◆ THE LAB · Models
Reflection's Beam is faster than every model on our table
The News
On October 16, 2026, Reflection released the weights of Beam, its first open-weight model, built for coding, reasoning and agentic workloads, under an Apache 2.0 licence. Reflection's own API, in public beta, is “free to start,” it says, and its docs list no price. On OpenRouter, host Nebius charges $0.25 per million input tokens and $1.00 per million output.
Chandni Melwani
Founder & Editor
Oct 23, 2026 · 1 MIN READ
Know More
- — Size: 501 billion parameters, 23 billion active, in a sparse mixture-of-experts design with 256 experts.
- — Context: 1 million tokens, with up to 128K output tokens per request on Reflection's API.
- — Training: over 100 million reinforcement-learning rollouts on 10.5K Nvidia GB300 GPUs over four weeks, by Reflection's account.
- — Other OpenRouter hosts, per million tokens: CoreWeave $0.30 in, $1.00 out; BaseTen, DigitalOcean and Together $0.40 and $1.20.
On October 22, Artificial Analysis measured Beam at 327 tokens a second, a median across its hosts. The quickest model on New in AI’s comparison table that day was Meta’s Muse Spark 1.3 at xhigh, at 285 on Meta’s API. On LiveBench’s reasoning suite, which the table also uses, Beam scores 80.2, below all nine table models. Meta’s open-weight Muse Spark Open, which isn’t on the table, scores 73.7.
Vals.ai’s Terminal-Bench 4.0, the table’s coding measure, lists Beam 1.0 46th of 54 at 4.04% as of October 22, with its cost shown as “N/A” and a run time of 19 minutes 15 seconds. Every model ranked above it took at least 29 minutes. Reflection reports 80.1% on Terminal Bench v2.1, but Vals says 4.0’s 66 tasks are all new and “scores are not comparable across versions.”
Reflection’s own launch table has DeepSeek V4.1 Flash ahead of Beam on seven of the nine benchmarks where it gives both scores.
The Read
Reflection pitches Beam on efficiency: “Where frontier open models like Kimi K3 remain ahead on raw capability, Beam's advantage is efficiency at inference time.” Under Apache 2.0 anyone can serve it, and five OpenRouter hosts already charge $0.40 or less per million input tokens, so Reflection's own price, if it sets one, will probably sit close to theirs.
Room for Disagreement
Open weights haven't pulled Moonshot down to the cheapest hosts' prices. On OpenRouter it charges $3 in and $15 out per million tokens for Kimi K3, also open-weight, the same as seven other hosts, while Decart charges $1.65 and $8.25 and Sail Research $1.20 and $5.50.
Related
Frequently Asked Questions
How does Beam's price compare with DeepSeek V4.1 Flash's?
Nebius's OpenRouter rate for Beam is below DeepSeek's own peak-hours price for its Flash model, now DeepSeek-V4.1-Flash-1017: $0.30 per million cache-miss input tokens and $1.20 per million output. It is above DeepSeek's off-peak price, which is half the peak rate. Peak hours are 01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese public holidays. New in AI's comparison table uses each lab's own standard-tier price; Beam's figure here is a host's.
How does Beam score on Artificial Analysis's Intelligence Index?
Artificial Analysis gives Beam 1.0 an Intelligence Index of 30. DeepSeek V4.1 Flash 1017 at max and Qwen3.8 Max (0902) score 45, and Kimi K3 at max scores 44. Beam's cost per Index task, $0.23, is the lowest of the four: DeepSeek's is $0.26, Kimi K3's $2.00 and Qwen3.8 Max's $5.41.
Can Beam be fine-tuned, distilled or run on-premises?
Reflection's model card says it “does not prohibit distillation, use of outputs to train other models, fine-tuning,” among other uses, and Reflection publishes the weights in BF16, FP8 and NVFP4 with a speculative-decoding drafter model. For on-premises, sovereign and air-gapped environments, Reflection says, Dell and HPE deliver Beam “behind the firewall.”
