A developer used to judge a coding tool by how well it finished a line. Now the same product edits twelve files, runs the test suite, opens a pull request and bills by the credit, so the buyer is really choosing a contractor with shell access. The questions are practical. Who sees the code, and for how long? What does a heavy week of agent use cost on top of the seat price? Can the model be swapped when a better one ships, or does the contract tie you to one maker? Is there legal cover if generated code resembles someone else's? This ranking covers eleven tools a small or mid-sized engineering team can buy today, from open-source agents to editors, IDE plugins and cloud-platform assistants. We read each vendor's own pricing, documentation, terms and privacy pages, and ordered them by what a buyer can confirm in writing.
Visibility in this ranking can be paid for. Payment moves a vendor's position within the
shortlist; it never adds a vendor, and it never changes a word of the review. The largest vendors in AI coding assistants
cannot hold places 1 to 3. How it works: placement disclosure ·
editorial process.
How we ranked these
Setup: an hour for a plugin, a week for a policy
Installing is the easy part. Zed, Cursor and Windsurf are editors, so a team adopting them also adopts a new set of shortcuts, extensions and settings files. Cline, Augment, Claude Code and GitHub Copilot arrive as extensions or a terminal command inside the editor a developer already uses, and JetBrains AI Assistant sits inside IntelliJ-family IDEs. The slower step is deciding what the agent may do. Cline asks for approval before each action in its Plan and Act flow, Claude Code reads a CLAUDE.md file and runs hooks, and Cursor has project rules. Someone has to write those rules, choose which commands are allowed, and decide whether the tool may touch the production branch. Budget a week for that policy, not an hour for the install.
The real price: the seat is the floor, the meter is the bill
Every tool here publishes a number, and almost none of them is the whole number. Cursor lists $20 for an individual and $40 per user for Teams, then charges usage beyond included amounts. GitHub Copilot's Pro plan comes with $15 of monthly AI credits. Augment adds a flat 40% service fee to model cost. OpenAI Codex reports that a typical task can burn 2 to 15 credits. Cline and Zed's own-key route cost nothing to the vendor and shift the full bill to the model provider. Ask each vendor what happens at the limit: throttling, a top-up or an overage, and have a developer run a full week of real tickets before the budget is set.
Getting your data out: where the code went and what remains
Leaving a tool is easy because your code stays in git. The harder question is what the vendor kept. Claude Code states a 30-day retention period on commercial plans, with zero retention only for qualified Enterprise accounts, and a five-year period for consumer users who allow training. Zed says it does not retain prompts or code context by default. JetBrains says it will not train on inputs without consent. Cursor offers privacy mode, with a guarantee that code is not used for training. Also check what lives locally: Claude Code keeps session transcripts in plaintext on disk for 30 days by default. Ask for deletion terms for indexes and embeddings of your repository, not just chats.
Independence from the vendor: one model maker or many
The tools split along a line that matters for the next three years. Claude Code is built on Anthropic models and Codex on OpenAI's, so a price rise or a model retirement from that maker lands directly on you. Cline, Zed and JetBrains AI Assistant accept keys from several providers, and Cline is Apache 2.0 licensed, so the agent itself cannot be withdrawn. Cursor, Windsurf and GitHub Copilot offer a menu of models inside a product the vendor controls. The Windsurf name now redirects to Cognition's Devin pages, a reminder that products get renamed and re-owned. Prefer contracts that let you keep rules files, prompts and project instructions in your own repository.
Who it's for: match the tool to how the team ships
Start with the constraint that cannot move. A team that lives on GitHub gets the least friction from Copilot and its IP indemnity on Business. An AWS shop gets transformation help from Amazon Q Developer, and a Google Cloud shop gets Gemini Code Assist in one console. A Java or Kotlin group already inside IntelliJ should look at JetBrains AI Assistant first. Developers who want to delegate whole tasks from a terminal suit Claude Code or Codex. Teams that care about open code and their own keys suit Cline or Zed. Large repositories with many services are the audience Augment's context engine is written for.
