model comparison

vs

Beam vs Kimi K3

Reflection against Moonshot AI, compared on listed price, real workload cost, context and capabilities.

the short answer

  • on Terminal Bench 2.1, as each maker reports it, Beam scores 80.1% and Kimi K3 88.3.
  • Beam lists lower input and output prices.
  • Kimi K3 supports vision; Beam does not.
  • Beam reads cached input for less ($0.015/M against Kimi K3's $0.3/M), which decides the bill for agents that resend long context.

side-by-side specs

specBeamKimi K3
providerReflectionMoonshot AI
input price$0.125/M$3/M
output price$0.4/M$15/M
cached input$0.015/M$0.3/M
blended price (3:1)$0.194/M$6/M
context window1M tokens1M tokens
max output128k tokens—
inputtexttext, images
structured outputyesyes
weightsopen (Apache 2.0)open (Kimi K3 License)
releasedOctober 2026July 2026
Terminal Bench 2.1 (maker-reported)80.1%88.3
GPQA Diamond (maker-reported)90.6%93.5

same work, both models

workloadBeamKimi K3
one chat reply3,000 tokens in, 500 outKimi K3 costs 29× as much as Beam$0.00058$0.02
summarise a 100-page report50,000 tokens in, 1,500 out, one requestKimi K3 costs 25× as much as Beam$0.0069$0.17
1,000 classification calls500 tokens in, 20 out, per callKimi K3 costs 26× as much as Beam$0.07$1.80
coding-agent session2M tokens in across ~60 turns, 60,000 out, no cacheKimi K3 costs 25× as much as Beam$0.27$6.90
same session, cached85% of input read from the prompt cache, the rest written to itKimi K3 costs 27× as much as Beam$0.09$2.31

at listed rates. token counts are typical sizes, not measurements.

Kimi K3 pricing notes

  • not included in any plan: every request bills from credits.
  • context caching is automatic: cache hits bill at $0.30 per 1M input tokens, a tenth of the input rate.
  • thinking is on by default at max effort, and reasoning tokens bill at the output rate.

when to pick which

Beam

the first open-weights model from Reflection, a 501B-total, 23B-active mixture-of-experts built for agentic tool use, coding and general reasoning. it reads text only and its weights ship under Apache 2.0.

pick it if

  • you want a long-window text model priced with the flash tier
  • you want Apache 2.0 weights, with BF16, FP8 and NVFP4 builds to self-host
  • you run tool-calling agents on text and repeat long prompts, so cached input matters

skip it if

  • you need image, audio or video input; Beam reads text only
  • long autonomous coding is the job; its reported DeepSWE v1.1 score sits below GLM-5.3 Flash and DeepSeek V4.1 Flash

Kimi K3

Moonshot AI's most capable Kimi model: an open-weights 2.8T-parameter mixture-of-experts with native vision and a 1M-token context. it sits above Kimi 2.6 and Kimi 2.7 Code, which have a 256K context and cost less per token.

pick it if

  • you want the model Moonshot calls its most capable
  • your prompts outgrow the 256K context of Kimi 2.6 and 2.7 Code
  • you resend long prompts, so cache hits bill at a tenth of input

skip it if

  • Kimi 2.6 or 2.7 Code covers your workload at a lower rate

faq

which is better, Beam or Kimi K3?

on Terminal Bench 2.1, as each maker reports it, Beam scores 80.1% and Kimi K3 88.3. these are each maker's own numbers, not a controlled head-to-head; the spec table lists every benchmark both publish.

which is cheaper, Beam or Kimi K3?

Beam lists lower input and output prices. on a coding-agent session (2M tokens in, 60,000 out) Beam costs $0.27 and Kimi K3 $6.90.

which one accepts images?

Kimi K3 supports vision; Beam does not.

can i switch between Beam and Kimi K3 mid-task?

yes. in an octomind session, /model octohub:beam and /model octohub:kimi-k3 switch models without losing the conversation, so you can draft on one and review on the other.

try either model

switch between models inside the same octomind session.

open the panel all models

prices from the catalog of 2026-11-09