Mercury 3 vs Reflection Beam: The Fastest Closed Model Meets the Fastest Open One

Reflection Beam arrived with a number that could be checked, and Mercury 3 arrived with a number that could not. Beam is scored at 335.1 output tokens per second by Artificial Analysis, with a time to first token of 0.80 seconds and an Intelligence Index of 30 — fourth fastest of 78 models measured, twentieth of 121 on quality. Mercury 3 claims 1,068 tokens per second and has no independent score of any kind, one day after Inception released it on October 19, 2026. So the fair way to open this comparison is not "which is faster" but "what happens to the gap when the unmeasured side gets measured". 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.

The reason this pairing matters beyond a spec sheet is that both models are selling the same thing. Mercury 3 is a diffusion language model that refines blocks of tokens in parallel; Reflection Beam is a sparse mixture of experts with 501 billion total parameters and only 23 billion awake per token. Different tricks, same bet: that the cost of inference, not the ceiling of capability, is where the next fight is. Beam was announced on 5 October 2026 and its weights landed under Apache 2.0 on 16 October; Mercury 3 shipped nine days later as a closed API. One of them you can download and measure yourself. The other you can only call and take on trust.

What each side has actually put on the record

Beam's record is detailed enough to argue with. An Intelligence Index of 30, ranked twentieth of 121 models and above the median of 19 that Artificial Analysis records for comparable open-weight models. A measured 335.1 output tokens per second and 0.80 seconds to first token, both against a peer median near 77 tokens per second and 1.70 seconds. A published rate of $0.70 per million input tokens and $2.10 per million output with an 84 per cent cache discount, working out to $0.57 of compute per index task. And a verbosity figure of 79 million output tokens to complete the index, tenth leanest of 47 evaluated models, against a 140-million median. That last number is the efficiency claim showing up in somebody else's harness for the first time, and it is the most interesting figure in either model's file.

Mercury 3's record, by contrast, is three claims and one small experiment. The claims: 1,068 tokens per second on NVIDIA GPUs; more than 40 times the parameters of Mercury 1 with roughly threefold the measured intelligence; and quality "on par with or better than frontier speed-optimized models like GPT-6 Luna and Claude Haiku 5.5". The experiment: 50 live Mind2Web navigation tasks in Inception's own web-agent harness, 50 per cent completed, "significantly more than GPT-6.1-Sol", at the same token volume and step count and therefore a fraction of the cost. None of it is independently reproduced, and Artificial Analysis has no Mercury 3 page at all.

That asymmetry is not a knock on Inception's engineering. It is a statement about what you can decide today. For Beam you can compute cost per completed task, because somebody measured completion and verbosity. For Mercury 3 you can compute cost per token and nothing else.

The spec sheet, with the sourcing attached to each line

• Price per million — Mercury 3 $0.20 in and $0.70 out at list, $0.10 and $0.35 during the launch discount vs Beam $0.70 in and $2.10 out

• Cached input — Mercury 3 $0.02 per million at list vs Beam an 84 per cent discount on the input rate, or roughly $0.11 per million

• Throughput — Mercury 3 1,068 tokens per second, vendor claim, hardware unspecified vs Beam 335.1 tokens per second, measured by Artificial Analysis

• Time to first token — Mercury 3 not published vs Beam 0.80 seconds, measured

• Intelligence Index — Mercury 3 no entry as of October 20, 2026 vs Beam 30, ranked 20th of 121

• Cost per index task — Mercury 3 cannot be computed without a score vs Beam $0.57

• Output verbosity on the index — Mercury 3 not published vs Beam 79 million tokens, 10th leanest of 47

• Context window — Mercury 3 1,000,000 tokens with 131,072 maximum output vs Beam 1,048,576 tokens with 131,072 maximum output

• Parameters — Mercury 3 undisclosed, described only as more than 40 times Mercury 1 vs Beam 501B total, 23B active per token

• Input modality — both text only

• Reasoning control — Mercury 3 four settings from instant to high, defaulting to low vs Beam a mandatory reasoning effort with five levels, and Reflection's documentation defaults to max

• Weights — Mercury 3 closed vs Beam Apache 2.0, in BF16, FP8 and a 273 GiB NVFP4 build

• Self-hosting — Mercury 3 enterprise dedicated capacity only vs Beam a reference deployment of four NVIDIA B300 GPUs on NVLink, with day-zero vLLM, SGLang, Transformers and Dynamo support

One caution on the price line. Mercury 3's discount has no published end date, so the number that survives planning is $0.20/$0.70, not $0.10/$0.35. Beam's $0.70/$2.10 is a live rate card published by the evaluator and by Reflection's own platform.

