
Qwen4-Max vs Reflection Beam: $1.20 Output Against $4.50, and the Weights That Decide It
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 90 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 97 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-0717 tok/s
- OrcaNEWOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 87 tok/s
- OrcaNEWOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 133 tok/s
- OrcaNEWOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$4.00 / $20.00 per 1M tokens · 72 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 130 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 65 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 92 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 43 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 193 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 150 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
- OrcaOrca: OrcaCyber Zero 1.52026-10-10$3.00 / $7.50 per 1M tokens · 119 tok/s
- openaiOpenAI: GPT-6.1 Sol2026-09-2955Intelligence
- anthropicAnthropic: Claude Sonnet 5.52026-09-2855Intelligence
- typesafeTypeSafe: Jev 1.132026-09-24$0.04 / $0.00 per 1M tokens · 219 tok/s
Reflection Beam 1.0 will generate a million output tokens for $1.20 on its own API. Qwen4-Max charges $4.50 for the same million in Singapore, $4.24 in Frankfurt, Virginia, Tokyo, Hong Kong and Beijing. That is a 3.75× gap in favour of the open-weight model, and it is the sort of number that ends most comparisons before they start. It does not end this one, because the two models are not the same kind of product: Reflection Beam 1.0 is a 501-billion-parameter sparse mixture-of-experts model with Apache 2.0 weights you can download and run on four GPUs, and Qwen4-Max is a closed Alibaba flagship that became generally available with a published rate card on 16 November 2026 and publishes no weights at all. One is a model you can own. The other is a model you can rent, at a price that buys you a much larger context window, native video input and a vendor with a support contract.
There is also a measurement question, and it cuts against the cheaper model hard. Beam 1.0 scores 17.7 on the current Artificial Analysis Intelligence Index. Qwen4-Max has no score at all. So the honest framing is not "cheap open model versus expensive closed model" — it is "a measured model that is cheap and not very strong, against an unmeasured model that is expensive and claims a great deal."

What Beam 1.0 actually is
Reflection, the New York lab founded in 2024 by two former Google DeepMind researchers, announced Beam on 5 October 2026 and shipped the weights on 16 October. The model is Beam 1.0-501B-A23B: 501.3 billion parameters total, 23 billion active per token, an activation ratio of about 4.6%. Reflection states it does not restrict distillation, fine-tuning, quantisation or merging.
The release is unusually complete for a first model. BF16 at 188 safetensors shards, FP8 in E4M3 at 468.72 GiB, and an NVFP4 build at 273 GiB that Reflection's own serving guides recommend downloading. A separate 1.93-billion-parameter speculative decoding drafter, DFlash2, published as its own repository. Day-zero support in Transformers, vLLM, SGLang, Unsloth, Axolotl, TokenSpeed and NVIDIA Dynamo, with step-by-step guides for SGLang, vLLM and TokenSpeed plus a Helm chart. A conformance suite, beam-verifier, that runs against any live Beam endpoint. And a reference self-hosting requirement of four NVIDIA B300 SXM GPUs on NVLink — one box, not a cluster, which is a much lower bar than a 500B model normally implies and follows directly from the 23B activation ratio.
Context is 1,048,576 tokens. The model is text-oriented and pitched at coding, reasoning and agentic work.
What Qwen4-Max actually is
Alibaba's fourth-generation flagship, in general availability since 16 November 2026 under the model ID qwen4-max. A 1,000,000-token context, a 131,072-token maximum output outside thinking mode, a shared 262,144-token budget for thinking plus answer, and text, image and video input with text output. Thinking is on by default with a 131,072-token budget and reasoning effort at high. Six regions. OpenAI, Anthropic and DashScope wire protocols. Fine-tuning and model deployment are documented as unsupported, and no weights are published.

Side by side, on the dimensions that decide whether you rent or own:
• Input price — Qwen4-Max $1.50 per 1M in Singapore and $1.413 elsewhere, versus Beam 1.0 $0.40
• Output price — Qwen4-Max $4.50 Singapore, $4.24 elsewhere, versus Beam 1.0 $1.20
• Context — Qwen4-Max 1,000,000 tokens versus Beam 1.0 1,048,576 tokens
• Max output — Qwen4-Max 131,072 non-thinking, 262,144 shared with thinking, versus Beam 1.0's ceiling, which Reflection's public material does not break out separately from context
• Modality — Qwen4-Max text, image and video in, versus Beam 1.0 text in
• Weights — Qwen4-Max none, versus Beam 1.0 Apache 2.0, downloadable in three precisions
• Self-hosting — Qwen4-Max not permitted and not possible, versus Beam 1.0 on four B300 SXM GPUs by the vendor's own reference build
• Independent score — Qwen4-Max none published, versus Beam 1.0 at 17.7 on Artificial Analysis Intelligence Index v5.0.1
The index-version trap in Beam's score
When Beam shipped in October, the first independent number attached to it was 30, on index version v4.3.2, ranked nineteenth of 120 models. The current board reads 17.7 on version v5.0.1. Neither number is wrong; they are not comparable, because Artificial Analysis re-based the index between those vintages and every model on the board moved. Anyone quoting "Beam scored 30" today is quoting a retired vintage, and anyone quoting "17.7" without the version is inviting the same mistake in reverse.