Sources
- Reflection, “Beam is now available” (October 16, 2026): the weights released on Hugging Face under Apache 2.0, Beam as Reflection's first open-weight model, the BF16, FP8 and NVFP4 formats and drafter model, the public beta that is “free to start,” and Dell and HPE on-premises
- Reflection, “Introducing Beam: Reflection's 501B open-weight model” (October 5, 2026; results updated October 8): the launch benchmark table with DeepSeek V4.1 Flash, the RL run on 10.5K NVIDIA GB300 GPUs, and the line on Kimi K3 and efficiency at inference time
- Reflection, Beam model card (checked October 22, 2026): 501B total and 23B active parameters, 256 experts, a 1M-token context, 80.1% on Terminal Bench v2.1 at BF16, and the uses Reflection does not prohibit
- Hugging Face, reflection/Beam-1.0-501B-A23B (README and LICENSE checked October 22, 2026): the weights and the Apache 2.0 licence
- Reflection developer docs, “Models” (checked October 22, 2026): Beam's 1M context window and 128K maximum output
- Reflection developer docs index (checked October 22, 2026): every documentation page, none of them a pricing page
- OpenRouter, Reflection: Beam 1.0 501B A23B endpoints (checked October 22, 2026): Nebius, CoreWeave, DigitalOcean, BaseTen and Together prices per million tokens
- OpenRouter, MoonshotAI: Kimi K3 endpoints (checked October 22, 2026): Moonshot AI and seven other hosts at $3 and $15 per million tokens, Decart at $1.65 and $8.25, Sail Research at $1.20 and $5.50
- Artificial Analysis, Beam 1.0 (checked October 22, 2026): 326.7 output tokens a second, based on the median across providers serving the model
- Artificial Analysis, Muse Spark 1.3 (Xhigh) (checked October 22, 2026): 285.4 output tokens a second, based on Meta's API
- Artificial Analysis, Beam 1.0 providers (checked October 22, 2026): five providers benchmarked, Nebius (FP8) the slowest at 185.3 tokens a second
- Artificial Analysis, LLM leaderboard (checked October 22, 2026): Beam 1.0's Intelligence Index of 30 and $0.23 per task, with DeepSeek V4.1 Flash 1017, Qwen3.8 Max (0902) and Kimi K3 for comparison
- LiveBench 2.0, October 14, 2026 results table (checked October 22, 2026): reasoning scores for Beam 1.0, Muse Spark Open and the nine models on New in AI's comparison table
- Vals.ai, “Terminal-Bench 4.0” leaderboard (updated October 22, 2026; checked October 22): Beam 1.0 46th of 54 at 4.04% with cost N/A and a 19m15s duration, the durations of the models ranked above it, and the note that scores are not comparable across versions
- New in AI, model comparison table (updated October 8, 2026; checked October 22): the nine models and the leaderboards it uses for coding, reasoning and speed
- DeepSeek API docs, “Models & Pricing” (checked October 22, 2026): deepseek-flash, model version DeepSeek-V4.1-Flash-1017, its peak-hours cache-miss input and output prices, off-peak at half, and the peak-hour windows
- Hero photo: Nvidia HGX B200 8-GPU board, by Pokiiri, Wikimedia Commons, licensed CC BY-SA 4.0 (cropped)
- Creative Commons Attribution-ShareAlike 4.0 International licence, which applies to the hero photo
Chandni Melwani
Chandni Melwani is the founder and editor of New in AI, covering AI agents, M&A, and enterprise adoption. She holds a Master's in Management of Artificial Intelligence from Queen's University and brings a practitioner's perspective from her work in Data and AI leadership.
Follow ↗The Lab Briefing
Get the five-beat briefing in your inbox.
The big story, frontier labs, AI at work, e-commerce and the capital crossing borders, each with the part that tells you why it matters, not just that it happened.
Free · No spam · Unsubscribe in one click
More from The Lab
Models
Moonshot's Kimi K3.1 Pro ranks third on Vals' coding test
Moonshot AI's Kimi K3.1 Pro scores 63.64% on Vals.ai's Terminal-Bench 4.0, third of 57 models, at $4.57 a test; Kimi K3 scores 17.17% at $12.48. Moonshot charges a third less for standard K3.1 Pro than for K3, and it promises the weights by November 4.
Chandni Melwani · China · Oct 29, 2026
Research
Anthropic's Claude Mythos 5 designs cheaper quantum circuits
Anthropic says Claude Mythos 5, working in a loop with an evaluator program, found more efficient quantum circuits. In seven of eight tasks, including chemistry and optimization, the accompanying paper reports both fewer Toffoli gates and fewer qubits than the baseline it compares each against.
Chandni Melwani · Global · Oct 22, 2026
Labs
Anthropic is turning away distillers' paying traffic
Anthropic says early evidence suggests its new defences have cut the data harvested by distillers, who covertly extract Claude's capabilities to replicate them in other models, by around 90 to 99%, and that it has turned away meaningful revenue and trillions of tokens' worth of paying traffic from distillers.
Chandni Melwani · Global · Oct 21, 2026