The gap that matters is 335 against 1,068

If Inception's figure holds under a neutral harness, Mercury 3 is roughly three times faster than the fastest measured open-weight model in the field, and it is doing that at a third of Beam's input price and a third of its output price. That combination — faster and cheaper — is the case for Mercury 3 in a single sentence, and it is a strong sentence.

The problem is that the sentence rests entirely on a number with no stated methodology. Inception does not say which GPU, what batch size, which reasoning effort setting, or whether the measurement is single-stream or aggregate. All four choices move a throughput figure by multiples. The discount to apply is not hypothetical: when Mercury 2.5 launched, the vendor throughput number was materially higher than the 706 tokens per second Artificial Analysis subsequently measured under its own harness. That is the normal vendor-to-independent spread in this category, not bad faith — but it is the spread to assume.

Apply a three-to-one haircut and Mercury 3 lands near 350 tokens per second: level with Beam's measured figure, at a third of the price, with an unknown quality score. Apply a two-to-one haircut and it is comfortably the fastest model either lab has shipped at this price. Either way the comparison changes from "three times faster" to "somewhere between as fast as Beam and three times as fast, and cheaper than Beam either way".

There is a second asymmetry hiding in the same pair of numbers. Beam's 0.80-second time to first token is measured on a reasoning model that defaults to its maximum effort setting — which means the model is doing real thinking before it emits anything, and still returning the first token in under a second. That is a more impressive engineering result than the raw throughput number, and it is the kind of thing only an independent harness finds.

Twenty-three billion active parameters is the real story on Beam's side

The 501B figure on Beam's card is mostly not working. At 23 billion active per token, the model wakes roughly 4.6 per cent of itself for any given token, an activation ratio that is aggressive even by current open-weight standards. The architectural notes name interleaved local and global attention, fine-grained routed experts, SandwichNorm and FP8/FP4 serving builds, but the practical consequence is simpler: Beam fits on one four-GPU node. That is a serving profile, not a research curiosity.

It also explains the verbosity result. A model that spends 79 million output tokens completing an index against a 140-million median is doing the same work in roughly half the output tokens, and output tokens are where the bill lives. Reflection's own argument is efficiency at inference time rather than parity at the frontier, and the company says as much in its launch material — it concedes ground to Kimi K3 and G​LM on raw capability. The independent data supports the concession and the efficiency claim at the same time.

Mercury 3's efficiency argument is different in kind. It is not about spending fewer tokens; it is about producing them faster, and Inception's scaling story — 40 times the parameters of its first model, three times the measured intelligence, same speed — is a claim that the parallelism of diffusion buys throughput that autoregressive models cannot match. That is a plausible and interesting argument. It is also entirely self-reported, and it says nothing about whether the model spends more tokens than a comparable autoregressive model to reach the same answer. Nobody has published a Mercury 3 verbosity figure, because there is no Mercury 3 evaluation to take it from.

Cost per completed job, worked through

Take an agentic coding task: a 200,000-token repository context, a 4,000-token patch and explanation out, and 5,000 runs a day across a team.

• Mercury 3 at list — 1,000 million input tokens at $0.20 is $200; 20 million output at $0.70 is $14. About $214 a day.

• Mercury 3 at the launch rate — about $107 a day.

• Beam at its published rate — 1,000 million input at $0.70 is $700; 20 million output at $2.10 is $42. About $742 a day, before the 84 per cent cache discount, which matters enormously here because repository context is exactly the kind of preamble that gets cached. With a realistic cache-hit rate on the context, the input line collapses toward $112 and the daily total lands near $154.

