Best AI coding assistants for defined coding tasks

Editor's note: This is a documentation-led edition. We have not yet run our own hands-on tests of GitHub Copilot, JetBrains AI, Cursor, Devin Desktop, Google Antigravity, Claude Code or OpenAI Codex. What follows rests on official documentation, published pricing and independent studies we read in full, all accessed on 16 October 2026. A tested update will follow.

"AI coding assistant" now covers five jobs: finishing the line you are typing, editing files inside your editor, running commands in your terminal, working alone in a cloud machine, and reviewing pull requests.

The more a tool does alone, the more its defaults matter. So this shortlist is organised by job, with what each vendor documents and charges. Independent research follows, with its limits.

The word defined matters too. These tools are easiest to supervise on a task with a clear finish line: a failing test, a reproducible bug, a refactor covered by tests.

How we chose: the selection criteria

No scores here. A tool made the list when its official pages showed all six of these on 16 October 2026:

1. A defined coding job: completion, agentic editing, terminal work, cloud delegation or review.
2. Where the work runs: what stays on your machine and what goes to vendor servers.
3. Permission controls: approval prompts, sandboxing or limits on what the agent can change.
4. Data terms by plan: whether your code can be used for training, and how to opt out.
5. Published pricing on an official page.
6. On sale to new customers: this rules out Gemini Code Assist, which Google stopped selling on 9 October 2026.

Inclusion means a tool passed these checks, not that we have confirmed it works well.

Job 1: help inside the editor you already use

GitHub Copilot

What it is for: inline suggestions and chat in most major editors, plus agent mode, a CLI, a cloud agent and pull-request review.[1]

How it runs: the extension sends nearby code and open-file context to GitHub's models.[1] The cloud agent edits files and runs tests "in an ephemeral cloud development environment" and can open a pull request.[2]

Permissions: the cloud agent pushes only to its own branch, cannot approve or merge its pull requests, and its GitHub Actions workflows wait for human approval. A firewall restricts its internet access.[3]

Data terms: since 24 April 2026, GitHub says, it may train on Free, Pro, Pro+ and Max interactions, including code snippets, unless users opt out.[1][4] It does not train on Business or Enterprise data.[1]

Price (16 Oct 2026): Free (2,000 inline suggestions a month), Pro $10, Pro+ $39, Max $100 a month. Chat, agents and review draw on "GitHub AI Credits" worth $0.01 each.[1]

"The product is called “Copilot” not “Autopilot” and it’s not intended to generate code without oversight."

— GitHub, GitHub Copilot plans and pricing FAQ (2026).[1]

JetBrains AI (with the Junie agent)

What it is for: completion, chat and the Junie coding agent in JetBrains IDEs, calling cloud models through JetBrains AI or your own OpenAI or Anthropic key.[5][6]

Permissions: by default Junie asks before running bash commands, changing code, file operations or external tools. A "Brave" setting runs everything without asking, which JetBrains warns "can lead to data loss or security issues".[6]

Data terms: JetBrains says full prompts and code sent to models ("detailed code-related data") can be used to train AI models, but it does not collect this from commercial users without explicit consent and keeps it for one year.[7]

Price (16 Oct 2026, individuals): AI Free (3 credits per 30 days), AI Pro $10 and AI Ultimate $30 a month, where one credit equals $1. Organisation seats are $20 and $60.[5]

Job 2: agent-first editors

Cursor

What it is for: an agent-first editor, plus cloud agents and "Bugbot" code review.[8]

How it runs: locally, with requests routed through Cursor's backend even when you use your own API key.[9] Cloud agents "run in isolated VMs in the cloud" and push to a separate branch.[10]

Permissions: the agent may modify workspace files without approval, except configuration files. Cursor's Agent Security page says terminal commands need approval by default, but its Run Modes changelog says the Auto-review classifier "shipped as the recommended default" in version 3.6 on 29 May 2026, so check which mode your install uses.[11][12] Cursor calls its auto-run modes "best-effort guardrails rather than a hard security boundary".[11] Cloud agents have internet access by default, which teams can restrict.[10]

Data terms: with Privacy Mode on, code is not used for training and model providers work under zero-data-retention agreements. With it off, Cursor may store and train on codebase data and prompts.[9] Cursor says Privacy Mode is on by default for team members, including Enterprise teams, and admins can enforce it; the pages we read did not say whether it is on by default for a new individual account.[13]

Price (16 Oct 2026): Free; Plus $8 (Grok models only); Premium $30; Super $100; Ultra $200 a month; Teams from $30 per user. Third-party models cost extra at API list price, and subscriptions are sold through X Subscription Corp with X and Grok features bundled.[8]

Devin Desktop (formerly Windsurf)

What it is for: Cognition's renamed Windsurf editor (since 2 June 2026), pairing a local agent, Devin Local, with Devin Cloud, which runs "on its own machine in the cloud".[14]

Permissions: Devin Local uses allow, ask and deny rules. In its default mode, writes and shell commands need your approval and MCP tool calls prompt first.[15]

