Practical AI

AI Tool Selection & Secure Use

We test Copilot, Claude, Codex and the other tools you have or are considering on your real tasks. Then we check what each one can reach before anyone connects it to your mail, files or systems.

Discuss AI Tool Selection
Before connecting an AI assistant to its files, the eagle switches off email and delete permissions and limits reading to the Policies folder, so the answer comes only from Policies.

The best model keeps changing.

Last spring’s choice may already be wrong.

The #1 changed 15 times in 24 months

Today Sep 2026

Anthropic’s Claude Opus 5.5 is #1. A team choosing now would pick Claude.

Coding now: Claude Opus 5.5 in Claude Code (Anthropic), since 23 Sep.

GPT-6 Astra in Codex (OpenAI) led 5 to 23 Sep.

Also new in Sep 2026: Claude Fable 5.1, GPT-6 Sol, GPT-6 Luna, Claude Sonnet 5.5, GPT-6.1 Sol.

Best at coding:Not yet measuredAnthropicGoogleOpenAI

Short leads:Google Gemini 3 Pro, 16 to 26 Nov 2025OpenAI GPT-5.6 Sol in Codex, 12 to 16 Jul 2026Moonshot AI Kimi K3, 16 to 27 Jul 2026Alibaba Qwen3.8 Max, 1 to 2 Sep 2026OpenAI GPT-6 Astra, 5 to 23 Sep 2026

See the data
MonthIndependent tests: #1People’s votes: #1Best at coding
Oct 2024o1-preview OpenAIChatGPT-4o OpenAI, tiedNot yet measured
Nov 2024o1-preview OpenAIGemini Exp 1121 Google, tiedNot yet measured
Dec 2024o1-preview OpenAIGemini Exp 1206 Google, tiedClaude 3.5 Sonnet Anthropic
Jan 2025o1 OpenAIGemini 2.0 Flash Google, tiedClaude 3.5 Sonnet Anthropic
Feb 2025o3-mini OpenAIGrok 3 (early) xAIClaude 3.7 Sonnet Anthropic
Mar 2025o3-mini OpenAIGemini 2.5 Pro GoogleClaude 3.7 Sonnet Anthropic
Apr 2025o4-mini OpenAIGemini 2.5 Pro GoogleClaude 3.7 Sonnet Anthropic
May 2025o4-mini OpenAIGemini 2.5 Pro Google, tiedGemini 2.5 Pro Google
Jun 2025Gemini 2.5 Pro GoogleGemini 2.5 Pro GoogleGemini 2.5 Pro Google
Jul 2025Grok 4 xAIGemini 2.5 Pro Google, tiedGemini 2.5 Pro Google
Aug 2025GPT-5 OpenAIGemini 2.5 Pro Google, tiedGPT-5 OpenAI
Sep 2025GPT-5 OpenAIGemini 2.5 Pro Google, tiedGPT-5 OpenAI
Oct 2025GPT-5 Codex OpenAI, tiedGemini 2.5 Pro Google, tiedGPT-5 OpenAI
Nov 2025Gemini 3 Pro Preview GoogleGemini 3 Pro Google, tiedClaude Opus 4.5 AnthropicGemini 3 Pro (Google) 16 to 26 Nov
Dec 2025Gemini 3 Pro Preview GoogleGemini 3 Pro Google, tiedClaude Opus 4.5 Anthropic
Jan 2026GPT-5.2 OpenAIGemini 3 Pro GoogleClaude Opus 4.5 Anthropic
Feb 2026Gemini 3.1 Pro Preview GoogleClaude Opus 4.6 Anthropic, tiedClaude Opus 4.6 Anthropic
Mar 2026Gemini 3.1 Pro Preview Google, tiedClaude Opus 4.6 Anthropic, tiedClaude Opus 4.6 Anthropic
Apr 2026GPT-5.5 OpenAIClaude Opus 4.7 Anthropic, tiedClaude Opus 4.7 Anthropic
May 2026Claude Opus 4.8 AnthropicClaude Opus 4.6 Anthropic, tiedClaude Opus 4.7 Anthropic
Jun 2026Claude Fable 5 AnthropicClaude Fable 5 Anthropic, tiedClaude Fable 5 Anthropic
Jul 2026Claude Opus 5 AnthropicClaude Fable 5 Anthropic, tiedClaude Opus 5 AnthropicGPT-5.6 Sol in Codex (OpenAI) 12 to 16 JulKimi K3 (Moonshot AI) 16 to 27 Jul
Aug 2026Claude Opus 5 AnthropicClaude Fable 5 Anthropic, tiedClaude Opus 5 Anthropic
Sep 2026Claude Opus 5.5 AnthropicClaude Opus 5.5 Anthropic, tiedClaude Opus 5.5 AnthropicQwen3.8 Max (Alibaba) 1 to 2 SepGPT-6 Astra (OpenAI) 5 to 23 Sep
  • A company’s best-ranked model is not always its newest.
  • Independent tests: the Artificial Analysis Intelligence Index. Independent: Artificial Analysis runs every evaluation itself rather than relying on vendor-reported numbers, and publishes its methodology.
  • People’s votes: the LMArena / Arena text leaderboard (style-controlled default view). Measures preference in chat, not task success; top models are usually statistically tied.
  • Each line is a company’s best model at each month end, placed by its rank on that leaderboard. Ranks, not scores, because both leaderboards changed their scoring during these two years.
  • Best at coding: the #1 on LMArena’s WebDev Arena, then its Code Arena from November 2025.
  • Chosen for credibility and relevance, not for how often the top changes: the #1 model changed 15 times on Artificial Analysis (8 changes of company) and 11 times on LMArena (4).
Sources (59)

