Summary: Deepnote's Agent Initiative Benchmark tests something most model leaderboards skip: whether an AI teammate in Slack can tell when to step in and when to say nothing. Across 12 models and 50 workplace situations, GPT-6 Astra (8.97) and GPT-6.1 Sol (8.65) lead, Claude Fable 5.2 is third at 8.17, and Sol does it for about one sixth of Astra's cost per task. The bigger lesson for anyone building a proactive agent: almost every model is good at helping and noticeably worse at staying quiet.
What the benchmark measures
Most agent benchmarks hand the model a task and grade the answer. This one hands the model a company and waits. The agent sits inside the Slack workspace of PageMantle, a synthetic software company with 71 employees and 100 enterprise customers, and receives one standing instruction: assess what is going on and help if it is useful. Nobody tags it.
From there it has three jobs. It has to dig up evidence in systems the conversation never mentions, decide whether a message is needed at all, and, if so, pick the right audience: a channel post, a thread reply or a direct message. In 10 of the 50 situations the correct move is silence, because the issue has already been handled and the records say so.
Deepnote's example makes the idea concrete. Sales says in one channel that a launch is on Friday; engineering says in another that the release slipped a week. A useful agent spots the mismatch without being asked, finds the release update and brings it into the sales thread.
The setup in numbers
- 11 simulated services: Slack, Linear, Notion, HubSpot, Stripe, Google Calendar, PostHog, Google Analytics, Google Search Console, BigQuery and Snowflake, each behind an API shaped like the real one.
- About 10 GB of data covering six years of company history, with one live month of Slack activity (roughly 9k messages and 2.5k threads).
- 50 tasks, 8 task types: Find, Explain, Analyze, Verify, Decide, Diagnose, Coordinate and Guard.
- Same conditions for every model: identical instructions and tools, high reasoning effort, and a budget of 180 seconds of model response time (tool execution doesn't count). A run that hits the limit without sending a needed message scores as a miss.
- Grading: two LLM judges, blind to which model they are scoring, rate each run 0 to 10 against a reference solution on four things: the decision to intervene, content and evidence, clarity and tone, and audience and location.
The headline results
Here is how the field lines up on the dimensions that matter when you pay for an always-on agent. All figures come from Deepnote's results page; the "month of Slack" column is Deepnote's own estimate of one investigation per thread across about 2,500 threads.
| Model | Score (0-10) | Helped when needed | Stayed silent when right | Cost per task | Avg. response time | Month of Slack |
|---|---|---|---|---|---|---|
| GPT-6 Astra | 8.97 | 40/40 | 8/10 | $0.693 | 54 s | $1,732 |
| GPT-6.1 Sol | 8.65 | 39/40 | 8/10 | $0.112 | 54 s | $281 |
| Claude Fable 5.2 | 8.17 | 40/40 | 6/10 | $1.87 | 1 min 0 s | $4,684 |
| Claude Opus 5.5 | 7.84 | 40/40 | 8/10 | $1.28 | 29 s | $3,196 |
| GPT-6 Luna | 6.38 | 38/40 | 4/10 | $0.0070 | 50 s | $18 |
Three things stand out once you read past the ranking.
1. Astra and Sol are statistically tied
Sol's 8.65 sits inside Astra's 95% interval (8.30 to 9.52), so Deepnote itself says this task set does not separate them. If you are choosing between the two, the deciding factor is price: $0.112 versus $0.693 per task.
2. The priciest models are not the best ones
The three most expensive models per task (Fable 5.1 at $3.02, Fable 5.2 at $1.87 and Opus 5.5 at $1.28) finish third, fourth and fifth. GPT-6 Luna and Claude Fable 5.1 are about 430 times apart in price and only 1.1 points apart in score.
3. More tool calls did not mean better answers
Gemini 4 Argon averaged 34.5 tool calls per task and Grok 4.7 averaged 35.5, the most in the field, yet they scored lowest (5.60 and 4.51). Both kept investigating until the clock ran out: 27 and 18 of their 50 runs hit the time limit before any message went out, and they scored zero on 17 and 20 tasks. Meanwhile Opus 5.5 answered in about 29 seconds with 10 tool calls.
Why "knowing when not to act" is the real test
When help was needed, models sent a message about 9 times out of 10, and nearly every miss was a timeout rather than a deliberate choice. When silence was right, they held back only about 6 times out of 10. Even the best (Astra, Sol and Opus 5.5) still spoke up in 2 of the 10 settled situations. GPT-6 Luna, DeepSeek V4.1 Flash and Mistral Large 4 spoke in 6 of 10.
The uncomfortable detail is that those unnecessary messages were accurate and well placed. The matter was simply already closed. So a filter that checks message quality won't catch them; you need a separate check for "is this already handled?". Deepnote also found that speaking out of turn cost about 1.8 times as much as staying quiet for the same model, and took about 1.8 times as long, so false alarms hurt both your team's attention and your bill.
Gemini 4 Argon is a useful warning about reading one number in isolation. It stayed silent correctly in 10 of 10 cases, the best in the field, but it sent a message in only 23 of the 40 situations that needed one. Perfect restraint that comes from running out of time is not judgment.
