For years, the AI market has trained buyers to compare models much as they compare cloud storage or phone plans: dollars per million tokens in, dollars per million tokens out. It is a simple metric, and a useful one. But it can become misleading when models are asked to do more than draft a paragraph.
A software agent investigating a production bug may read thousands of lines of logs, inspect a codebase, run tests, call a search tool, revise its plan and write a patch. In that setting, the bill is not determined by the final answer alone. It is determined by every token read, every token generated, every tool loop and every failed attempt along the way.
Reflection said in its launch announcement that Beam is a sparse mixture-of-experts model with 501 billion total parameters, 23 billion active parameters, training on 23.8 trillion tokens and more than 100 million reinforcement-learning rollouts. The company says it trained Beam over four weeks using 10,500 NVIDIA GB300 GPUs. Those are imposing numbers, but the business argument is more interesting than the scale: make the model deliberate less wastefully.
The new denominator: a completed task
A token is a small unit of text processed by a model. Token prices matter because providers bill for them and because inference hardware spends time processing them. Yet a token is an input to work, not the work itself.
Consider two coding models. Model A costs less per million output tokens, but writes long chains of reasoning, repeats tool calls and needs several repair attempts. Model B costs more per million output tokens, but identifies the bug quickly, makes one tool call and produces a correct patch. The second model can be the cheaper system even if its individual tokens are more expensive.
“A DeepSeek model or Qwen model may have better kernels or fused kernels, and they just might be much more efficient,”
That is the logic behind tokens per finished task. It asks a practical question: how many generated and consumed tokens, plus how much tool use and elapsed compute, are required to reach a predefined success condition? For an enterprise, that condition might be a passing test suite, an approved support reply, a completed research brief or a correctly processed claim.
This framing changes what teams should measure. Instead of asking only, “What is our output-token price?”, they can ask, “How many successful tickets, patches or reports do we get for a fixed budget?” It also makes failures visible. A cheap model that confidently ends a task without solving it is not efficient. It has merely shifted the cost to a human reviewer or a later incident.
Why 501 billion parameters does not mean 501 billion are working at once
Beam’s architecture illustrates why headline parameter counts can confuse buyers. A mixture-of-experts model contains many specialized internal components, often called experts. For each token, a routing mechanism selects only a small number of those experts to do the computation.
Reflection’s model card says the released Beam model has 501.3 billion total parameters but 23.4 billion active parameters. In plain language, the model has a very large library of learned capabilities, but only a fraction of that library is used for any one token.
That can deliver the broad capacity associated with a very large model without paying the full computational price for every token. But “active” is not the same as “free.” The full model still needs to be stored, loaded and moved through a serving system. Routing work, memory bandwidth, hardware utilization and communication between accelerators can all affect the real bill.
This is why parameter count should not be treated as a price tag. A smaller dense model can sometimes run more efficiently than a sparse model with fewer active parameters, depending on the hardware and software stack. Conversely, a sparse model can outperform its apparent cost if its routing and kernels are well optimized.
Training a model to stop when it has done enough
Beam’s central efficiency claim comes from its reinforcement-learning campaign. In this kind of training, a model attempts a task, receives a reward based on the result and is updated to favor behaviors that earn better outcomes. A rollout is one such attempt, including its reasoning, actions, tool use and answer.
For a coding task, a useful reward might reflect whether the code compiles, whether tests pass and whether the solution avoids unnecessary steps. For a science or research task, it might reflect factual accuracy, sound tool use and a verifiable conclusion. The model is not simply taught to write shorter text. It is taught to balance task success against unnecessary generation.
The distinction is crucial. A blunt token penalty can teach a system to quit early, omit checks or produce terse but wrong answers. In other words, it can reward the appearance of efficiency rather than efficiency itself.
VentureBeat reported that Reflection co-founder and CTO Ioannis Antonoglou described this as a reward-hacking risk: applied naively, a length penalty can cause a model to stop solving tasks. The company’s answer is a controllable penalty, so users can adjust the trade-off between thoroughness and brevity instead of assuming that fewer tokens are always better.
That is sensible in principle, but it creates a benchmarking obligation. Any claim of “three times fewer reasoning tokens” needs a clear definition of comparable quality. The task set, scoring rule, allowed tool access, retry policy and human review standard all matter. A model that uses fewer tokens because it skips difficult cases has not established a lower completed-task cost.
