Inkling 2 Small costs $0.26 per agent task, 4.7× cheaper than leader Kimi K3.1 Pro
Thinking Machines has released the open weights for Inkling 2 Small. On Agent Arena, a leaderboard that compares agents on real user work sessions, the model ranks second among open models. A typical agent task costs $0.26 (median) on it, compared with $1.22 for Kimi K3.1 Pro, the leader of the open list, which makes it 4.7 times more expensive. If your agent handles thousands of tasks a day, run the model on your own cases and compare price and success rate.
AI-processed from X: @arena; edited by Hamidun News
If you run an AI agent in production, every task it performs, whether it's triaging a ticket, answering a customer or editing code, costs money. Agent Arena has published figures for Thinking Machines' open model Inkling 2 Small: a typical (median) task costs $0.26, while Kimi K3.1 Pro, the leader among open models, costs $1.22. That is 4.7 times more. After its open weights were released, Inkling 2 Small took second place among all open models on Agent Arena and first among American ones.
How Agent Arena evaluates agents
Agent Arena is not a classic blind vote where people pick the more pleasing answer. The platform uses real usage traces: what the agent did in a work session, whether it helped the user, and what it cost. Arena itself stresses that it wants to move away from human-preference scores and rely on objective signals from live usage collected by the global community.
The Inkling 2 Small numbers are as follows:
- net improvement score is +2.9% across 11.5 thousand real agent sessions, which places it 22nd in the overall ranking;
- steerability is +10.9%, the best result among all open models;
- Inkling 2 Small remains the only American model in the top ten open models on Agent Arena.
The summary gives no detailed definition of steerability, but in essence it is the model's ability to adapt to a user's clarifications and corrections during a task. This matters for agents: they don't always succeed on the first try, and it is valuable when a short remark can put them back on track.
Money: $0.26 versus $1.22
The most practical figure here is the median cost of a single task. Picture a service where an agent handles tens of thousands of requests a day. At that volume, the gap between 26 cents and $1.22 becomes a significant line in the budget. If quality on your scenario is no worse than the expensive alternative, the savings cost you almost nothing.
Caveats. This is a median across Agent Arena tasks, not a guarantee for your product: the cost depends on context length, the number of agent steps and where you host the model. Moreover, Kimi K3.1 Pro remains first among open models, so Inkling 2 Small is cheaper but not yet better.
Code and web development
Thinking Machines also appeared in Code Arena's WebDev category, which evaluates building web applications. The model scored 1647 points. That is first place among American open models, fourth among all open models in the world and 23rd in the overall standings. For teams that hand front-end and layout work to a neural network, this is another reason to consider an open alternative to closed APIs.
What to do with this
First, teams building agents on open weights have another candidate, especially if price and the ability to deploy the model in-house matter. Second, an American model near the top of the open ranking is of interest to companies that, for compliance or procurement reasons, prefer not to depend on Chinese vendors. This is our observation, not a conclusion from Arena.
The steps are simple: take 50–100 real tasks from your product, run them through Inkling 2 Small and through your current model, then compare the share of successful completions, the number of agent steps and the final price. The Arena ranking is useful as a reason to run a test, but the decision should rest on your own sample.
What the Arena summary does not yet include: model size, license and hardware requirements are not specified. Check these details in Thinking Machines' release before building the model into your product architecture.
- Nov 21, 2026FDE job postings grew 4.6x in a year: what a forward deployed engineer actually does at the client's site
- Nov 21, 2026Sakana AI Opens Applied Research Lab for Market and Economic Modeling
- Nov 21, 2026Meta Is Reportedly Preparing the $300 Muse Charm AI Tamagotchi With a Year of Verizon 5G
- Nov 21, 2026Dreamteam Raises $53M for a CRM That Fills Itself from Email and Calls
Key Questions Answered
How Agent Arena evaluates agents?
Agent Arena is not a classic blind vote where people pick the more pleasing answer. The platform uses real usage traces: what the agent did in a work session, whether it helped the user, and what it cost. Arena itself stresses that it wants to move away from human-preference scores and rely on objective signals from live usage collected by the global community.
What are the key details on Money: $0.26 versus $1.22?
The most practical figure here is the median cost of a single task. Picture a service where an agent handles tens of thousands of requests a day. At that volume, the gap between 26 cents and $1.22 becomes a significant line in the budget. If quality on your scenario is no worse than the expensive alternative, the savings cost you almost nothing. Caveats.
What are the key details on Code and web development?
Thinking Machines also appeared in Code Arena's WebDev category, which evaluates building web applications. The model scored 1647 points. That is first place among American open models, fourth among all open models in the world and 23rd in the overall standings. For teams that hand front-end and layout work to a neural network, this is another reason to consider an open alternative to closed APIs.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.