The 11 tools, reviewed
#1 Zed
A fast code editor with agents, built by its own team · United States · zed.dev
Own-key friendly Editor first Low entry price
Zed takes first place because it is the cheapest way to try AI-assisted coding without handing a vendor your workflow. The pricing page lets a developer run the editor free, plug in keys from Anthropic, OpenAI, Mistral, Ollama or others, and pay Zed nothing for that traffic. The privacy documentation states that prompts and code context are not retained by default, and that Zed-hosted model providers agree to zero data retention, with one exception for safety review. Pro at $10 adds unlimited edit predictions and a small token allowance. For a team of five that wants control over which model sees which repository, the Business plan adds organization-wide AI policies.
Where it falls short
Zed is an editor first, so a team standardized on a JetBrains IDE or Visual Studio gets nothing from it. The $5 monthly token credit is small and usage beyond it is billed separately. Some Anthropic models in the hosted route keep data for 30 days for safety review, which matters for strict retention rules.
Wrong for
Teams that live inside IntelliJ-family IDEs, or buyers who want one flat seat price covering heavy agent usage. JetBrains AI Assistant or a subscription tool with a larger included allowance fits better.
Pricing: Personal is free with 2,000 accepted edit predictions; Pro is $10 per month with a 14-day trial and $5 of tokens; Business is $30 per seat per month. (free plan)
Visit Zed →
#2 Cline
Open source coding agent that uses your own model keys · United States · cline.bot
Apache 2.0 Bring your own key VPC and air-gapped
Cline ranks second because it separates the agent from the model bill more cleanly than anything else on this list. The repository carries an Apache 2.0 license, the pricing page says individuals pay only for inference, and the keys come from Anthropic, OpenAI, Google, Bedrock, Vertex or others. Its Plan and Act toggle forces a review step before files change, which suits teams that fear surprise edits. The enterprise page describes deployment in the customer's own VPC, on premises or air-gapped, where inference goes straight to the chosen provider and Cline does not see the code or prompts.
Where it falls short
There is no flat monthly price, so a heavy agent session can produce a surprising API invoice and nobody at Cline caps it for you. The privacy notice says data flows to Cline when its own provided keys are used. Enterprise controls such as SSO sit behind a sales conversation, and the JetBrains extension is listed as an enterprise feature.
Wrong for
Finance-led buyers who need predictable per-seat spend, and developers who want autocomplete without managing keys. GitHub Copilot or Cursor give a fixed monthly figure.
Pricing: Free for individual developers, who pay only for model inference by usage; Enterprise is quote only. (open source)
Visit Cline →
#3 JetBrains AI Assistant
AI chat and agents inside JetBrains IDEs · Czech Republic · jetbrains.com
JetBrains IDEs Credit tiers Czech vendor
JetBrains AI Assistant sits third because for a Java, Kotlin or Python shop already paying for IntelliJ or PyCharm it adds AI without a second editor. The documentation says it can be activated through a JetBrains subscription, personal API keys, a provider account or external agents, so the model choice is not locked to one maker. The service terms state that JetBrains will not use inputs, outputs or suggestions to train language models without consent, though instructions do travel to third-party providers. The company is registered in Prague, a rare European address in this category.
Where it falls short
Credits are opaque: one credit buys roughly ten chat generations or forty editor generations, so a month of agent work can drain 20 credits quickly. The free tier of 3 credits per 30 days is a sampler. Value depends on using a JetBrains IDE, and Mainland China is excluded from some tiers.
Wrong for
Developers who work in VS Code or a terminal, and teams that want unlimited autocomplete for a fixed fee. Cursor or GitHub Copilot fit an editor-agnostic crowd better.
Pricing: AI Free gives 3 credits per 30 days; AI Pro is $20 per month for 20 credits; AI Ultimate is $60 for 70; AI Enterprise is $60 per month. (free plan)
Visit JetBrains AI Assistant →
#4 Cursor
AI-native editor built around a coding agent · United States · cursor.com
AI-first editor Privacy mode Usage on top
Cursor lands in fourth place under the rule that dominant products cannot take the top three. It is the clearest example of the editor-as-product approach: the docs describe a coding agent that understands a repository, plans changes and works with rules, plugins and MCP servers, and it connects to GitHub, GitLab and Azure DevOps. The Teams plan at $40 per user adds shared context, usage analytics and team-wide privacy mode, where Cursor says code is not used for training by it or its model providers. Enterprise adds SCIM, audit logs and invoice billing.