Leaderboards

Vendor sources

Same task, different tools.

None wins every task. Pick a task.

Draw our office network as a diagram for the board pack.

Claude

Every link and VLAN labelled.

Copilot

Slide-ready, but the VLANs are missing.

Gemini

Right order; VLANs merged into one box.

Codex

A Python script it ran to render office.png: a plain diagram with no VLAN labels.

network.py

  1. from diagrams import Diagram
  2. from diagrams.generic import network
  4. with Diagram("Office"):
  5. fw = network.Firewall("fw")
  6. fw >> network.Switch("core")

Rendered office.png

Rendered, but plain; no VLAN labels.

Best fit for this task: Claude

Illustrative, not a benchmark. Copilot means Microsoft 365 Copilot.

Same model, different tool.

The app or agent around a model changes what it can do.

Vendor’s own tool:OpenAIAnthropicGoogleTest agent (the benchmark’s own, neutral)

  • Claude Code also cost about half as much to run.

    Test: CORE-Bench Hard. Test agent: CORE-Agent.

    Source: HAL: CORE-Bench Hard Leaderboard, Princeton HAL, September 2026

Effects run both ways: a vendor’s own tool sometimes wins and sometimes trails.

Sources (28)

Leaderboards

Vendor sources

Research

What a connection can see.

Switch the scope and watch what it reaches.

Connection permissions

Assistant would like to:

  • Read, edit and delete all your files
  • Read and send email as you
  • See and change your calendar

5 of 5 in reach, including payroll and your inbox

  • PoliciesFull access
  • HRFull access
  • FinanceFull access
  • ClientsFull access
  • MailboxFull access

A decision your team can follow.

Where the review lands. Open a row.

AI tool decision matrixSample, illustrative
  • Email and meeting follow-ups, use Copilot
    Why
    Already has the thread and the Teams notes
    Watch for
    It sees everything your account can open
    Human verification
    Check names, dates and commitments
  • Spreadsheets and long documents, use Claude
    Why
    Works on the file directly and cites pages
    Watch for
    Upload and retention settings
    Human verification
    Recalculate the headline figures
  • Scripts and code changes, use Codex
    Why
    Works in the repository and opens a pull request
    Watch for
    Access to production secrets
    Human verification
    Review the diff and run the tests
  • Finding policy answers, use Your assistant, scoped
    Why
    Answers with the source named
    Watch for
    Keep it to Policies, read only
    Human verification
    Open the source document

Yours is built from your tasks, and dated.

How the review runs.

  1. Map what you have

    Tools, licences and connections, official or not.

  2. Agree the tasks

    Three to five real tasks, and how we judge them.

  3. Test and inspect

    Each tool on each task, and what it can reach.

  4. Recommend

    Which tool for which task, and which connections are safe.

What you receive

  • Dated tool comparison
  • Decision matrix
  • Permissions review for each connection
  • Short user guidance

Why have a security team do this?

“What can this tool see once we connect it?”

Connecting an assistant hands it your permissions.

How we do it

We test what it can reach as a real user and scope each connection to the task.

  • Data questions firstWhat could leak is settled before switch-on.
  • Access set up properlyWe configure and test it ourselves.
  • Verified with your peopleSessions in plain English, not slide decks.

Related Practical AI services

All Practical AI services
Choosing or connecting a tool?

Check what it can reach before anyone clicks Allow.

Discuss AI Tool Selection