Best AI coding agents in 2026: Claude Code, Codex, Devin and more compared

Published by 3Gen Internet Corporation, maker of OpenAssistant, Skej and AGNT. Comparisons use public documentation, without hands-on testing.

Claude Code is an assistant we recommend among the best AI coding agents in 2026, followed by OpenAI Codex and the GitHub Copilot cloud agent. We ranked seven coding agents from vendor pages only: what each one does, what each plan costs and what each vendor says about approvals and your data. Specialists lead this list. If you also want an assistant for the email, texts, calls and Slack messages around your engineering work, OpenAssistant has its own short section below. All prices are in US dollars as shown by the vendor.

Options at a glance

The ranking is an editorial judgment from vendor pages, organized by five criteria: whether the documented capabilities cover the job, the vendor-published evidence and its limitations, price for what you get, where it works, and privacy, security and admin controls. Vendor result claims are labelled as vendor claims and carry no weight in the order.

• 1. Claude Code. It covers the terminal, IDE, desktop app, browser and cloud sessions, plus scheduled tasks and CI use. Anthropic lists a SOC 2 Type 2 report and ISO 27001 certificate, and documents its permission modes and sandbox in detail. Entry price is $20 a month. It loses points for no free plan and for a permission default that changed (see its section).
• 2. OpenAI Codex. It is included in ChatGPT plans from Free to Enterprise, runs in a CLI, IDE extension and cloud, and ships with network access off by default. It ranks below Claude Code because the Codex pages we read state no SOC 2 or other certification, and the per-plan feature table could not be read in full.
• 3. GitHub Copilot cloud agent. It is the lowest-priced paid plan with a cloud agent ($10 a month), and its documented guardrails are strict: a separate branch, no merging its own pull request, an internet firewall. The pages read state no SOC 2 certification, and agent features draw down credits.
• 4. Devin. The vendor pitches the most hands-off workflow (a vendor claim), with Slack and Teams triggers, parallel cloud agents, SOC 2 Type 2 and ISO 27001 listed. It drops to fourth for a conflict between its homepage and docs on data handling, a Teams price that reads two ways, and quota sizes it does not publish.
• 5. Cursor. Certifications are strong (SOC 2 Type II, ISO 27001, ISO 42001, AIUC-1) and Privacy Mode is documented. Cloud agents are listed from the $30 Premium plan up, and the vendor’s own pages disagree on how cloud agents are billed.
• 6. Replit Agent. It is built around Replit’s cloud workspace and shipping apps, and lists SOC 2 Type II. The pages read state no data-use policy, and it is the least repo-centred pitch of the seven.
• 7. Google Jules. The free tier is real (15 tasks a day) and plans are approved before code changes. It is labelled a public beta, prices for the paid Google plans are not on Jules pages, and no compliance statement is stated.

1. Claude Code (Anthropic)

The product page says: “Hand Claude a bug fix, test, or multi-day migration.” (Claude Code page)

What it does. The documentation describes “an agentic coding tool that reads your codebase, edits files, runs commands” and says it is available in the terminal, IDE, desktop app and browser (Claude Code docs). The product page lists entry points including Slack, GitHub, VS Code, JetBrains, iOS and Android, and says it plans, writes code, runs tests and opens pull requests (Claude Code page). Cloud sessions run in an Anthropic-managed VM or a self-hosted environment and keep running after you close your laptop. They are available on Pro, Max and Team plans and share rate limits with all other Claude usage (cloud sessions docs). The docs also list cloud scheduled tasks, parallel agents, and CI use through GitHub Actions or GitLab CI/CD (Claude Code docs).

Price. (Claude pricing and Claude Code page)

• Free: Claude Code is not included.
• Pro: $17 a month on an annual plan ($200 billed up front) or $20 billed monthly.
• Max 5x: $100 a month, includes $100 a month in API credits. Max 20x: $200 a month, includes $200 a month in API credits. Max is billed monthly only.
• Team (2 to 150 seats): Standard seat $20 per seat per month annual or $25 monthly. Premium seat $100 annual or $125 monthly, with 5x the usage of a standard seat.
• Enterprise: the pricing page shows US$20 per seat per month billed annually with a 20 seat minimum, and says usage cost scales with model and task.
• Every plan has usage limits that reset on a rolling five-hour session window, with weekly caps on paid plans, shared across Claude apps and Claude Code.

