, no timed window, no published commitment about whether it will be sold per token at all. The comparison is not $0.66 against something cheaper; it is a priced model against an unpriced one, and until Reflection publishes a rate card there is no break-even to compute.

Benchmarks: the two models have never been run head to head

This is the part where care matters, because there is an obvious trap available and it is worth naming explicitly.

Reflection's launch tables compare Beam against Inkling, Nemotron 3 Ultra, GLM 5.2, GLM 5.3, Kimi K3, Qwen 3.8-Max and DeepSeek V4.1 Flash — the sibling, not V4 Pro. So the DeepSeek model that appears in Beam's own tables is a different one: the cheap, fast member of the family, at $0.15 and $0.60, which posts a vendor-table Terminal-Bench v2.1 of 90.6 against Beam's 80.1 and a DeepSWE v1.1 of 74.2 against Beam's 44.4. Reading a V4 Pro figure off a V4.1 Flash column is the single most likely mistake anyone writing this comparison will make.

What does exist for V4 Pro, from Artificial Analysis via our catalogue, sits on different harnesses again — so these are directional, not controlled.

• tau-squared Bench — DeepSeek V4 Pro records 96.2 against a vendor figure of 78.7 for Beam on MCP Atlas and 38.0 on tau3 banking. Different evaluations in the same family of agentic-tool-use measurement, and the two sets are not interchangeable.

• Artificial Analysis Intelligence Index — DeepSeek V4 Pro sits at 36.0, ranked 41st of 147 in its run; Beam has no index row at all, with the model pages returning 404 and no Beam entry in the leaderboard payload as of this morning.

• GPQA Diamond — DeepSeek V4 Pro at 92.8 per Artificial Analysis against Beam's vendor-reported 90.5.

• Humanity's Last Exam — DeepSeek V4 Pro at 41.0 per Artificial Analysis against Beam's vendor-reported 36.2, the latter explicitly without tools.

• SciCode — DeepSeek V4 Pro at 51.0 per Artificial Analysis against Beam's vendor-reported 49.7.

• Terminal-Bench v2.1 — DeepSeek V4 Pro at 78.7 per Artificial Analysis against Beam's vendor-reported 80.1. The only row in this list where Beam is ahead, and the margin is under two points.

• Long-context recall — DeepSeek V4 Pro at 80.3 per Artificial Analysis against Beam's vendor-reported 79.3 on AA-LCR.

The pattern is consistent and it should be uncomfortable for Beam's positioning. On every row where both models have a number, DeepSeek V4 Pro is ahead except Terminal-Bench v2.1, where it is behind by under two points — and every DeepSeek figure is from a third-party harness while every Beam figure is the vendor's own. A model already on general release, at a published price under a dollar in, is not obviously losing to an unreleased one on the benchmarks they share.

What Beam's efficiency claim has to do with any of this

Reflection's actual claim is not a benchmark claim, and repeating the tables as if it were misses the point. The claim is that Beam scores comparably to GLM 5.2 on advanced reasoning while using 3 to 4× less inference compute, with the gap widening against 2-trillion-parameter models like Qwen 3.8-Max. The supporting chart estimates generated FLOPs as twice the active parameter count times mean generated tokens per attempt, across DeepSWE, HLE and Terminal-Bench 2.1, and its own caption concedes that prompt prefill, context-dependent attention and serving overhead are excluded.

Set that against V4 Pro's numbers and the argument gets more interesting than the headline tables suggest. V4 Pro activates 49 billion parameters per token to Beam's 23 — roughly twice as many — so on the formula Reflection is using, Beam should be ahead by a factor of about two on generation compute per token before any difference in token count. That is a real and legible advantage, and it is the one Beam has.

Against it: V4 Pro's activation ratio is more aggressive, its context is four times larger, and its maximum output is three times larger, which means a V4 Pro session that solves a task in fewer, longer turns can end up cheaper in total even at a higher per-token estimate. FLOPs per token is not FLOPs per task, and the chart is per token.

The independent measurement does not exist yet. The only third-party statement on the record is Artificial Analysis saying on 5 October that Reflection had given it access and that it was benchmarking Beam, with early indicators suggesting Beam will be one of the most token-efficient open models it has seen for its level of intelligence. No number has followed.

Where V4 Pro's context actually matters

There is one dimension in this matchup where the gap is large enough to be decisive rather than incremental, and it is not a benchmark.

DeepSeek V4 Pro serves 1,048,576 tokens of context and 384,000 tokens of maximum output. Beam serves 256,000 and 128,000, with a footnote on the context figure warning it may move during the beta. For a legal contract review, a whole-repository refactor or a synthesis task over thousands of pages, that difference is not a performance delta — it is the difference between one call and a chunking pipeline you have to write and validate. Beam's midtraining stage reportedly extends its effective context to 1 million tokens, which is a training-side capability, while the serving limit stays at 256,000; the two figures are frequently conflated and they are not the same thing.

The output ceiling compounds it. A 384,000-token maximum output lets a single call produce a large structured artifact; 128,000 does not, and neither model offers a batch endpoint or an embeddings route through the surfaces described at launch.

Choosing, and where routing fits

For the specific workloads where a million tokens of context and a 384,000-token output matter, DeepSeek V4 Pro is not competing with Beam — it is competing with nothing else on this page, and Beam cannot enter the comparison at any price until its window changes.

For high-volume work that fits inside 256,000 tokens, the decision reduces to a bet on whether Beam's claimed efficiency survives a third-party harness, priced against a V4 Pro rate that is already an order of magnitude below the frontier and further discounted by cache reads at $0.022. Against that, DeepSeek's own cheaper sibling is sitting in the same catalogue at $0.15 and $0.60 — and it is the model Beam's own launch tables put Beam behind on DeepSWE and Terminal-Bench. The floor for a new entrant here is lower than it looks.

This is where a routing layer earns its place rather than being a slogan. DeepSeek V4 Pro is on our catalogue, with the provider's list rate passed through and nothing added, behind one OpenAI-compatible key at api.orcarouter.ai/v1 alongside 200-plus other models. The routing DSL lets you express the decision instead of hard-coding it: pin long-context requests to the model with the million-token window, pin high-volume short requests to whichever cheap model is passing your eval, and compose the two into a single call where the task needs both. Because the routing handles failover across providers, the cheap tier is safe to put on a share of traffic rather than all of it — which is exactly the posture you want towards a model like Beam, whose benchmarks nobody outside Reflection has reproduced.

The verdict, and the two dates that would move it

Reflection Beam is losing this comparison on everything that is knowable today. It is smaller, less sparse, has a quarter of the context, has no independent score, and — decisively for a model sold on cost — has no price. DeepSeek V4 Pro posts a better Artificial Analysis index, a better HLE, a better GPQA Diamond, a better SciCode and a better long-context recall figure, all measured by somebody other than its vendor, and all while activating twice as many parameters per token at a published rate under a dollar.

What Beam has is a single legible claim: 23 billion active parameters is a third of V4 Pro's, so a generation-side compute comparison should favour it by roughly two to one. That claim is plausible, it is the reason to take Beam seriously, and it has not been measured by anyone outside the lab that made it.

Two dates settle it. The weights, promised under Apache 2.0 later this month, which would let anyone count the FLOPs themselves. And the Artificial Analysis row, which would replace a vendor caption with a number. Until one of them lands, the model you can price, cache and call at a million tokens of context is the one to build on.