Get started
Research

Research notes from agent developers

How we think about building, deploying, and governing AI agents inside the data workspace. Engineering insights, industry analysis, and original quantitative research from the team shipping Deepnote.

Industry reports

To build a pioneering product, our team conducts industry studies to understand and anticipate the needs of the market. Get access to exclusive reports.

Agent Initiative Benchmark

Which frontier models know when to help, what to say, and when to stay silent? We built an extensive benchmark to test 12 models on everyday business tasks. The most expensive models aren’t always the best choice for always-on agents.

12

frontier
models

50

workplace
situations

8.97

top score,
GPT-6 Astra

Read the report

Research

Agent Initiative Benchmark

Proactive help in a synthetic company with 6 years of history and a rich data footprint across 11 services.

01 | Leaderboard

Two models at the top

Astra 6 8.97 | Sol 6.1 8.65 | Fable 5.2 8.17

GPT-6 Astra: 8.97 / 10GPT-6.1 Sol: 8.65 / 10Claude Fable 5.2: 8.17 / 10Claude Opus 5.5: 7.84 / 10Claude Fable 5.1: 7.50 / 10Claude Sonnet 5.5: 6.65 / 10GPT-6 Luna: 6.38 / 10DeepSeek V4.1 Flash: 6.35 / 10Mistral Large 4: 6.09 / 10Claude Haiku 5.5: 6.05 / 10Gemini 4 Argon: 5.60 / 10Grok 4.7: 4.51 / 10010

02 | Cost

A 430× price range

$0.007 to $3.02 per task. The three priciest models finish third to fifth.

GPT-6 Astra: 8.97 / 10, $0.693 per taskGPT-6.1 Sol: 8.65 / 10, $0.112 per taskClaude Fable 5.2: 8.17 / 10, $1.87 per taskClaude Opus 5.5: 7.84 / 10, $1.28 per taskClaude Fable 5.1: 7.50 / 10, $3.02 per taskClaude Sonnet 5.5: 6.65 / 10, $0.420 per taskGPT-6 Luna: 6.38 / 10, $0.0070 per taskDeepSeek V4.1 Flash: 6.35 / 10, $0.026 per taskMistral Large 4: 6.09 / 10, $0.082 per taskClaude Haiku 5.5: 6.05 / 10, $0.075 per taskGemini 4 Argon: 5.60 / 10, $0.403 per taskGrok 4.7: 4.51 / 10, $0.425 per task$0.01$0.1$1

03 | Restraint

Knowing when not to help

Models struggle more when their help isn’t needed.

Help90%Help: 90.4% across all 12 modelsSilence63%Silence: 62.5% across all 12 models

04 | Tool use

More tool calls, lower scores

Gemini 4 Argon and Grok 4.7 made the most tool calls and finished last.

GPT-6 Astra: 8.97 / 10, 19.3 tool calls per taskGPT-6.1 Sol: 8.65 / 10, 18 tool calls per taskClaude Fable 5.2: 8.17 / 10, 20.9 tool calls per taskClaude Opus 5.5: 7.84 / 10, 10 tool calls per taskClaude Fable 5.1: 7.50 / 10, 14.5 tool calls per taskClaude Sonnet 5.5: 6.65 / 10, 8.9 tool calls per taskGPT-6 Luna: 6.38 / 10, 13.7 tool calls per taskDeepSeek V4.1 Flash: 6.35 / 10, 21.1 tool calls per taskMistral Large 4: 6.09 / 10, 22.5 tool calls per taskClaude Haiku 5.5: 6.05 / 10, 12.1 tool calls per taskGemini 4 Argon: 5.60 / 10, 34.5 tool calls per taskGrok 4.7: 4.51 / 10, 35.5 tool calls per task02040

The state of agent readiness 2026

An evidence-based benchmark of 4,350 companies across 22 industries, scored on what we can verify in their technology, hiring, teams, and deployments, not what they announce. Median score: 37.7. Only 1 in 27 companies is AI mature.

4,350

companies
benchmarked

22

industries
covered

37.7

median
score

Read the report

Research

The state of agent readiness 2026

Scoring the evidence, technology, hiring, teams, and deployments behind 4,350 companies’ AI claims.

01 | The landscape

Readiness is bottom-heavy

1 in 5 is Operationalizing | 1 in 27 is AI mature.

02 | Talent & controls

People lag the stack

70% of companies are weakest on talent & controls.

03 | Say-do gap

Announcements vs production

Of 3,075 AI announcements, 2.8% report a production deployment.

04 | Company size

Readiness falls mid-market

2,000–10,000 employees is the market’s low point.

The state of data leadership 2026

A quantitative study of ~40K data, analytics, and AI leaders across 140+ countries. Who leads data today, where they came from, where they go next, and how organizations are absorbing agentic AI.

~40K

leaders
analyzed

140+

countries
represented

Read the report

Research

The state of data leadership 2026

Mapping every major flow between ~15k companies and the careers of ~40K data leaders across 140+ countries.

01 | Career velocity

Speed to director by education

This major hits director 2.2× faster than CS/IT grads.

02 | AI era takeover

Leaders with AI/ML in title

10.5 → 34.9% in seven years | +232%.

03 | What drives promotion

Effect on promotion probability

Masters+ +1.4pp | Geo move +0.8pp | Tenure +0.3pp.

04 | Gender across the ladder

Share of women by seniority

Manager 38% → C-suite 23% | monotonic.

Development insights

Sharing our learnings from our forward-deployed engineers who help leading companies building data workspaces for agents & humans safely and scalably.

Your data team is already building a context layer. They just don't call it that.

How separating planning from execution — with cell-level reactivity — lets the agent self-correct without touching production data.


JK

Jakub Jurovych

Apr 28, 2026

11 min read

How data teams are actually building the context layer

The teams making the most progress aren't buying another layer. They're writing down what they know and keeping context close to the data and code.


JK

Jakub Jurovych

Sep 1, 2026

8 min read

Data notebooks as the atomic unit for RL

Why a reactive notebook — with its data, code, and outputs in one place — is the right atomic unit for training and evaluating RL agents.


JK

Jakub Jurovych

Mar 31, 2026

7 min read

AI briefings & outlook

Curated bi-weekly roundup of how AI is reshaping the industry, academic signals, and the latest on AI agents, LLMs, and AI infrastructure.

Read on Substack
Rich ants, LLM-built quantum circuits, and an unlikely fertilizer operator

I hate SF Halloween because what do you mean you’re the two opportunities you had to work at Anthropic, survivorship bias, or Claude reaching the daily limit?

Data Deep Dives

Deepnote's Substack

Cryptographic mystery finally solved, decision models go brrr, FTC & EU vs labs

The 49ers went 6–0, Mercedes and Teslas queued at a San Francisco food bank, and the FBI’s arrest of an alleged OpenAI spy gave “the year of agents” an unexpected second meaning.

Data Deep Dives

Deepnote's Substack

Agents off the leash. Who holds the keys?

It’s been a great fortnight for struggling, second-tier competitors firing on all cylinders.

Data Deep Dives

Deepnote's Substack

Footer

Solutions

  • Python notebook
  • SQL notebook
  • R notebook
  • Data apps
  • Machine learning
  • Data teams

Product

Company

Comparisons

Resources

Tools

Footer

  • Privacy
  • Terms

© 2026 Deepnote. All rights reserved.