Strengths by task type
Rankings shift by task type, and each type has only 5 to 8 tasks, so treat these as tendencies to test on your own work rather than fixed rules:
- Analyze is the hard one. Pulling numbers from PostHog, BigQuery or Snowflake and checking them against a claim in Slack is the weakest type for 8 of 12 models (field average 5.17). Astra (8.44) and Sol (8.38) are the clear exceptions.
- Verify is the easy one. Checking a claim against its sources averages 8.00 across all models.
- Astra leads five of eight types, including Coordinate (9.33) and Guard (9.83).
- Sol leads Explain (9.08) and Find (8.60) but is about 3 points behind Astra on Coordinate.
- Fable 5.2 has the top Decide score (9.58), with Sol (9.50) and Astra (9.42) close behind.
How to choose a model for your own proactive agent
Putting the numbers together, this is how we would approach the choice. It is our reading of Deepnote's data, not a recommendation from Deepnote, and the benchmark tested models, not routers.
- Start from your task mix. List what people would actually ask the agent for. A support team that mostly needs "who owns this?" (Find) has different needs from a finance team reconciling numbers (Analyze).
- Use a cheap capable model as the default watcher. Sol gives you near-top scores at $281 for a month of Slack in Deepnote's estimate. Luna and DeepSeek V4.1 Flash are far cheaper but held back in only 4 of 10 silence cases, so use them for triage, not the final decision to post.
- Escalate the hard cases. Numbers to reconcile, a decision to weigh or a thread that may already be closed are where Astra's extra cost pays off.
- Pick speed for live threads, depth for background work. Someone waiting in a thread benefits from Opus 5.5's roughly 29-second replies; a nightly digest can afford Sol's 54 seconds and deeper digging.
- Cap the search. Give your harness a tool-call or time budget and a point where it must commit to a message or to silence.
- Track false alarms as their own metric, separately from answer quality.
Run it against your own model
The code is on GitHub under Apache 2.0 and the PageMantle dataset is on Hugging Face under CC BY 4.0. According to the repository's quickstart, you need Python 3.13 on Linux or macOS, plus a C++ compiler for the hnswlib dependency. The flow is:
- Clone the repository, create a Python 3.13 virtual environment and install from
requirements.lock, then install the package in editable mode. - Run
benchmark init. It createsbenchmark.json, where you set one agent model and two different judge models. Each profile names an API format (OpenAI Responses, OpenAI-compatible Chat Completions, Anthropic Messages, or a custom Python adapter), an endpoint and a model ID. The model must support function tools. - Run
benchmark data download, thenbenchmark check, which validates settings without calling any model. - Run
benchmark run --task task-01for one example, or--task allfor all eight bundled examples (one per task type), andbenchmark resultsto see the scores.
The docs also note that changing the model, API protocol or generation settings counts as a different experiment, so keep everything fixed when you compare.
Editor's note
- Who it's for: teams building or buying a Slack-style agent that watches conversations and decides on its own when to chime in.
- Not for: picking a coding model or a chat assistant you prompt directly. This benchmark doesn't measure either.
- Gotcha 1, one effort level: every model ran at high reasoning effort. Deepnote says it has not run lower or higher effort sweeps, and both score and cost would move.
- Gotcha 2, costs are historical: the results page says recorded costs are observations, not live price quotes. Recompute with current prices before you budget.
- Gotcha 3, small sample per type: 50 tasks overall, 5 to 8 per type. The overall ranking is more reliable than any single type.
- Gotcha 4, Slack only: agents reply only through Slack in this version; email, documents and actions beyond messaging are out of scope.
- Common setup error: the install fails building hnswlib if Python 3.13 development headers or a C++ compiler are missing. The quickstart points to
xcode-select --installon macOS andbuild-essentialplus Python 3.13 headers on Ubuntu. - Common config error: using the same model ID for both judges. The docs say the two judge model IDs must differ.
FAQ
Which model won the Agent Initiative Benchmark?
GPT-6 Astra had the highest mean score at 8.97, but GPT-6.1 Sol (8.65) falls inside Astra's confidence interval, so Deepnote does not treat them as separated. Claude Fable 5.2 is third at 8.17.
What does "staying silent" mean in this benchmark?
In 10 of the 50 situations the problem had already been handled and the records showed it. The correct action was to send nothing. Models were judged on whether they held back, not just on whether their message was good.
Can I test my own agent or a local model?
Yes. The harness accepts OpenAI Responses, OpenAI-compatible Chat Completions and Anthropic Messages endpoints, or a custom Python adapter, and the docs describe a setting for a local server without authentication. The model must support function tools.
Related on lightmind24: When a Small Decider Beats a Frontier LLM: Strands Decider · Mem0 vs Zep vs Letta: Choosing an AI Agent Memory Layer · AI by Zapier Playbook: Sort Form Leads and Alert Slack
If you are building an agent that lives in your team's chat, test the moments it should stay out of as carefully as the ones where it should help.
Screenshots of Deepnote's results page and our own model-fit table are at the end of this post. Screenshots: deepnote.com/research/benchmark, captured 2026-11-14 (vendor icons removed); model-fit table by lightmind24.
Last verified: 2026-11-14
Sources: Deepnote Research: Agent Initiative Benchmark · GitHub: deepnote/agent-initiative-benchmark · Benchmark quickstart · Benchmark model settings · PageMantle dataset on Hugging Face
lightmind24 is an independent review site, not affiliated with Deepnote, OpenAI, Anthropic or any other lab named here.
댓글
댓글 쓰기