Reflection Beam explained: what it changes for SaaS buyers

What Reflection launched and what it changes

Reflection AI launched Beam, a 501 billion-parameter open-weight coding model. For a small B2B software team, it is a model to evaluate this month, not a tool to switch to, because no per-token price is published and self-hosting runs on multi-GPU AWS instances that need quota.

Reflection's launch post, dated October 16, 2026, says the weights are on Hugging Face under the Apache 2.0 license and that its API platform is "free to start," according to Reflection (2026). Beam was first announced on October 5, 2026, and DataCamp checked Hugging Face on October 6, 2026 and found no weights listed yet, according to DataCamp (2026), so the weights arrived with the October 16 post, after the initial announcement. For a service business, what changes is the option: a coding model you can host inside an AWS account you already control, rather than only calling a third party's hosted API.

Beam is a sparse mixture-of-experts model, which means it stores a large set of weights and activates only part of them for each token. That design is why the active-parameter count matters more than the headline size when you think about per-token compute.

TL;DR: Beam is an open-weight coding model you can call through Reflection's public beta or deploy on AWS. Its hosted price is not published, so the cost comparison with GLM-5.2 stays incomplete until a rate appears.

Key Takeaways

• Beam is an open-weight coding model, and its weights and AWS deployment guide were both public by the October 2026 launch window.

• Beam has no published per-token price, while GLM-5.2 has a posted rate, so the cost comparison is one-sided for now.

• The AWS route runs on SageMaker JumpStart real-time endpoints or HyperPod Inference, and both depend on GPU quota.

• Both models advertise long context windows, but Beam's model card does not say whether its context is effective or maximum.

• Beam's coding score is vendor-reported and the model card names no rival, so run your own evaluation before any change.

• The launch post says Apache 2.0, but the model card names no license, so check the license file in each repository you deploy.

Rival status: GLM-5.2 is not discontinued, but its model page points to a newer GLM-5.3-BF16 checkpoint, and Z.ai's pricing page still lists GLM-5.2 among its latest models. Confirm which model ID your contract covers before you compare.

Who this is for

This is for a small software company whose engineers already use an AI coding agent, want an open-weight option their cloud account can host, and have someone who can own a GPU endpoint. It is a weaker fit for a team that needs a fixed monthly bill today. For related background, see how SaaS teams save on workspace tools, GPT-6 Astra vs Claude Fable 5.1 for coding and GPT-5.6 Sol alternatives after the Astra launch.

Red flags: no one on the team can run a GPU endpoint or a Kubernetes cluster; client code or data must stay out of any hosted API; or you need a published per-token rate before the next budget cycle.

How we evaluated Beam against GLM-5.2

The weights below reflect what a small software buyer decides on this month: spend, license and control, fit with an existing cloud, the quality of evidence behind vendor claims, and maturity. They are our judgment, not a vendor or analyst standard, and they sum to 100%. Price carries the most weight because a rate you can forecast is the first thing a buyer needs before a trial becomes a commitment.

Feature matrix: Beam against GLM-5.2

Reflection's model card gives Beam BF16 SWE-bench Verified: 81.6% according to Reflection (2026), and says the model uses approximately 3x fewer inference-time reasoning tokens than leading open models, without naming them. The matrix below restates figures cited in this article; a blank or "not stated" cell means the vendor pages reviewed did not publish that figure.

Pricing and total cost

Pricing checked October 17, 2026.

Z.ai lists GLM-5.2 at GLM-5.2 input price: $1.40 per million tokens according to Z.ai (2026). Output is $4.40 and cached input $0.26 per million tokens, according to Z.ai (2026), and that pricing table carries no date.

Reflection's launch post gives no paid tier and no per-token rate, so the hosted Beam price has no published rate beyond "free to start" during the beta (ask Reflection for a written quote). The SageMaker side is different: the SageMaker AI pricing page says real-time hosting is billed by the instance type you choose, but the page we read does not print an hourly rate for the Beam instance types, so this review gives no self-hosted rate. The same page says charges for services connected to HyperPod clusters, such as EKS, FSx and S3, are not covered by its prices.

