Model usage is the cost that grows with adoption, because the open-source agent itself is free. Shortlist models from the OpenHands Index and test them on your own tasks. Choose between your own provider keys and the at-cost OpenHands provider, and cap spend per conversation and per organization before the pilot widens.
When you need this
- A pilot is choosing its first model and wants a defensible reason.
- Finance wants a monthly figure for agent usage.
- Engineers want to try open-weight or local models to cut cost or keep data in-house.
Evaluation criteria
Shortlist
Which models have published results with OpenHands?
Look forThe OpenHands Index averages five benchmarks per model. Its recommendations, generated June 18, 2026, list Claude Opus 4.8 at 71.9, GPT-5.5 at 65.9, and Gemini 3.5 Flash at 62.6. The strongest open-weight options are GLM-5.1 at 58.2, MiniMax-M3 at 57.2, and Kimi-K2.6 at 57.1. For local use the docs start with Qwen3.6-35B-A3B.
Keys or provider
Who bills the model usage?
Look forYour own key works with any provider LiteLLM supports, so existing enterprise agreements keep applying, and Enterprise deployments always use your own keys. The OpenHands provider charges vendor rates with no markup. Claude Sonnet 4.5 costs US$3 per million input tokens and US$15 per million output tokens. GPT-5 costs US$1.25 per million input tokens and US$10 per million output tokens. OpenHands' docs state that All Hands AI may collect minimal usage metadata when the CLI runs on its provider, so check that against your data rules.
Cost per task
What does a finished task cost on each model?
Look forStart from the Index's cost per SWE-bench Verified instance, then measure your own. All Hands AI's runs cost US$0.75 per instance for Claude Opus 4.8, US$1.52 for GPT-5.5, and US$0.17 for MiniMax-M3. Those are costs on benchmark issues, so record the cost per merged pull request on your own repositories before you budget.
Budgets
What caps spend before it surprises anyone?
Look forSet a maximum conversation budget so OpenHands stops and asks before it continues. In Cloud organizations, add a monthly organization budget, a default for new members, and per-user overrides where a team needs more. Admins get alerts at 80, 90, and 100 percent. Enterprise sets budgets in its LiteLLM gateway.
Routing
Does every task need the most expensive model?
Look forNo. OpenHands advises keeping the strongest model for planning, debugging, and review, and a lower-cost LLM profile for routine edits. An engineer can switch profiles in the middle of a conversation.
Open-weight and local models
Will a smaller model hold up on agent work?
Look forTest tool calls before you choose an open-weight or local model, because OpenHands warns that their tool-use reliability varies widely. Open-weight SWE-bench Verified scores run from 76.4% for MiniMax-M3 down to 34.2% for Nemotron-3-Nano. If responses slow down or tool calls come back as malformed JSON, move to a stronger model or a larger context window.
Questions to ask vendors
- Which model did each merged pull request use, and what did the conversation cost?
- Which teams use their own keys, and which use the OpenHands provider?
- What share of conversations reach their budget before they finish?
- Has a lower-cost profile matched the stronger model on our routine tasks?
Red flags
- No budget per conversation or per organization while usage grows.
- A model chosen from a leaderboard average without a test on your own repositories.
- Open-weight models adopted for cost before anyone measured their tool-use failures.
- Provider keys pasted into personal settings with no record of who spends what.
Products in this piece
Sources
- Official documentation
- 13
- Public source
- 4
- OpenHands docs: LLM overview and model recommendations
- OpenHands docs: LLM settings
- OpenHands docs: OpenHands LLM provider and pricing
- OpenHands docs: Local LLMs
- OpenHands docs: Application settings
- OpenHands docs: Organizations
- OpenHands docs: Organization budgets
- OpenHands docs: OpenHands Enterprise
- OpenHands docs: Enterprise vs. open source
- OpenHands Index
- OpenHands Index results repository
- OpenHands Index: claude-opus-4-8 results
- OpenHands Index: GPT-5.5 results
- M. Chechulin: Open-source agentic coding systems, what can they deliver for US$10 (Sep 1, 2025)
- Daniel Manzke: Burning tokens with AI coding agents (Mar 1, 2025)
- knail1, GitHub issue #9259: a new session hits Anthropic's 20,000 input tokens per minute limit (Jun 20, 2025)
- tbiens, GitHub issue #10039: the CLI with Devstral and Qwen3 Coder locally (Aug 1, 2025)