Permissions: read this before you rely on approval prompts. Anthropic’s own pages disagree. The product page FAQ says Claude Code “asks for permission before making changes to your files or running commands”, while the same page and the docs say auto mode is now the default starting mode on Pro, Max and Team plans and in terminal and VS Code sessions on current versions (Claude Code page and permission modes docs). In auto mode a separate classifier model reviews actions instead of you (security docs). We follow the docs. Manual mode starts with read-only permissions, other modes include acceptEdits, plan and dontAsk (for locked-down CI), and an organization can switch auto mode off with the disableAutoMode setting (security docs and permission modes docs).

Data. Anthropic lists a SOC 2 Type 2 report and ISO 27001 certificate (security docs). For Team, Enterprise and API use, Anthropic says it does not train generative models on code or prompts sent to Claude Code; for Free, Pro and Max accounts, training depends on a setting, with a 5-year retention period if you allow it and 30 days if you do not. Standard commercial retention is 30 days, and zero data retention is per organization on Enterprise only (data usage docs).

Watch out for. No free plan. Usage limits are shared with the rest of Claude. The vendor says “no system is completely immune to all attacks” (security docs). Session transcripts are stored locally in plaintext, 30 days by default (data usage docs). Whether standard Enterprise seats include Claude Code is unclear on the pages we read, and payments are generally non-refundable (Claude pricing).

2. OpenAI Codex

The product page says: “Codex is included in your ChatGPT plan.” (Codex page)

What it does. The vendor describes Codex running many agents in parallel and says “Codex keeps going when you close your laptop” (Codex page). Cloud tasks get their own workspace, can keep working while your computer sleeps, and end in a diff review where you commit or open a pull request (cloud docs). Surfaces are the ChatGPT or desktop app, cloud, an IDE extension and a CLI. The pricing doc’s feature table lists @codex delegation in GitHub, GitHub code review, and Slack and Linear cloud integrations (Codex pricing).

Price. (Codex pricing)

• Included in ChatGPT Free, Go, Plus, Pro, Business, Edu and Enterprise plans. Codex and ChatGPT Work share usage.
• Free: $0 a month, GPT-6 Luna in the desktop app, subject to rollout. Go: $8 a month, for lightweight coding tasks. Plus: $20 a month.
• Pro: from $100 a month, with tiers at $100, $200 and $500. The page says Pro plans currently have no five-hour limit.
• Business: $20 per user per month billed annually, $25 billed monthly, 2 or more users. Enterprise and Edu: contact sales.
• API key option: usage at API rates, CLI, SDK and IDE only, with no cloud features such as GitHub code review and Slack.

Approvals and data. By default the agent runs with network access off, inside an OS-enforced sandbox with an approval policy that controls when it must stop and ask. Edits outside the workspace and network commands ask first, and destructive MCP or app tool calls always require approval (security docs). The pricing doc says there is no training on your business data by default (Codex pricing).

Watch out for. The pricing doc lists Free, Go and the $200 and $500 Pro tiers, but the product page shows only Plus, Pro at $100 and Business. The pricing doc’s feature table lost its check marks when read, so we do not state which plan has which feature. The product page note “Unlimited subject to abuse guardrails” is not explained on the pricing doc. Cloud tasks may use more of your allowance than local messages. The Codex pages we read state no SOC 2 or other certification and no training statement for Free, Go, Plus and Pro users (Codex pricing and Codex page). OpenAI’s site blocked automated readers, so the product page was read in a browser.