That arithmetic is the reason both of these models exist. On a well-cached agentic workload the two land within 40 per cent of each other — Beam slightly cheaper in a good cache scenario, Mercury 3 cheaper in a bad one — and both are a fraction of what a frontier model would charge for the same traffic. The cost argument does not separate them. What separates them is that Beam's number is reproducible and Mercury 3's is a list price against an unknown completion rate.

The licensing question, which is not close

Beam's weights are downloadable under Apache 2.0 in three precisions, plus a 1.93B speculative-decoding drafter, and Reflection states it does not restrict distillation, fine-tuning, quantisation or merging. There is a conformance tool — beam-verifier — that runs a test suite against any live Beam endpoint and then evaluates it on the main benchmarks, which means a buyer can check a provider's numbers rather than trust them. Self-hosting removes the per-token question entirely and converts it into a GPU-hours question.

Mercury 3 offers none of that. It is closed, it is API-only, and its enterprise tier adds dedicated capacity and configurable data retention rather than weights. For a workload where the model is a component in a product you sell, that is a dependency rather than a supply chain.

There is also the vendor-history signal. Inception's own migration note says Mercury 2 and Mercury 2.5 are closed to new users, with existing customers losing access on November 15, 2026 — less than a month after Mercury 3 shipped. A company that retires two generations that quickly is telling you how it expects its models to be consumed: as short-lived rented services. Reflection, for now, has done the opposite and handed over the artefact.

Where a router fits, given that neither model is on our catalogue

Neither Mercury 3 nor Reflection Beam is served by OrcaRouter. A check of our model API returns nothing for either, and we will not describe a route we do not have. What we do carry is the layer around them: more than 200 models behind one OpenAI-compatible endpoint, at provider list price with zero markup, with automatic failover and a routing DSL for composing models into a single call.

For this specific pair, that layer does two jobs. The first is comparison. If you want to know whether a 300-token-per-second open-weight model is fast enough for your product before you commit to any of this, the honest test is to run the workload against models you can already call — Kimi K3 at $3.00/$15.00 with 1,048,576 tokens of context, GLM 5.2 at $1.40/$4.40, Qwen 3.8-Max at $2.00/$6.00 — and measure what your users actually notice. The second is fallback. A self-hosted Beam deployment and a vendor-hosted Mercury 3 are two different failure modes; putting the production path behind a router with failover means an endpoint degradation becomes a logged event rather than an outage. Because provider list prices pass through unchanged, a Reflection price cut or an Inception discount reaches your bill on the day it is announced.

The decision, by constraint rather than by score

• If you need the model to be an asset — auditable, fine-tunable, runnable on your own hardware, survives the vendor — Beam is the only option here. Apache 2.0, a 273 GiB build that fits on one node, and a verifier to check any provider against Reflection's numbers. That is a complete answer to a question Mercury 3 cannot answer at all.

• If you need the lowest possible latency on a closed API and a human is waiting, Mercury 3 is the more interesting bet, and its price makes the evaluation cheap. Pin the effort setting explicitly when you test it, because the default is low and Inception's claims do not say which setting produced them.

• If you need to know the quality before you commit, Beam has a number and Mercury 3 does not. A 30 on a 121-model index is mediocre — it is not a frontier score and Reflection does not claim it is — but it is a floor you can design against.

• If your workload is agentic and long-running, Beam's mandatory reasoning at a default of maximum effort is expensive per call, and Mercury 3's low-effort default is cheap per call. Which one is cheaper per completed job depends on a completion rate nobody has measured for Mercury 3.

The single most useful thing that could happen to this comparison is an independent Mercury 3 measurement. If it arrives near 1,068 with a competitive quality score, Inception has shipped the fastest cheap model in the field and Beam's efficiency story becomes a story about openness rather than raw speed. If it arrives near Beam's 335 with a mid-table score, then the fastest open-weight model and the fastest closed one are the same speed, and the deciding questions become price, licence and whether you want to own the artefact — three questions to which Beam currently has better answers.