Where it falls short
The headline price is a floor. The pricing page notes usage-based charges beyond included amounts, and the Individual tiers run Pro, Pro+ and Ultra. Some models need data retention approval, and Anthropic stores agent inputs and outputs for harm prevention. Switching away means leaving the editor.
Wrong for
Organizations that need a fixed annual bill with no metering, or developers who want to keep their current IDE. A plugin such as Cline or GitHub Copilot keeps the existing editor.
Pricing: Hobby is free with limited agent requests; Individual starts at $20 per month; Teams is $40 per user per month; Enterprise is custom. (free plan)
Visit Cursor →
#5 Claude Code
Terminal and IDE coding agent from Anthropic · United States · claude.com
Agent first Subscription or API Runs on Bedrock and Vertex
Claude Code takes fifth place. It reads a codebase, edits files, runs commands and opens pull requests, and the docs show it working from a terminal, VS Code, JetBrains, a desktop app and the web. Its configuration is plain files: a CLAUDE.md for standards, hooks that run shell commands around actions, and skills a team can share. Where data goes depends on the route, since the docs list Bedrock, Google Cloud and Microsoft Foundry as alternatives to the Anthropic API, and commercial plans keep code out of training by default. Zero data retention exists for qualified Enterprise accounts.
Where it falls short
It speaks only to Anthropic models, so a team that wants to compare vendors mid-task cannot. On consumer Free, Pro and Max plans, data may train future models when that setting is on, with five-year retention for those who allow it. Standard commercial retention is 30 days, and zero retention is not part of the plain Enterprise plan.
Wrong for
Developers who want inline autocomplete as they type, and buyers who must keep every vendor swappable. Cursor or Zed fit the first need; Cline or Kilo-style open agents fit the second.
Pricing: Included in Claude Pro at $17 per month billed annually ($20 monthly), Max from $100, Team from $20 per seat; the free Claude plan does not include it. (none)
Visit Claude Code →
#6 GitHub Copilot
GitHub's assistant for completion, chat and agents · United States · github.com
GitHub native IP indemnity Free tier
GitHub Copilot is sixth, a major vendor placed by the rule rather than by weakness. For a company whose code, reviews and issues already live on GitHub, the plan page shows how little there is to set up: assign a pull request to the agent, get a review, accept inline suggestions. Business and Enterprise carry IP indemnity, which the individual plans do not, and Enterprise adds codebase indexing and custom model access. The free tier of 2,000 monthly suggestions lets a developer test it before asking for a budget.
Where it falls short
Plans now include a dollar amount of AI credits, such as $15 on Pro, so heavy agent users can run out and need a bigger tier. Pro and Pro+ carry no IP indemnity. It is strongest where GitHub is the source host, and weaker for teams on GitLab or Bitbucket.
Wrong for
Teams on another Git host, and individual freelancers who need legal cover for generated code. Look at Amazon Q Developer Pro, which lists IP indemnity, or a business plan.
Pricing: Free includes 2,000 inline suggestions a month; Pro is $10 per month, Pro+ $39, Max $100; Business and Enterprise add policy controls and IP indemnity. (free plan)
Visit GitHub Copilot →
#7 Windsurf (Devin Desktop)
Cognition's agent editor, formerly Windsurf · United States · devin.ai
Cognition product Local plus cloud agents Renamed
Devin Desktop is seventh. The windsurf.com address now redirects to Devin pages, and the product page calls it the renamed Windsurf, built by Cognition. It is a full editor with debugging, diff review and a Kanban board for supervising several agents, plus a handoff to cloud execution for heavy tasks. The pricing is easy to read: Free keeps unlimited Tab completions and inline edits on a light quota, Pro at $20 adds cloud agents, and Teams puts a flat $80 base on top of $40 per full developer seat.