3. GitHub Copilot cloud agent

The vendor says agents “execute work autonomously in the background”. (Copilot page)

What it does. GitHub’s docs call it the “cloud agent” (the plans page still says “coding agent” in one place). It can research a repository, plan changes and implement them in the background, runs tests and linters in an ephemeral cloud environment, and handles branch creation, commits and pushes. It can run on a schedule or on repository events. Surfaces include github.com, mobile, VS Code and other IDEs, the CLI, Raycast, Slack, Teams, Jira and Linear (GitHub docs). On GitHub you can also assign tasks to Claude and OpenAI Codex (Copilot page).

Price. (Copilot page and Copilot plans)

• Free: $0, 2,000 completions a month and the Copilot CLI. The cloud agent is not included.
• Pro: $10 per user per month, with the cloud agent and code review, and $15 in total monthly credits ($10 base plus $5 flex).
• Pro+: $39 per user per month, premium models, $70 in total monthly credits ($39 base plus $31 flex).
• Max: $100 per user per month, priority access, $200 in total monthly credits ($100 base plus $100 flex, which “may change over time”).
• Business: $19 per user per month, 1,900 credits per user, with access control, budget control, governance and IP indemnity. Enterprise: $39 per user per month, 2,000 more credits per user than Business.
• One AI credit is $0.01. Code completions and next edit suggestions do not use credits; agent mode, code review, the cloud agent and the CLI do. You can set a dollar budget for extra usage, with alerts at 75%, 90% and 100%.

Approvals and data. The cloud agent pushes only to a copilot/ branch it creates (or to the branch of a pull request where you mention @copilot), cannot approve or merge a pull request, needs a person to click “Approve and run workflows” by default, and works behind an internet firewall. The requester cannot approve their own pull request (risks and mitigations docs). GitHub says Business and Enterprise data is not used to train its models, and individual plan users can opt out. Prompts and suggestions outside the IDE are retained for 28 days for Business and Enterprise (Copilot page).

Watch out for. The features page lists third-party agents (Claude Code and Codex) under Pro, while the plans page shows them as not included on Pro and included, in preview, on Pro+ and Max. Code review workflows also consume GitHub Actions minutes from 1 June 2026. The pages we read link a Trust Center but state no SOC 2 or ISO certification. The plans page we read showed only the individual tab, so cloud agent coverage on Business and Enterprise is not recorded here.

4. Devin (Cognition)

The homepage says: “Devin runs in the cloud or on your machine”. (Devin homepage)

What it does. The docs call it an “autonomous AI software engineer that can write, run and test code” and give a vendor rule of thumb: “if you can do it in three hours, Devin can most likely do it”. The same docs say it excludes extremely difficult tasks (Devin docs). Cloud sessions run in their own Windows, Mac or Linux virtual machines, split work across parallel cloud agents, and can be started by tagging @Devin in Slack, Teams and more. Automations handle bugs and failed CI (Devin homepage). There is also a Devin CLI for local terminal use (Devin docs).

Price. (Devin pricing and Devin billing docs)

• Free: $0, light quota. The pricing page comparison shows no cloud sessions on Free.
• Pro: $20 a month, with cloud agents (Devin Cloud) and extra usage bought at API pricing. Up to 10 concurrent sessions.
• Max: $200 a month, “significantly higher quotas”. Up to 10 concurrent sessions.
• Teams: see the price reading below. Up to 200 users, unlimited concurrent sessions, central billing and an admin dashboard.
• Enterprise: custom pricing, SAML/OIDC SSO and a dedicated deployment option. Enterprise is billed in Agent Compute Units at the rate in the order form (Devin enterprise billing docs).
• Usage allowances refresh daily and weekly (Max has a weekly allowance only). The exact quota sizes for Pro, Max and flex seats are not published.

Flag: the Teams price reads two ways. The pricing page shows “$80 /month for team plan + $40/mo per full dev seat”, which reads as a base fee plus a per-seat fee. The docs say every Teams account pays at least $80 a month, and give a table: zero, one or two full seats cost $80, three cost $120, and two or more cost the number of seats times $40. Read literally, the pricing page would make three full seats $200 a month ($80 plus 3 x $40), while the docs table gives $120. We follow the docs table, which is the clearer of the two (Devin pricing and Devin billing docs).

