Standard Compute
Model comparison 2026

Beam 1.0 vs Step 5 Preview

Beam 1.0
Reflection · Balanced · Open weights
$1.60/1M · 1.0M ctx
Step 5 Preview
StepFun · Balanced · Proprietary
$1.51/1M · 1M ctx

Community vote

Pick a winner in each dimension. You can change your vote any time.

Output Quality

Correctness and depth of what it produces

Agentic Ability

Tool calls, instruction-following, and multi-step tasks

Speed

Tokens per second and time-to-first-token

Value for Money

How much capability you get per dollar

Reliability

Consistent results: fewer refusals, loops and format breaks

Specs and pricing side by side

Spec
Beam 1.0
Step 5
Maker
Reflection
StepFun
Blended price / 1M
$1.60
$1.51
Input / output
$1.00 in · $3.00 out / 1M tokens
$1.00 in · $2.70 out / 1M tokens
Context window
1.0M
1M
Open weights
Yes
No
Tool use
Yes
Yes
Reasoning
Yes
Yes
Output Quality
7.0
8.2
Agentic Ability
5.0
7.6
Speed
7.0
7.5
Value for Money
5.7
7.1
Reliability
8.0
8.6

Pricing and capabilities synced from the OpenRouter catalogue. Scores are editorial (0–10).

Verdict: Beam 1.0 or Step 5?

Updated 2026-10-17

Choose Beam 1.0 if you want trying a permissively licensed open-weight coding model through OpenRouter, or self-hosting it where code must stay on your own hardware. Choose Step 5 Preview if you want mid-priced agent loops that want more capability than a Flash-class model.

In our editorial scoring, Step 5 Preview leads in 5 of five dimensions (output quality, agentic ability, speed, value for money and reliability), while Beam 1.0 leads in 0. On price, Beam 1.0 runs about $1.60 per 1M tokens (blended) and is open-weight; Step 5 Preview is about $1.51 and proprietary.

Where Beam 1.0 falls short
  • Every published benchmark figure is Reflection's own so far
  • Text only: no image input
Full Beam 1.0 breakdown →
Where Step 5 falls short
  • Still a preview, so behavior and price can change
  • Only StepFun serves it on OpenRouter, so there is no provider fallback
Full Step 5 breakdown →

The agent matters as much as the model. Compare AI agents →

Current Claude, GPT, Gemini and open models run on Standard Compute

Standard Compute

Current frontier models Claude Fable 5.1 · GPT-6 Astra

More intelligence
per dollar.

Each step goes to the cheapest model that can do it, on providers we vet for price, uptime and speed, with automatic fallback when one fails.

Compliance
Opt-in rules keep a company, team or key on European or US providers
Privacy
We don’t train on your prompts or store them

Fixed monthly plans from $19 · cancel anytime

7-day fair refund Free trial, no card Billing by Stripe3,500+ agent users

Related comparisons