Here is an illustrative workload, not a measured one. Suppose a four-person team runs 200 agent runs a month, each sending 150,000 input tokens and returning 20,000 output tokens, and prices that workload on the GLM-5.2 input and output rates cited above. The month totals 30 million input tokens (200 × 150,000) and 4 million output tokens (200 × 20,000). At $1.40 per million input tokens, the input side costs $42.00 (30 × $1.40), and at $4.40 per million output tokens, the output side costs $17.60 (4 × $4.40), for $59.60 in total at list rates. The same runs sent to https://api.reflection.ai/openai/v1 produce no cost line, because no per-token rate is published, so any Beam comparison has to wait for a written quote. A self-hosted run through the AWS guide's jumpstart_real_time.py script would be billed by instance-hour rather than by token, and the pages read print no rate for it.

Reflection Beam: best fit, limits and implementation

Best fit: teams that want an open-weight coding model in an AWS account they already control and can accept a beta API while the rate is unset. AWS presents the managed path first: a SageMaker JumpStart real-time endpoint, which lets you use the model without operating Kubernetes. Beam active parameters: 23 billion according to Amazon Web Services (2026), and the same guide names FP8 on ml.p5.48xlarge, ml.p5e.48xlarge and ml.p5en.48xlarge, NVFP4 on ml.p6-b200.48xlarge, and two US East regions for JumpStart at launch.

Limits: the vendor has not published a per-token rate, and the API overview page reviewed did not state rate limits or maximum output tokens. Reflection's own benchmarks are not independent, and the card names no rival models. Launch coverage also quotes Reflection's efficiency claim in different terms from the card, which is covered in the signal section below.

Implementation: The AWS guide's JumpStart sample runs a script named jumpstart_real_time.py with a --variant flag, and NVFP4 is the default variant when you omit it. Use --endpoint-name to give a smoke-test endpoint its own name. The script deploys the endpoint, sends a short request, prints the response and cleans up; use the README functions if you need to keep an endpoint running. For the API route, Reflection's developer docs give the base URL and the model ID Beam-1.0-501B-A23B, and the overview shows chat completions as the endpoint in use. A secondary article used a shorter model ID, so copy the ID from the docs rather than from coverage.

The AWS guide's deployment paths, summarized:

The Hugging Face collection lists four repositories: a base checkpoint, an FP8 build, an NVFP4 build and a fourth named DFlash2, according to Hugging Face (2026). The card's BF16 numbers apply to the base checkpoint, and the collection page shows no license for any of the four, so check each repository's license file before you deploy.

GLM-5.2 on Z.ai: best fit, limits and implementation

Best fit: teams that want a hosted, per-token option with a posted rate and an MIT license, and that accept the rival's benchmark claims as vendor-reported. The model page lists MIT and 753 billion parameters in its Safetensors metadata, according to Hugging Face (2026), and describes a 1M-token context with 128K maximum output in its evaluation notes.

Limits: the model page does not state active parameters per token, and it points to a newer GLM-5.3-BF16 checkpoint. A trade article gives a different total parameter count, so quote the model card rather than secondary coverage. TechCrunch treats GLM-5.2 as the main Chinese comparison point for Beam, so a procurement or security reviewer may want to see the model's origin in your approval record before use.

Implementation: use Z.ai's hosted API for a per-token budget line, or use the Hugging Face weights under MIT if you want to self-host. For a managed API, the budget line is the per-token rate in the pricing section above. For self-hosting, your cost is GPU hours and staff time, and the model page does not price that.

Build it yourself or run a managed workflow

Most small teams will stitch this together in Zapier, Make or n8n, or build it in-house, so compare on that basis. Those tools can keep run histories, retries, error branches and audit evidence when you configure them. You still design and own idempotency, so a retried run does not open two pull requests, plus escalation paths, access controls and maintenance when the model version changes. If you would rather have someone configure this than build it, US Tech Automations can propose the design below for your review.

A proposed US Tech Automations design for engineering would start when a new issue carrying an agent label appears in your Git host. The action sends the issue text and a list of relevant files to a model endpoint you control, with a token cap per run, and writes a draft branch and a pull request marked as model-generated, plus a run log. Prerequisites are scoped Git host API access, either a SageMaker endpoint or a Reflection API key, and a rule that secrets and client code are excluded from prompts. Human review points: an engineer reviews the diff and test results before merge, and a second reviewer signs off on any change that touches license or security files.

A proposed design for the operations side of a service business would start when a client change request arrives through a form export or an API. The model drafts a scoped estimate and a requirements checklist, and the output is a draft in your ticketing system. Prerequisites are export or API access to the form and ticket tools, and a data-handling review, because client details reach the model. Human review: an account lead edits and approves the draft before anything reaches a client. Neither design reflects a live customer or a measured result; both are configurable proposals.