Flag: the homepage and the docs disagree on data retention. The homepage lists “Zero data retention enforced with model providers” under enterprise security (Devin homepage). The docs security page says: “By default, we may use your data for model training purposes”. Paid plans can opt out on the Data Controls settings page (admin only on Teams), and on Enterprise the vendor says it will never train on your data without express prior written consent (Devin security docs). We follow the docs page and do not treat zero data retention as a default for every plan.

Security and approvals. The trust center lists SOC 2 Type 2, ISO/IEC 27001:2022 and CCPA, and gives documents only after an NDA (Cognition trust center). The docs mention only SOC 2 Type II, since September 2024. The vendor advises branch protections so checks are enforced before Devin can merge, and admits that “it can still experience hallucinations, introduce bugs into code” (Devin security docs).

Watch out for. The two flags above, plus: the docs say the Free plan includes Devin Review and DeepWiki, while the pricing page comparison shows DeepWiki as not on Free. Purchased credits never expire, but the quota behind Pro and Max is described only as “light” or “significantly higher”.

5. Cursor

The homepage says: “Cursor is your coding agent for building ambitious software.” (Cursor)

What it does. Cloud agents “run in isolated VMs in the cloud with full development environments”, can run in parallel, produce merge-ready pull requests with artifacts, and can control the desktop and browser (Cursor cloud agent docs). The pricing page lists them from Premium up (Cursor pricing). Cursor also lists a desktop IDE, a CLI, mobile, automations and Bugbot reviews (Cursor).

Price. (Cursor pricing and Cursor pricing docs)

• Free: $0, limited agent requests.
• Plus: $8 a month, Grok models only, with no cloud agents, automations or third-party models.
• Premium: $30 a month (marked “Most popular”), 4x the usage of Plus, adds cloud agents and automations and third-party models at API list price. Annual billing is $300 a year.
• Super: $100 a month, 15x Plus usage ($1,000 a year). Ultra: $200 a month, 50x Plus usage ($2,000 a year).
• Teams: “$30 +/user/month” on the pricing page, and $30, $100 or $200 per seat per month in the docs depending on tier. Enterprise: custom.
• Usage refreshes weekly. Third-party model usage is not included in the monthly prices shown. Cursor is sold as part of a bundled subscription with Grok and X.

Approvals and data. The vendor lists AIUC-1, ISO/IEC 27001:2022 and ISO/IEC 42001:2023 certifications and a SOC 2 Type II attestation (Cursor security). With Privacy Mode on, it says it will not train on your data and keeps zero data retention agreements with providers; with it off, it may store codebase data and prompts (Cursor data use). By default terminal commands and all MCP connections need your approval, but agents can edit workspace files without approval except configuration files. The vendor calls its run modes “best-effort guardrails rather than a hard security boundary” (Cursor agent security docs).

Watch out for. The pricing page lists cloud agents as included from Premium, the cloud agent docs say they are “charged at API pricing for the selected model”, and the pricing docs say the weekly pool covers cloud agents. Which meter applies is not consistent across the three pages. The pricing page and the docs also use different plan names for older plans.

6. Replit Agent

The Agent page says: “Agents run independent tasks simultaneously, with progress always visible.” (Replit Agent page)

What it does. Agent 4 runs independent tasks at the same time, splits single tasks into forks, and creates mobile and web apps, landing pages, decks and videos. The vendor claims “10x faster builds”, which is a vendor claim (Replit Agent page). It runs in Replit’s cloud workspace rather than a local terminal tool. The vendor also warns that “its behavior is probabilistic” (Replit pricing).

Price. (Replit pricing)

• Core: $20 a month, or $18 a month billed annually, with $20 in monthly credits.
• Pro: $100 a month, or $90 a month billed annually, with credits starting at $100, 10 parallel agents and up to 15 collaborators. Higher credit tiers are listed, and the credit amounts for the upper tiers are not labelled.
• Enterprise: custom, with SSO, SCIM and audit logs. Extra usage is charged at pay-as-you-go rates, and you can set a usage budget.
• The docs describe a free Starter plan with daily credits up to a monthly cap and one published app that goes down after 30 days (Replit Starter docs). The pricing page shows no Starter card.