Where it falls short
The rename and move to a new domain make old Windsurf documentation, reviews and contracts harder to match to the current product. The pages we read give no quota numbers, and no self-hosted option appeared. The Teams base fee means a three-person team pays more per head than the seat price implies.
Wrong for
Buyers who need a documented self-hosted deployment or a stable vendor identity for procurement. Cline with a VPC deployment fits the first need better.
Pricing: Free at $0; Pro $20 per month; Max $200 per month; Teams is $80 per month plus $40 per full developer seat; Enterprise is custom. (free plan)
Visit Windsurf (Devin Desktop) →
#8 Augment Code
Coding agents with a codebase context engine · United States · augmentcode.com
Context engine Pooled credits SOC 2 Type II
Augment Code is eighth: its pricing is the most unusual here and rewards teams that read it carefully. Usage is pooled across a team and billed in dollars as model inference at provider list price, a flat 40% service fee on that inference, and compute time. The docs describe an agent platform with deep codebase intelligence, a terminal tool called Auggie and automated review on every pull request. The pricing page states that paid plans exclude AI training on customer data, that all plans hold a SOC 2 Type II report, and that Enterprise adds CMEK.
Where it falls short
The 40% fee on top of model cost is a real surcharge that grows with use. Unused monthly credits do not roll over, though top-ups last twelve months. The fixed seat ceiling of 50 on Standard and Business means larger teams must go to Enterprise. No free plan appeared on the pricing page.
Wrong for
Solo developers who want a free tier, and teams that can run their own keys more cheaply. Cline or Zed with your own keys avoids the service fee.
Pricing: Standard is $20 per month flat with $20 of usage, Business $100 with $100 of usage, both up to 50 seats; a 40% service fee applies to LLM usage. (none)
Visit Augment Code →
Bundled with ChatGPT Web, CLI and IDE Credit based
OpenAI Codex is ninth, partly because it is a feature of a broader subscription instead of a standalone developer product. That cuts both ways. A company already paying for ChatGPT Business at $20 per user gets Codex with SSO, MFA and no default data training, with no second vendor review. The pricing page lists a free tier and an $8 Go plan, and shows credit rates per million tokens by model, so a buyer can estimate cost. Enterprise adds audit logs, priority processing and data residency options.
Where it falls short
Usage is metered in credits that vary by model and task, and the page itself says a typical task can use anywhere from 2 to 15 credits, which makes forecasting a monthly bill difficult. The free and Go tiers run on the smaller model. It is tied to one model family, so no comparison across makers.
Wrong for
Teams that want a model-neutral tool or an editor-centered workflow, and buyers who dislike credit meters. Zed or Cline with your own keys is the opposite approach.
Pricing: Free at $0; Go $8 per month; Plus $20; Pro from $100; Business $20 per user per month on annual billing ($25 monthly); Enterprise is custom. (free plan)
Visit OpenAI Codex →
#10 Amazon Q Developer
AWS's developer assistant for IDEs and console · United States · aws.amazon.com
AWS native IP indemnity Java upgrades
Amazon Q Developer is tenth. It earns its place on two specifics: code transformation and AWS depth. The pricing page lists 1,000 lines of Java upgrade per month free, and 4,000 pooled lines per user on Pro, with extra lines at $0.003 each, which is a concrete number a platform team can budget. Pro lists IP indemnity and says data collection is automatically opted out. The product page covers JetBrains, VS Code, Visual Studio and Eclipse, plus the AWS console, so it reaches developers wherever they work.
Where it falls short
Outside AWS work it has no edge over the other tools. The free tier of 50 agentic requests a month runs out within a few sessions. Our sources did not confirm the roadmap beyond the pages read, and benchmark claims on the product page come from AWS itself.
Wrong for
Teams not built on AWS, and anyone who needs a generous free allowance for open-ended agent work. GitHub Copilot or Zed suit that profile better.
