Reflection AI's Beam now available on AI Gateway for free

Reflection AI launched Beam on Vercel AI Gateway as a low-latency open-weight reasoning model tailored for programming workloads and autonomous agent tasks, according to details shared in Vercel Blog's report. The model previously operated in stealth under the codename stealth/cat-jam and is now free to use for a limited time during its launch period when serviced directly by Reflection AI. Released under the Apache 2.0 license, Beam provides US-developed open weights combined with architecture choices aimed at reducing wait times on demanding interactive sessions.

This report covers the architecture specifications, parameter activation footprints, integration paths, and billing transitions for engineering teams evaluating open-weight reasoning options. Software engineers building multi-step agent pipelines, development teams deploying coding assistants, and technical leads monitoring infrastructure spend across tools like Cursor, Claude Code, and Codex will find the specific technical constraints and deployment steps detailed below.

What shipped in Reflection AI's Beam now available on AI

Reflection AI packaged Beam with an Apache 2.0 open-weight license and an expansive 1M-token context window designed to ingest large codebases, technical reference manuals, and extensive conversational logs. The release includes native support for tool calling, adjustable reasoning effort settings, and structured outputs, allowing developers to constrain generated responses into exact program-ready schemas. The architecture is explicitly tuned for agentic loops and software engineering workflows where tools must execute reliably across persistent workspaces.

Under the hood, the release utilizes a sparse activation profile comprising 501B total parameters, of which only 23B parameters activate per generated token. This mixture-of-experts dynamic prevents unnecessary token generation on tasks that require succinct answers, allowing the system to arrive at verified solutions faster than dense open architectures. Vercel opened access across its AI Gateway infrastructure alongside unified API endpoints, software development kit bindings, and playground testing interfaces.

What improved in Reflection AI's Beam now available on AI

Beam replaces the earlier stealth implementation stealth/cat-jam with direct, production-grade serving infrastructure across AI Gateway routing layers. The primary performance advantage stems from its sparse execution design, which slashes token generation latency while maintaining the reasoning depth of a 501B-parameter model. Developers gain fine-grained programmatic control over execution length by dialing reasoning effort up or down depending on task difficulty.

The sparse 23B parameter activation path ensures reasoning runs bypass processing overhead seen in competing large open setups, yielding faster time-to-first-token and lower response delays during iterative debugging sessions.

What you gain from Reflection AI's Beam now available on AI

Engineering teams integrating Beam through AI Gateway receive infrastructure reliability safeguards without incurring additional middleman costs. AI Gateway reflects provider pricing with zero markup, adds no platform fee on raw inference calls, and extends the same terms to Bring Your Own Key configurations. Teams running heavy test matrices benefit from automated retries, dynamic provider failover, and uptime layers designed to exceed standalone upstream provider stability.

The platform provides administrative governance features including Zero Data Retention compliance, per-key usage budgets, usage and cost tracking dashboards, custom analytics reporting, and programmable routing rules. For developers building autonomous agents, the combination of a 1M-token context window, tool calling support, and structured output guarantees eliminates parsing errors when orchestrating complex file modifications or command execution sequences across external shells.

How to get Reflection AI's Beam now available on AI

Developers can connect local environments and coding agents by upgrading the Vercel CLI and initializing gateway credentials through the terminal.

npx vercel ai-gateway setup

Once installed, coding tools such as Claude Code, Cursor, and Codex can target the model directly through standard configuration interfaces. To switch models dynamically inside interactive agent sessions, select the designated identifier:

/model reflection/beam

To prevent billing surprises once the launch promotional window expires, select the explicitly capped free tier identifier instead:

/model reflection/beam-free

CLI agents supporting runtime arguments can launch pinned directly to the model instance:

claude --model reflection/beam

Applications integrating directly via the Vercel AI SDK invoke the stream through standard completion calls:

```javascript
import { streamText } from 'ai';

The standard slug reflection/beam automatically transfers over to standard usage pricing with no service downtime once the launch period concludes, whereas the reflection/beam-free slug immediately terminates requests when the promotion finishes.

What to watch after Reflection AI's Beam now available on AI

Teams deploying the standard reflection/beam identifier must actively audit usage metrics and spend caps within the AI Gateway dashboard. Because reflection/beam shifts automatically into metered billing at the conclusion of the promotional phase, unmonitored production pipelines risk unexpected invoice charges if budgets are not configured in advance. Setting hard API key limits avoids surprises during the cutover window.

Developers who cannot absorb immediate inference costs should route critical tasks through reflection/beam-free while monitoring Reflection AI announcements regarding commercial pricing schedules. Because reflection/beam-free stops responding entirely when the free promotional phase ends, production environments running that identifier require fallback routing rules configured inside AI Gateway to maintain continuous application availability.