Approvals and data. Replit says it holds SOC 2 Type II certification, gives each customer an isolated cloud project, keeps secrets inaccessible to the Agent, and scans apps before publishing. It says it is working toward ISO 27001 (Replit security). Replit asks for confirmation before a paid action starts (Replit AI billing docs).

Watch out for. The pricing page says “10 parallel agents” for Pro, while the docs say up to 10 tasks at once (Replit Pro docs). The pages read state no training or data-use policy and no approval controls beyond Plan Mode and paid-action confirmation.

7. Google Jules

The homepage says: “Jules does coding tasks you don’t want to do.” (Jules)

What it does. Each task runs in a fresh virtual machine. Jules proposes a plan you can review and approve before any code changes, then creates a pull request. You can assign a task with the “jules” label on a GitHub issue, and Jules reads an AGENTS.md file if present (Jules and Jules docs).

Price. (Jules and usage limits)

• Free: 15 tasks a day, 3 concurrent.
• Jules in Pro: 100 tasks a day, 15 concurrent, included with the Google AI Pro plan.
• Jules in Ultra: 300 tasks a day, 60 concurrent, included with the Google AI Ultra plan.
• The price of the Google AI Pro and Ultra plans is not stated on any Jules page we read. Paid plans are currently for individual Google Accounts only, and limits can change.

Data. The FAQ says “Jules does not train on private repository content.” The VM has internet access, and the docs tell you not to commit secrets to your repo (Jules FAQ).

Watch out for. The FAQ says Jules is in public beta and the getting-started docs call it experimental (Jules FAQ). The home page names two different Gemini models in different places. Long-running processes such as dev servers are not supported in setup. The pages read state no compliance certification and no data-training policy for public repositories.

OpenAssistant: the pick for the work around your engineering

OpenAssistant is a general work assistant, not a specialist coding agent, and the coding specialists above beat it at writing code. It fits a different job: an assistant that acts across email, text, calls and Slack around your engineering work, such as status updates, follow-ups and scheduling.

• Where it works. The marketing page lists calls, iMessage, SMS, WhatsApp, email and a web app (OpenAssistant executive assistant page). In Slack you DM the assistant or add it to channels. The help center’s example requests include scheduling, meeting prep and project status from connected tools such as Asana, and scheduling with internal OpenAssistant users is automatic while external participants are handled by email (OpenAssistant Slack help). The help center lists GitHub and Linear among its integrations, notes that some integrations are in beta and that capabilities vary (integrations overview).
• Approvals. Autonomy can be set globally or per action type (Manual, Confirmation Window or Autonomous). A Confirmation Window is not Manual approval, because the assistant proceeds if you do nothing (OpenAssistant autonomy settings).
• Security. The trust page states SOC 2 Type II, audited annually by an independent CPA firm, and says “We never train on your data” (OpenAssistant trust page).
• Price. Starter is free: 1,000 credits at signup, then 100 bonus credits a day once used up, up to 1,000 a month. Pro is $72 monthly or $60 annual, and Team is $120 monthly or $100 annual, both per user per month at the 5,000-credit tier; higher credit tiers cost more (OpenAssistant pricing; per-user pricing confirmed by 3Gen). Plan credits do not roll over (pricing details).
• Limits. It does not join group texts or group WhatsApp chats, not every country is supported for texting, and it does not work in Slack Connect channels with outside companies (OpenAssistant ways to work).

For the broader field of general-purpose agents, see Best AI agents in 2026.