Pricing: Free tier with 50 agentic requests per month; Pro is $19 per user per month, adding higher limits, an admin dashboard and IP indemnity. (free plan)
Visit Amazon Q Developer →
#11 Gemini Code Assist
Google Cloud's coding assistant for teams · United States · codeassist.google
Google Cloud Two editions Daily request quotas
Gemini Code Assist closes the list at eleventh with the least verified detail, which says something about how a buyer meets it. The developer documentation names two managed editions, Standard and Enterprise, and the quota page puts agent mode and the Gemini CLI under one shared daily cap of 1,500 requests per user on Standard and 2,000 on Enterprise. One prompt can use several model requests, so the cap is looser than it sounds. For a team already administering Google Cloud, it keeps billing and identity in the same console.
Where it falls short
The pages we could read did not give a list price per seat, a free plan, or data-training terms, so the buyer must get those from Google in writing. Daily caps apply across all models together, and one prompt can use several requests. It sits inside Google Cloud administration, which is heavier than installing a plugin.
Wrong for
Small teams that want to see a price and sign up with a card in five minutes, and companies outside Google Cloud. GitHub Copilot or Cursor publish a clearer entry price.
Pricing: Sold as Standard and Enterprise editions through Google Cloud; the documented quotas are 1,500 requests per user per day on Standard and 2,000 on Enterprise. (none)
Visit Gemini Code Assist →
What the data says about this market
Of the eleven tools ranked, ten are published by companies based in the United States and one, JetBrains, by a company registered in Prague. No tool here comes from Germany, France or the Nordics, so a buyer who wants a European supplier has one option in this list, and JetBrains still sends instructions to third-party model providers. Measured by how the products reach a developer, the field splits three ways: standalone editors (Zed, Cursor, Windsurf), plugins and agents that sit inside an existing IDE or terminal (Cline, Augment, Claude Code, JetBrains AI Assistant, Copilot), and assistants bundled into a platform account (Codex, Amazon Q Developer, Gemini Code Assist).
Pricing is more open than in most enterprise categories. In the pages we read, ten of the eleven vendors publish a number, either a seat price, a credit tier or a quota, and only Gemini Code Assist leaves its price to Google Cloud documentation we could not confirm. Seven offer a free plan: Zed, JetBrains, Cursor, GitHub Copilot, Windsurf, Codex and Amazon Q Developer, with Cline free as open source. Claude Code and Augment Code have no free plan. The free tiers are samplers, though: JetBrains gives 3 credits per 30 days, Amazon Q Developer 50 agentic requests per month, and Copilot 2,000 inline suggestions.
The commercial shape of the market is moving from a flat seat price to metered credits. Copilot, JetBrains, Codex, Augment and Windsurf all describe a credit, quota or usage layer, and Cursor says usage-based pricing applies beyond included amounts. For a European company the data question matters more than the sticker price. ICT services rose from 10.58% of EU service exports in 2013 to 17.55% in 2023 (World Bank), and the software those exports pay for increasingly sends source code across the Atlantic. None of the eleven tools published a European-only processing guarantee in the pages we read.
The 11 ranked vendors, counted
- Headquarters by region: North America 10, Europe 1
- By country: United States 10, Czech Republic 1
- Pricing model: Per user / month 2, Credits / month 1, Flat monthly + usage 1, Open source + paid tiers 1, Per seat / month 1, Per seat / month + base fee 1, Per seat / month + credits 1, Per seat / month + usage 1, Per seat / month or API usage 1, Per user / month + credits 1
- Free option: Free plan 7, None 3, Open source 1
Counted from the 11 vendors on this page. More in our market data.
For the wider market behind AI coding assistants, read our report The Global Shift to ICT Services,
or browse all industry reports.
Questions and answers
What is the best AI coding assistant in 2026?
Zed leads this ranking because a developer can start free, bring keys from the model provider of their choice, and pay Zed nothing for that traffic, while its privacy page says prompts are not retained by default. Cline is the better pick for a team that wants an Apache 2.0 agent with VPC deployment, and JetBrains AI Assistant is the better pick for IntelliJ-family shops. Cursor, Claude Code and GitHub Copilot are widely used but, as dominant products, are placed from fourth down. A team already on GitHub will often still choose Copilot.
