Sparse mixture-of-experts reasoning model with 23B active parameters for coding and agentic workflows
Beam 1.0 501B A23B by reflection is offered by 9 providers on this page. Public prices are shown for 4 of them. Its official list price is $0 per 1M tokens. The lowest paid channel is Kilo Gateway at $0.25 / $1.00 per 1M, about 4.0× below the list price. That channel is a third-party gateway, so confirm its availability and rate limits before depending on it. It also has 5 free ($0) channels; free tiers usually carry rate limits, and subscription-covered access bills $0 per token only after the subscription fee.
The context window is 1,048,576 tokens at the reference host, but hosts report different limits, from 262,144 to 1,048,576, so the usable window depends on the provider you pick. It supports reasoning, tool use, and structured output. The weights are open, so it can also be self-hosted or served through a gateway of your choice. Providers report the capability flags inconsistently, so verify a specific feature against the host you plan to use. Its training knowledge cuts off at 2026-06-30.
| Provider | Tier | Input | Output | Cache read | Cache write | Context | Output limit | Status |
|---|---|---|---|---|---|---|---|---|
| ReflectionOfficial Beam-1.0-501B-A23B | Gateway | Free | $0 | — | 1,048,576 | 131,072 | ||
| Cloudflare AI Gateway | Cloud | Free | $0 | — | 262,144 ⚠ | 131,072 | ||
| OpenCode Go beam-free | Gateway | Free | $0 | — | 1,048,576 | 131,072 | ||
| OpenCode beam-free | Gateway | Free | $0 | — | 1,048,576 | 131,072 | ||
| Vercel AI Gateway reflection/beam | Cloud | Free | — | — | 1,048,576 | 131,072 | ||
| Kilo Gateway | Gateway | $0.25 | $1.00 | — | — | 1,048,576 | 943,718 | |
| NanoGPT reflection/beam-1.0 | Gateway | $0.45 | $1.25 | $0.04 | — | 1,048,576 | 131,072 | |
| OpenRouter | Gateway | $1.00 | $3.00 | $0.10 | — | 1,048,576 | 943,718 | |
| CoreWeave reflection/Beam-1.0-501B-A23B | Cloud | $1.00 | $3.00 | $0.10 | — | 1,048,576 | 1,048,576 |
Sorted by blended price (input×0.75 + output×0.25) asc. The official channel always shows regardless of rank. Whether a gateway's low price is actually usable can't be verified.
Interleaved thinking (reasoning between tool calls) is declared by 4 of 9 providers.
5 more channels offer $0 (Reflection, Cloudflare AI Gateway, OpenCode Go etc.); free tiers usually have rate limits and no SLA, excluded from ranking.
Price history accumulates from each data sync; currently only 1 sample(s) (2026-10-17). Each future sync adds a point, and once accumulated a line is drawn here.
| Name | Conditions | Score | Metric | Source |
|---|---|---|---|---|
| DeepSWE | variant: BF16 · v1.1 | 43.3 | — | Source ↗ |
| SWE-Bench Pro | variant: BF16 · dataset: hard · v2 | 76.5 | — | Source ↗ |
| SWE-Bench Pro | variant: BF16 · v1 | 66.1 | — | Source ↗ |
| Terminal-Bench | variant: BF16 · v2.1 | 80.1 | — | Source ↗ |
| SWE-Atlas Codebase QnA | variant: BF16 | 31.8 | — | Source ↗ |
| SWE-Bench Multilingual | variant: BF16 | 78.3 | — | Source ↗ |
| SWE-Bench Verified | variant: BF16 | 81.6 | — | Source ↗ |
| AIME 2026 | variant: BF16 | 97.7 | — | Source ↗ |
| Humanity's Last Exam | variant: BF16; no tools | 36 | — | Source ↗ |
| Humanity's Last Exam | variant: BF16; with tools | 51.1 | — | Source ↗ |
| SciCode | variant: BF16 | 49.2 | — | Source ↗ |
| CritPt | variant: BF16 | 16.3 | — | Source ↗ |
| GPQA Diamond | variant: BF16 | 90.6 | — | Source ↗ |
| LiveCodeBench | variant: BF16 | 88.3 | — | Source ↗ |
| AutomationBench | variant: BF16 · dataset: public | 39.3 | — | Source ↗ |
| MCP Atlas | variant: BF16 | 80.8 | — | Source ↗ |
| τ3 Banking | variant: BF16 | 36.1 | — | Source ↗ |
| BrowseComp | variant: BF16; with context management | 76.8 | — | Source ↗ |
| DeepSearchQA | variant: BF16; with context management | 80.5 | — | Source ↗ |
| AA-LCR | variant: BF16 | 78.2 | — | Source ↗ |
| LongBench | variant: BF16 · v2 | 67.5 | — | Source ↗ |
| IFBench | variant: BF16 | 79.6 | — | Source ↗ |
| AA-Omniscience | variant: BF16 · dataset: public split | 10.8 | index | Source ↗ |
The same benchmark scores very differently across harness / dataset, so the qualifying conditions must be shown together.