Guozhen AIGlobal AI field notes and model intelligence

Realtime AI News

MachinePulse open-sources Karpo Decision 31B and karpo-browser, taking browser-agent cost to $0.096 per task

MachinePulse AI has open-sourced Karpo Decision 31B, a decision model that finishes first on the 300-model Decision Index with a score of 73.9. The release also includes the karpo-browser agent framework and the KBB benchmark, with 89% success and $0.096 in model cost per task on Browserbase Benchmark v2.

Published

On November 21, MachinePulse AI released Karpo Decision 31B, a decision model built for the small judgments agents make constantly rather than for long-form generation: which button to click next, whether a task is finished, and which of dozens of available tools best fits the request.

The company's city-scanner test, reported by QbitAI, offers a direct comparison. Asked to pick 10 suitable outings from 36 options under a $50 budget and a 5 km commute, Karpo Decision 31B answered in 2.6 seconds; Gemini 3.5 Flash took 14.96 seconds. It got all 36 questions right, runs on a single RTX PRO 6000, and posted a median response time of 94 milliseconds across more than 140,000 requests from a public set.

Its leaderboard result is equally specific. On the Decision Index, a ranking of 300 models, Karpo Decision 31B took first place with a Full score of 73.9, judging up to 255 candidate options in a single question with calibration error down to 2.3%. MachinePulse notes that Karpo is excluded from the ranking and that its own run used all 102 tasks averaged over three attempts, a stricter setup than the 38-task single run applied to other entries.

For developers, the browser-agent stack released alongside the model may matter more. karpo-browser splits the work between System 2 and System 1: a general model understands the request, decomposes the task and prepares the information to be filled in, while Karpo Decision 27B Act, trained specifically on web actions, chooses fields, clicks buttons and decides the next step, with a median decision latency of 931 milliseconds.

The framework is written in TypeScript and can be paired with different open models and decision models. Through MCP, developers can attach the browsing capability to existing agents such as Claude Code and Codex instead of building a browser agent from scratch. An accompanying benchmark, KBB (Karpo BrowserBench), covers 584 real websites and 23,075 web-action steps recorded from live pages.

On the 102 core tasks of Browserbase Benchmark v2, the stack reported 89% average success, 201 seconds of average completion time and $0.096 in model cost per task. By comparison, the leaderboard-topping Claude Fable 5.2 plus Claude Code combination was about 1.9 times slower and cost roughly thirteen times as much in model spend.

About 78% of on-page actions were handled by the decision model while consuming only 14% of the agent's runtime and 13% of model cost. That ratio is the argument for the category: move the high-frequency, low-value judgments off the general model, leave planning and verification to the expensive one, and both latency and the bill fall.

Context matters here. Decision models became a hot Silicon Valley category after Jev went viral in September; TypeSafe AI, the company behind it, was reported to reach a $7.5 billion valuation within 24 days, and OpenAI, Microsoft, Amazon and Cloudflare have all moved into the space. Yet Jev and Google's Gemini Decision 1 remain closed, and the open-weight GLiDE does not fully deliver speed and judgment together. Karpo's release bundles weights, a browser framework and an evaluation set, giving teams building their own agents a self-hostable option.

Two things are worth watching: whether the KBB and Browserbase numbers can be independently reproduced, and whether the cost advantage survives long, multi-tool workflows once the stack is wired into Claude Code or Codex through MCP. Open weights are only the starting point; stability inside real workflows will decide whether decision models stay in the stack.

Why it matters

Shipping weights, framework and benchmark together turns decision models from a closed API into self-hostable infrastructure, which could push the cost curve of browser agents down across the board.

MachinePulse AIBrowser AgentOpen Source
Back to realtime news

Nearby Updates

All

11/21, 19:27

HUST and Alibaba Launch a Joint AI Safety Research Center

Huazhong University of Science and Technology and Alibaba have established a joint AI safety research center, according to a Chinese media report. The report gives no details on funding, staffing or specific research direction, leaving the center's operating model to be confirmed.

11/21, 19:40

Tencent Yuanbao and KCL release first agent memory benchmark built from group chats

Tencent Yuanbao and KCL have released an agent memory benchmark built from roughly 80,000 real group-chat messages, which the report describes as the first of its kind. The test targets a harder problem than single-turn recall: tracking who said what, when it changed, and which details still matter.

11/21, 17:40

Check Point discloses ARTEX, AI agent campaigns that launch attacks on their own

Check Point has disclosed a set of AI agent attack campaigns it names ARTEX, in which the attack workflow advances autonomously instead of being driven step by step by a human. The report argues this changes the economics of attacks and undercuts detection rules tuned to human operator tempo, though sample scale and chain details remain unpublished.

11/21, 21:53

Qwen4-Max reaches general availability as early tests stack it against Claude Opus 5.5

Qwen4-Max is now generally available, according to a report from Online Tech Tips, and early testing has placed it alongside Claude Opus 5.5 for comparison. The coverage confirms the model's move out of preview, but does not yet disclose the benchmarks, test sets or results behind that comparison.