What survives the re-basing is the relative picture. On v5.0.1, Beam 1.0's 17.7 sits well below Qwen3.8-Max at 33.7, Grok 4.7 at 35.5 and Kimi K3.1 Pro at 46.5. It is a working model that is not competitive with the current frontier on general capability, which is exactly what its price implies and exactly what Reflection's launch material did not say plainly. The vendor's claim at announcement was a frontier-adjacent reasoning model; the independent measurement says otherwise.
And Qwen4-Max has no number on either vintage. Its per-model page on Artificial Analysis returns a 404 and no Qwen4 entry appears in the index. Alibaba's documentation specifies the mechanics in detail — default temperatures of 1.0 with thinking and 0.7 without, top-p 0.95 and 0.8, top-k 20, per-mode input ceilings of 991,808 and 983,616 tokens — and publishes no benchmark table of any kind.
What self-hosting Beam actually costs
The $0.40 and $1.20 figures are Reflection's own API rates. They are not what Beam costs you if you run it yourself, and the difference is the whole reason open weights matter here.
Four B300 SXM GPUs on NVLink is a purchase, a lease or a cloud reservation, and none of those is free. The point of owning is not that the per-token cost is lower — it is that after the hardware is amortised, marginal inference cost approaches the electricity bill, and that nothing about your data leaves the building. For a team already running B300 capacity for other work, Beam 1.0's marginal cost is close to zero and its 17.7 score is a bargain. For a team that would have to buy the box, the same model is a capital commitment made on the strength of a mid-table score.
Qwen4-Max is the opposite trade in every respect. There is no capital cost, no capacity planning and no ops burden, and in exchange you pay 3.75× Beam's API rate, you cannot fine-tune, you cannot run it inside your own perimeter, and you are exposed to whatever Alibaba does to that rate card next quarter. The regional structure of the card is the interesting part: Singapore at $1.50 and $4.50 against $1.413 and $4.24 in the other five regions is a 6% difference for the same model, which is small enough to be an accounting detail and large enough to matter at volume.
Where each one earns its place
Beam 1.0 is the right answer when any of three things is true. Your data cannot leave your infrastructure — no contracted API clears that bar, and Apache 2.0 weights do. You want to fine-tune on domain data — Qwen4-Max's documentation explicitly marks fine-tuning unsupported, while Reflection permits distillation, fine-tuning, quantisation and merging without restriction. Or your workload is agentic coding where a 23B-active model at interactive latency is fast enough and cheap enough that a mid-table general score is not the binding constraint.
Qwen4-Max is the right answer when the input is not text. Beam takes text; Qwen4-Max takes text, images and video. If your pipeline ingests screenshots, scanned documents, frames from a camera or a screen recording, the comparison is over before price enters it, because one of these models cannot accept the request. The same applies to anything that needs a European, Japanese or US-East deployment: six Alibaba regions against a single first-party API is an eligibility question, not a cost one.
For everything else — general text generation at moderate volume, no fine-tuning, no residency constraint — the decision comes down to how much you trust an unmeasured model against a measured mediocre one, and the answer is that you should not trust either until you have run a hundred of your own prompts through both. At $4.50 and $1.20 per million output tokens, that evaluation costs cents.
Routing, and the case for a third option
OrcaRouter does not currently route Beam 1.0 or Qwen4-Max; both are reachable through their vendors' own APIs at the rates above. What the catalogue does offer is the middle of the market on one key and one base URL — GLM-5.3 at $1.40 and $4.40 per million with open weights and a 31.7 index score, Qwen3.8-Max at $2.00 and $6.00 with a 33.7, Kimi K3.1 Pro at $2.00 and $10.00 with a 46.5 — all passed through at provider list price with no markup.
That is the practical resolution of this particular matchup. Beam 1.0 asks you to accept a 17.7 in exchange for ownership and a low price; Qwen4-Max asks you to accept an unknown in exchange for modality and regions. A model with a published score, published weights and a price between the two asks you to accept neither. Routing both of the interesting-but-unproven options alongside a measured baseline is how you find out whether either deserves production traffic — and automatic failover means the experiment cannot take the service down while you run it.
What would settle it
One number: a Qwen4-Max entry on an independent leaderboard, on the current index vintage, next to Beam 1.0's 17.7. Everything else on this page is already known — the prices, the contexts, the licences, the GPU count. If Qwen4-Max lands anywhere near the 30s, its 3.75× premium is easy to justify for anyone who needs video input or a Frankfurt endpoint. If it lands near Beam, the premium is buying modality and compliance and nothing else, and the right answer is to route by requirement rather than by benchmark. Until that number exists, the comparison is between a model that has been measured and found ordinary and a model that has not been measured at all — and only one of those two positions is a reason to commit budget.