When NOT to use US Tech Automations

Skip a managed design when a single form-to-spreadsheet transfer covers the need, when your engineers want to own every prompt and every log themselves, or when no one can review model output before it reaches a client. In those cases a simpler existing tool or a short in-house script wins on cost and control, and a managed design adds a review burden you do not need.

Signal vs speculation

Signal (what the sources support): Reflection's launch post describes public weights, an Apache 2.0 license and a "free to start" beta; AWS's deployment guide describes the SageMaker JumpStart and HyperPod paths. The model card itself names no license. No source reviewed gives a per-token Beam rate. TechCrunch quotes Reflection on "3-4x less inference compute" against leading open models, according to TechCrunch (2026), and says the performance claims have not been independently verified. That is a compute claim, while the card's figure is about reasoning tokens, so do not merge the two when you brief a colleague.

Speculation (our reading, not vendor statements): we expect a per-token Beam rate to appear before a production decision is practical, but no source reviewed gives a date. We read self-hosting cost as mostly GPU hours and staff time rather than tokens, which follows from the deployment requirements and is not a vendor claim. We expect the GLM-5.3-BF16 mention to matter in contract reviews, because a buyer who signs for "GLM-5.2" may get a different checkpoint later.

Decision checklist for this month

1. Confirm the license file in the exact repository you plan to deploy, since the model card names none.

2. Ask Reflection for a written per-token rate and rate limits before any production use.

3. Check the AWS service quota for the instance family you need before you schedule a test.

4. Run a bounded evaluation on one repository you own, with a fixed task list and a named human reviewer.

5. Score GLM-5.2 on the same tasks and confirm which GLM model ID your contract covers.

6. Record the decision as switch, trial or wait, with the date you will revisit it.

Glossary

• Mixture-of-experts (MoE): a model design that routes each token through a subset of its expert sub-networks rather than the whole model.

• Active parameters: the share of a model's parameters used to process a single token, which is what the AWS guide cites for Beam.

• Open-weight: a model whose weights you can download and run, under a license that sets the terms of use.

• Context window: the amount of text a model can take in one request, which may be effective or maximum depending on the vendor.

• BF16, FP8, NVFP4: numeric formats for storing weights; lower-precision formats need less GPU memory.

• JumpStart: AWS's managed model catalog and endpoint service inside SageMaker.

• HyperPod Inference: SageMaker inference on GPU clusters orchestrated by Amazon EKS.

• Quota: the AWS limit on how many instances of a given type your account can run.

FAQs

Is Reflection Beam available to use today?

Yes, in two ways: Reflection's platform is in public beta, and the weights are on Hugging Face for deployment on AWS or elsewhere. The AWS route lists JumpStart in two US East regions at launch.

Does Beam have a published price?

No per-token rate is published; Reflection's launch post says only that the platform is "free to start," so no paid rate has been published; ask Reflection for a written quote.

Can a small team run Beam without Kubernetes?

Yes, the SageMaker JumpStart real-time endpoint is the path AWS describes for running Beam without operating Kubernetes, though it still needs GPU quota.

How does Beam compare with GLM-5.2 on cost?

GLM-5.2 has a posted hosted rate and Beam does not, so no like-for-like cost comparison is possible yet; the worked example above prices only the GLM side.

Should we switch models this month?

No production switch is justified by these sources; run a bounded evaluation, confirm the license and quota, and decide in writing when you will revisit.

Is the 81.6% SWE-bench figure independent?

No, it is Reflection's own figure from its model card, the card names no rival, and TechCrunch reports the performance claims have not been independently verified.

Bottom line for this month

Beam is worth a bounded evaluation this month, not a production switch, because the hosted rate is unpublished and the self-hosted path needs GPU quota. GLM-5.2 is the easier comparison on price, while Beam's advantage to test is whether your own cloud account can host the weights. If you want a proposed workflow that turns issues into reviewed drafts, see how US Tech Automations configures this and check the design against your own prerequisites.

Sources: TechCrunch, Oct 5, 2026 · Reflection launch post · Reflection model card · AWS deployment guide · Hugging Face Beam collection · Hugging Face GLM-5.2 · Z.ai pricing · DataCamp