Data terms: "By default, we may use your data for model training purposes," Cognition says. Paid plans can opt out, which also turns on zero data retention with model providers; Enterprise data is never used without written consent.[16]

Price (16 Oct 2026): Free; Pro $20 (includes Devin Cloud); Max $200 a month; Teams $80 a month plus $40 per full seat.[17]

Google Antigravity

What it is for: Google's agent app, IDE, CLI and extensions. It replaced Gemini CLI and the Gemini Code Assist extensions for individuals on 18 June 2026; Google stopped selling Gemini Code Assist Standard and Enterprise on 9 October 2026.[18]

Permissions: on macOS and Linux, the default preset runs terminal commands in a sandbox limited to the workspace and temporary folders, with no network, and asks before anything outside it, MCP tools or web pages. A "Turbo" preset removes those limits.[19]

Data terms (individual accounts): Google uses "Interactions" to improve its products and models, and "Google employees and contractors may access, view, review and use Interactions." You can change this in settings.[20] Teams use Google Cloud terms.[21]

Price (16 Oct 2026): free with weekly rate limits; Google AI Pro ($19.99 a month in the US) and Ultra raise limits. Gemini Enterprise seats start at $30 a month.[21][22]

Job 3: an agent in your terminal

Claude Code

What it is for: an agent that reads your repository, edits files and runs commands from the terminal, IDE or desktop app.[23]

How it runs: locally, sending prompts and outputs to Anthropic's API. Optional cloud sessions run in isolated Anthropic-managed virtual machines with network access limited by default.[23][24]

Permissions: interactive terminal and VS Code sessions start in auto mode, in which "A separate classifier model reviews actions instead of you." Manual mode asks before edits and commands. Shell sandboxing is available but "off by default", and is not offered on native Windows.[23][25]

Data terms: consumer accounts (Free, Pro, Max) are trained on when the training setting is on, which Anthropic's pricing page lists as "Opt-out"; retention is five years, or 30 days if you opt out. Team, Enterprise and API use is not trained on.[26][24]

Price (16 Oct 2026): included in Pro ($20 monthly, or $17 a month billed annually) and Max (from $100); Team standard seats $25 monthly; Enterprise $20 per seat plus usage at API rates.[26]

Job 4: delegate to a cloud agent

OpenAI Codex

What it is for: delegating a task from the ChatGPT desktop app, Codex CLI, IDE extension or Codex cloud, then reviewing the changes.[27]

How it runs: locally, the default "workspace-write" sandbox lets it edit and run routine commands inside the project and asks before going further; "full access" removes the boundaries.[28] Codex cloud blocks internet access during the agent phase by default.[29]

Data terms: for Free, Plus and Pro, OpenAI may train on content unless you opt out, and Codex has a separate "Include environments" setting. Business, Enterprise, Edu and API data is not used by default.[30]

Price (16 Oct 2026): included in ChatGPT Free and Go ($8) with limited models; Plus $20; Pro from $100; Business $20 per user a month billed annually; or API rates.[27]

Job 5: a second reader on pull requests

Review is a feature of three tools above. Copilot reviews pull requests on paid plans (using GitHub Actions minutes since 1 June 2026), and its cloud agent runs CodeQL and secret scanning on its own code.[1][3] Cursor's Bugbot comes with Premium and above.[8] Codex reviews pull requests when tagged on GitHub, from Plus.[27] At default settings, none replaces human approval: GitHub stops whoever asked the agent from approving its pull request, and Copilot's reviews do not count toward required approvals. An opt-in public preview, Copilot approvals, lets Copilot submit an approving review that satisfies a required-approval rule once it is enabled in repository, organisation and enterprise settings.[3][31]

The tools at a glance

US list prices, 16 October 2026. Vendor documentation, not measured performance.

AI Chronicle

What independent research says

METR's randomised trial, early 2025. Sixteen experienced open-source developers worked on 246 real issues in projects they knew well, each randomly allowed or barred from AI. With AI, mainly Cursor Pro with Claude 3.5/3.7 Sonnet, tasks took 19% longer, though developers believed they had been 20% faster.[32] METR's page now warns: "These results are out of date."[32]

METR's follow-up, February 2026. With 57 developers across 143 repositories, raw estimates pointed to speed-ups, but 30% to 50% of developers said they were withholding tasks they would not do without AI. METR called the data "an unreliable signal" and is redesigning the study, while judging that developers were likely sped up more by early 2026.[33]

Cui and colleagues, field experiments. In three company experiments with 4,867 developers using 2022–23 Copilot autocomplete, completed tasks rose 26.08% (standard error 10.3%), with larger gains for less experienced developers.[34] Our workplace AI evidence review sets these studies side by side.

DORA 2025. Google Cloud's survey of nearly 5,000 technology professionals found 90% use AI at work and 30% have little or no trust in AI-generated code. AI adoption was linked to higher delivery throughput but still to lower stability.[35] These are survey associations, not experiments.

