OpenRouter Adds Reflection Beam 1.0 With 1M-Token Context

OpenRouter has added Reflection: Beam 1.0 501B A23B to its model catalogue, listing a 1,048,576-token context window and rates of $0.25 per million input tokens and $1 per million output tokens. Its Beam 1.0 model listing names Nebius Token Factory as an inference provider.

The addition on 16 October 2026 gives developers a hosted route to evaluate a large, text-only reasoning model aimed at coding and multi-step agent workflows. Reflection first announced Beam on 5 October; today’s wider availability release includes downloadable weights under Apache 2.0. OpenRouter is the access platform, not the model’s developer.

Verification boundary: DIY AI checked the public listing and release documentation. We have not completed an authenticated Beam inference request, measured its latency or tested the full context window.

API access is listed, but request settings still need checking

Developers can use OpenRouter’s chat-completions endpoint with an API key and the model ID above. Start with a short request before connecting a repository or agent. This example uses an existing OPENROUTER_API_KEY environment variable and is a starting configuration, not a recorded test:

curl --fail-with-body --max-time 120 \
  https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer ${OPENROUTER_API_KEY}" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "reflection/beam-1.0-501b-a23b",
    "messages": [
      {
        "role": "user",
        "content": "Explain the difference between a list and a tuple in Python."
      }
    ],
    "max_tokens": 4096,
    "provider": {
      "require_parameters": true
    }
  }'

Beam’s listing advertises function calling through tools and tool_choice, structured outputs through response_format, and configurable reasoning effort. These are integration features, not proof that every tool call or generated JSON object will satisfy your application.

OpenRouter documents that, under its default routing, unsupported parameters can be ignored. Setting provider.require_parameters to true excludes providers that do not declare support for the requested settings. It does not independently verify their implementation.

A recurring lesson from developer audits of routed APIs is to record the serving provider and configuration, not just the model name. That makes a later regression easier to investigate. This is a general routing precaution, not evidence of a Beam-specific fault.

The million-token window is a budget, not a recall guarantee

A large context window could help with repository analysis, lengthy specifications and extended tool histories. It does not demonstrate that Beam can reliably connect every relevant detail across those inputs.

The context and completion limits are not additive. OpenRouter’s parameter documentation says the generation budget cannot exceed the context length minus the prompt length. System instructions, conversation history and returned tool content therefore leave less room for new material and the answer.

Reasoning adds another constraint: on most reasoning-model providers, max_tokens covers reasoning and visible output together. An overly small allowance can end generation before a useful answer appears. Check the finish reason and token usage before diagnosing an empty response as an outage.

For a meaningful long-context check, place known facts in different parts of a document and require answers supported by specific passages. Then ask a question that needs evidence from several sections. Accepting the input and answering correctly are separate milestones.

Low token prices do not settle the cost of an agent task

At the listed rates, a request with 100,000 input tokens and 5,000 billable output tokens would cost $0.03 in model token charges: $0.025 for input plus $0.005 for output. This is an arithmetic illustration, not an observed Beam bill, and assumes no caching discount.

An agent may make several requests, resend growing histories and retry unsuccessful changes. Compare the total charge for a completed job, not a single response. Record reasoning usage, tool failures and human correction time alongside the bill. Our AI model cost comparison covers that wider buying decision.

23B active parameters do not mean a 23B-sized deployment

Beam’s sparse architecture activates a subset of its parameters for each token. The rest of the model still has to be stored and made available to the serving system. The active count helps explain per-token computation; the total count drives the weight-storage requirement.

Hugging Face’s launch guide puts the BF16 checkpoint at roughly 1 TB and the NVFP4 version at approximately 280 GB. Those figures exclude additional working memory and context-cache requirements. NVFP4 also has specific hardware requirements, so its smaller size is not a universal deployment shortcut.

For an initial evaluation, hosted access avoids committing to that infrastructure. Teams seeking workstation-scale options should compare local LLMs for coding against their actual memory budget instead of choosing by active parameters alone.

Published benchmarks still need to translate into your workflow

Reflection has published benchmark results in its model card, including separate results for different weight formats. These are developer-reported evaluations, not DIY AI results or proof of equivalent behaviour through every hosted endpoint.

DIY AI’s testing methodology and code-generation dataset are separate from these vendor evaluations. This news report does not rate Beam.

The immediate news is that Beam now has a listed OpenRouter provider, published token prices and downloadable weights. The remaining buying questions are practical: can it complete your task, honour your required output format and remain economical after retries? Test those before making it the default model in an unattended agent.