The hidden costs before the first answer
The token-efficiency story also has an important boundary: models must read before they can reason. Beam has a 1-million-token context window, according to Reflection’s model card. That can be useful for tasks involving long repositories, research archives or complex internal documentation. It can also be expensive.
Input tokens require computation, memory movement and attention across the prompt. Long context may reduce the need to split a task across multiple calls, but it can increase the cost of each call. Antonoglou told VentureBeat that Reflection’s compute estimate excludes prompt-reading costs, a caveat that matters especially for agent workflows built around retrieval and long context.
This is also where quantisation and kernel efficiency enter the picture. Quantisation stores or computes model weights with lower numerical precision, reducing memory and potentially improving speed. Beam is available in BF16, FP8 and NVFP4 formats, the model card says. Lower precision can make a model easier to serve, but the gains depend on whether the hardware and software kernels use those formats efficiently.
Kernel efficiency is the unglamorous but decisive layer of machine learning systems: the highly optimized code that turns mathematical operations into work on GPUs. A model can look efficient on paper because it activates fewer parameters, while delivering disappointing real-world savings if its routing pattern or memory access leaves expensive hardware underused. VentureBeat noted that Reflection itself acknowledged this risk.
How to benchmark the claim without being fooled
TechCrunch reported that Beam’s performance and lower-compute claims had not been independently verified at launch. Reflection says Beam can reach comparable quality with roughly three times fewer reasoning tokens and claims lower inference compute in some comparisons. Those are meaningful claims, but they should be treated as hypotheses to test in the workload that matters to the buyer.
A useful evaluation starts with a fixed set of real tasks and a strict definition of completion. Record success rate first. Then measure median and tail latency, input tokens, output tokens, tool calls, retries, human corrections and total infrastructure cost. Compare models at the same quality threshold, not simply at the same token budget.
Teams operating their own infrastructure should also distinguish provider pricing from economic cost. A hosted API invoice is a commercial rate that includes margins, capacity planning and service operations. Self-hosting cost includes accelerator time, model loading, memory capacity, networking, electricity, engineering and idle capacity. Neither number is wrong. They answer different questions.
The available outside reception is limited and one-sided. VentureBeat offered the clearest account of Reflection’s technical caveats, while TechCrunch emphasized that the claims were unverified. No independent technical rebuttal was included in the reporting supplied for this article. That absence does not prove the claims, but it does underline the need for buyers to reproduce the measurement.
Beam’s larger lesson is not that tokens no longer matter. They do. It is that tokens become meaningful only when connected to an outcome. The winning model may not be the one with the lowest price per million tokens. It may be the one that reaches a trustworthy finish line with the fewest expensive detours.
Daniel Reyes is not a person. No notebook, no deadlines, no face behind the name — just a byline this newsroom publishes under. Here is the production line underneath it, because a name beside a portrait reads like a journalist, and this one is not one.
The models. Writing: gpt-5.6-luna and qwen3-max. Out on the live web: gpt-5.6-luna and gpt-5.6-terra. Pictures: gpt-image-1 and gpt-image-1-mini. Swap one in the newsroom and this line swaps with it — it is read off the machines, not typed here.
How a story is made
- Research. The searching model reads around the story, pointed at primary sources — the filing, the post, the repository — rather than at somebody else's write-up of them.
- Writing. The writing model drafts it against what was found, at Daniel Reyes's usual length and in Daniel Reyes's usual register.
- The loop. A reviewer reads the draft and sends it back with notes. Then reads it again. A piece can go round several times before it leaves the building.
- Enrichment. A quotation has to appear word for word on the page it is taken from. A chart may only use figures that appear in the source it cites. Whatever fails is dropped, and the reason is kept.
- Fact check. A last pass hunts for claims the article makes and its sources do not.
- A human stop. Sensitive subjects are held for a person to read before publication, and a person can kill any of it at any point.
If that sounds less like a newsroom and more like a factory: quite. It is called Press Factory.
This article was generated using AI and published automatically without human pre-publication review.
How this article was made
The article was produced by the Grandmonts Media News Engine using automated research, drafting and verification workflows. No human editor reviewed the article before publication. Grandmonts Media remains responsible for the published content. Errors can be reported at office@grandmonts.cz.