Is GitHub Copilot still worth it when Cursor and Claude Code exist?
For a team whose repositories and reviews are on GitHub, yes. Its Business and Enterprise plans include IP indemnity, which GitHub says the individual plans do not, and the tool needs little setup. The Free tier gives 2,000 inline suggestions a month, and Pro costs $10. Cursor suits developers who want the whole editor built around an agent, and Claude Code suits people who delegate multi-file tasks from a terminal. Many developers pay for two of them.
Which AI coding assistant lets me use my own API keys?
Cline, Zed and JetBrains AI Assistant all document it. Cline is free for individuals and has you pay for inference by usage, with keys from OpenAI, Anthropic, Google, Bedrock and others. Zed's pricing page says unlimited use with your own keys or external agents. JetBrains lists personal API keys as an activation option. The benefit is a bill you can inspect at the model provider, and the risk is a bill that nobody caps.
Does an AI coding assistant train on my company's code?
It depends on the plan, not the product name. Anthropic says it does not train generative models on code sent to Claude Code under commercial terms, but consumer Free, Pro and Max accounts can allow training. Cursor says privacy mode means code is not used for training by it or its model providers. Augment says paid plans exclude AI training. JetBrains says it will not train on your inputs without consent. Read the plan-level wording and get it in your contract.
Is there a European AI coding assistant?
Only one in this ranking: JetBrains AI Assistant, from JetBrains s.r.o., registered in Prague. It is not European-only in practice, because its terms say instructions are sent to third-party language model providers. Cline, Zed, Cursor and the others are published by US companies. A team that needs processing inside the EU should ask each vendor for a written statement of where inference runs, since none of the pages we read promised it.
How much does an AI coding assistant cost per developer?
Entry prices cluster between $10 and $40 per month. Zed Pro is $10, GitHub Copilot Pro $10, JetBrains AI Pro $20, Cursor Individual from $20, Codex Plus $20, Amazon Q Developer Pro $19 and Cursor Teams $40 per user. These are floors. Credits, usage layers and top-ups can add to them, and heavy agent users on Claude Max or Copilot Max pay $100. Run a week of real work before approving an annual budget.
What happened to Windsurf?
The windsurf.com address now redirects to Devin pages operated by Cognition, and the Devin Desktop page describes the product as the renamed Windsurf. The editor still exists, with a Free plan, Pro at $20 per month, Max at $200 and a Teams plan at $80 per month plus $40 per full developer seat. If your team signed a Windsurf contract, ask Cognition to confirm in writing which entity and terms now apply.
Can an AI coding assistant run in my own cloud or on premises?
Cline's enterprise page describes deployment in a customer VPC, on premises or air-gapped, with inference going straight to the chosen model provider. Claude Code can use Amazon Bedrock, Google Cloud's agent platform or Microsoft Foundry as its model route, and the docs describe self-hosted environments for cloud sessions. Augment's Enterprise tier lists customer-managed keys. Check what stays in your infrastructure, because model calls may still leave it.
Do I need IP indemnity for AI-generated code?
It matters if your customers or investors ask who is liable when generated code resembles licensed code. GitHub states IP indemnity for Copilot Business and Enterprise but not for individual plans, and Amazon Q Developer Pro lists it too. Several other vendors did not mention it in the pages we read. Ask for the clause itself, including any conditions attached, before treating it as protection.
Should a small team pick an editor or a plugin?
Pick a plugin or terminal agent when the team already has a standard editor, because it avoids retraining and keeps one set of extensions. Cline, GitHub Copilot, Augment and Claude Code work that way. Pick an editor such as Zed or Cursor when you want the interface itself designed around agents and can afford the switch. A trial week with two developers in each camp usually settles it.
What should I watch for when the free tier ends?
Free tiers are small. JetBrains gives 3 credits per 30 days, Amazon Q Developer 50 agentic requests a month, and Copilot 2,000 suggestions. Check whether training is permitted on free-tier data, which differs by vendor, and whether paid plans reset monthly or roll credits over. Augment, for example, says unused monthly credits do not roll over. Move the team to a paid plan before a deadline week, not during it.