What to ask each vendor

1. Prompt injection. What happens when the agent reads an issue, a web page or a file that contains hidden instructions, and what stops it from acting on them? OWASP says “it is unclear if there are fool-proof methods of prevention for prompt injection” (OWASP LLM01), so ask which controls do not depend on the model behaving. For more, see Are AI agents safe?.
2. Approvals. Which actions run without asking by default, and can an admin lock that setting? For Claude Code, ask whether auto mode is on for your plan.
3. Data. Is my code used for training on my plan, how long are prompts retained, and what is the retention for cloud sessions? For Devin, ask which page is correct, the homepage or the docs.
4. Price. What does one realistic task cost, what is the usage allowance behind the plan, and what happens at the limit? For Devin Teams, ask for the monthly total at your exact number of full seats.
5. Access. Which repositories, branches, secrets and networks can the agent reach, and can it push or merge?
6. Compliance. Can you share the SOC 2 report or equivalent under NDA? Where a vendor page states no certification, ask.

For plan-by-plan costs, see AI coding agent pricing.

What to keep for humans

• Merging and releasing. Copilot’s documentation says its agent cannot approve or merge a pull request, and Devin’s advises branch protections. Keep a person on every merge.
• Secrets and production access. Jules’ FAQ tells you not to commit secrets to your repo. Give agents the least access that does the job.
• High-impact actions. OWASP’s guidance is to “require a human to approve high-impact actions” (OWASP LLM06).
• Design and security review. Devin’s docs say the agent can still introduce bugs. Treat agent output as a pull request from a new colleague.
• Vendor results. Rules of thumb such as “three hours of work” and “10x faster builds” are vendor claims. Judge them on your own repositories.

Related reading

• AI coding agent pricing in 2026: what Claude Code, Codex, Devin and Copilot cost: Plans and prices for Claude Code, Codex, Devin, Copilot, Cursor, Replit and Jules, the stated limits, traps and a 5-developer example.
• Best AI agent builders and platforms in 2026: no-code to code: Ten agent builders, including n8n, Copilot Studio, Make, Zapier and CrewAI, with credits, tasks and executions explained.
• Best AI agents in 2026: top picks by job, with prices: Our recommendations for each job (general work, coding, browser use, research, sales, support, enterprise and builders), with vendor prices.
• How to choose an AI agent: 10 questions to ask before you buy: Ten questions on job fit, channels, approvals, price, data and exit, with a job-to-product table.
• Are AI agents safe? Prompt injection, permissions and approval settings: Real risks from prompt injection to excess access, the controls to look for and a 10-question checklist.

Frequently asked questions

What is the best AI coding agent in 2026?

From public information, we rank Claude Code first, then OpenAI Codex and the GitHub Copilot cloud agent. Claude Code leads on breadth of surfaces, documented permission controls and listed certifications. Your answer may differ if you already pay for ChatGPT (Codex is included) or live in GitHub (Copilot Pro is $10 per user per month).

Which AI coding agent is cheapest?

Among paid plans, the lowest prices on the vendor pages are Cursor Plus and Codex Go at $8 a month, but Cursor Plus has no cloud agents and Codex Go is described as for lightweight tasks. Copilot Pro at $10 per user per month is the lowest-priced plan that lists a cloud agent. Free options exist for Codex, Copilot (no cloud agent), Devin (no cloud sessions), Cursor, Replit (per docs) and Jules. Claude Code has no free plan.

Do AI coding agents train on my code?

It depends on vendor and plan. Anthropic says it does not train on Claude Code data for Team, Enterprise and API use, but consumer plans depend on a setting. Devin’s docs say it may use your data for training by default, with an opt-out on paid plans. Cursor says Privacy Mode means no training, Jules says it does not train on private repositories, and Replit’s pages we read state nothing either way. Check the plan you will actually buy.

Can an AI coding agent merge code on its own?

Some vendors say no by design: Copilot’s documentation says its cloud agent cannot approve or merge a pull request. Others depend on your repository settings, and Devin’s docs advise branch protections so checks are enforced before it can merge. Set up branch protection and required reviews whichever agent you pick.

Try OpenAssistant

If your engineering work also runs through email, texts, calls and Slack, OpenAssistant can handle the scheduling, follow-ups and status updates while your coding agent handles the code. The Starter plan is free, and you can see all plans.