Benchmark scores. SWE-bench Verified, 500 human-checked GitHub issues from 12 Python projects, became the standard coding score in launch posts.[36] In February 2026 OpenAI, which helped create it, stopped reporting it: its audit found flawed tests in 59.4% of 138 hard problems, and every frontier model tested could reproduce some original fixes from memory.[37]

"improvements on SWE-bench Verified no longer reflect meaningful improvements in models’ real-world software development abilities."

— OpenAI, Why SWE-bench Verified no longer measures frontier coding capabilities (2026).[37]

None of this tested the current versions of these seven tools. Our benchmarks explainer covers reading coding scores, and our guide to measuring AI productivity shows how to count review time.

Documented limits and risks

• Approval defaults vary. Claude Code starts with a classifier approving actions and Cursor saves file edits without asking; Junie and Devin Local ask before commands, and Codex asks before leaving its sandbox.[23][11][6][15][28]
• Sandboxes are not universal. Claude Code's is off by default; Antigravity's is on by default only on macOS and Linux.[25][19]
• Cloud agents differ on internet access: off by default in Codex cloud, on in Cursor's.[29][10]
• Training is usually opt-out on individual plans of Copilot, Devin, Codex, Claude and Antigravity.[1][16][30][24][20]
• Costs vary by task, because agent work is metered in credits or rolling windows.[1][8][27]

Who it suits, and who should skip it

• GitHub Copilot: teams on GitHub wanting agents that work through branches and pull requests. Skip individual plans for client code unless training is off.
• JetBrains AI: developers who live in JetBrains IDEs. Skip it if you mostly work elsewhere.
• Cursor: developers wanting an agent-first editor with a choice of models. Skip it if X/Grok bundled billing will not pass procurement.
• Devin Desktop: teams wanting local and cloud agents in one place. Skip Free for sensitive code; only paid plans can opt out of training.
• Google Antigravity: Google Cloud users and anyone leaving Gemini CLI. Skip the individual tier for confidential work.
• Claude Code: terminal-first developers and scripted workflows. Skip auto mode on sensitive repositories until deny rules and the sandbox are set.
• OpenAI Codex: people handing off well-scoped tasks. Skip "full access" mode outside a disposable environment.

How to test it yourself

1. Pick three defined tasks with a clear finish line: a failing test, a reproducible bug, a covered refactor.
2. Time the whole job, review and rework included, against a similar task done without AI.
3. Note the defaults (permission mode, sandbox, network) before the first run.
4. Use a branch or throwaway copy, never production credentials.
5. Read every diff and command, and count how many you rejected.
6. Run the full test suite and linters, not just the targeted test.
7. Confirm the data setting on the plan you actually use.
8. Track the cost in credits or tokens per task.

A worked example (hypothetical)

This scenario is invented to show how the jobs fit together. No tool was run, and no outputs are reported.

Amira leads a four-person team maintaining a payments dashboard. Her rule: AI gets defined tasks only.

Small fixes stay in the editor, each suggestion reviewed.

Chores go to a cloud agent. A logging-library upgrade runs on a branch, with internet limited to the package registry, and returns as a pull request.

Payments code stays local and gated. A refund-calculation change uses a terminal agent in manual mode, sandbox on, no production secrets.

Review is layered. An AI reviewer comments first; a human approves every merge.

Her test for each task: can a reviewer see what changed and why, and roll it back in one step?

Questions readers commonly ask

Which AI coding assistant is best?

We do not name one: we have not tested them, and it depends on the job. Our comparisons of Cursor and GitHub Copilot and of Claude Code and Cursor go deeper on two common choices.

Will a coding assistant make my team faster?

Possibly. The evidence is mixed: a 26% gain in field experiments, a 19% slowdown in a trial, and a follow-up METR could not measure reliably.[34][32][33] Measure on your own tasks.

Is it safe to let an agent run commands on my machine?

Only within limits you set: a sandbox where available, deny rules for sensitive files and a branch you can throw away. Our research on what agent platforms disclose about permissions compares what vendors document.

Can I use these tools on client or company code?

Check your contract and the plan's data terms first. GitHub, Anthropic and OpenAI say their business plans do not train on your data by default; their individual plans do unless you opt out.[1][24][30]

The next step

Choosing between the two most common editors? Read our Cursor vs. GitHub Copilot comparison next. For the terminal approach, see our Claude Code review.

New to AI at work? Start with how to start using AI at work.

Sources and evidence note

Product descriptions, defaults, data terms and prices come from each vendor's own documentation, pricing, terms or help pages and describe intended behaviour, not tested performance. Prices are US list prices as displayed on 16 October 2026. Independent evidence comes from METR, Cui and colleagues, Google Cloud's DORA programme and OpenAI's benchmark posts, each read in full on the publisher's site. The worked example is invented. All pages were accessed on 16 October 2026.

Related reading

• What workplace AI studies actually show
• What AI agent platforms disclose about permissions