@@ -1,2 +1,2 @@ methodology.diff

How We Evaluate AI Coding Agents

We review AI coding tools the way engineers review pull requests: transparent criteria, sourced claims, and every trade-off left visible in the diff — no sponsored placements, no hidden hunks.

← Back to comparison

A developer reviewing a code diff on a laptop in low, green-tinted light

The Hunks We Review

Every tool on this site is diffed against the same six criteria. We don't assign fake composite scores — we tell you what each tool does well, what it doesn't, and where the trade-offs actually bite.

@@ -1,1 +1,1 @@ pricing-transparency

Pricing Transparency — Whether a vendor's published pricing page reflects what a working developer will actually pay, including usage-based credits, overage rates, and recent price changes. Several vendors moved from flat-rate to metered billing in 2026, and we flag that shift explicitly rather than quoting a headline number in isolation.

@@ -2,1 +2,1 @@ ide-editor-integration

IDE/Editor Integration — Which editors and IDEs a tool actually supports out of the box, versus which require a plugin, a fork, or switching your entire workflow. An agent that only works in one ecosystem is reviewed differently than one that plugs into your existing setup.

@@ -3,1 +3,1 @@ agentic-capability

Agentic Capability — Whether the tool can execute multi-step, multi-file tasks autonomously (planning, editing, running tests, reporting back) versus offering single-line or single-file completions only. We note where "agent mode" is a thin layer on top of chat versus a genuinely autonomous workflow.

@@ -4,1 +4,1 @@ context-handling

Context Handling — How much of a codebase the tool can reason about at once — single file, open tabs, or a full repository index/code graph — and how that context is built (local indexing vs. cloud-side graph vs. large context windows).

@@ -5,1 +5,1 @@ privacy-compliance

Privacy & Compliance — Whether code and prompts are retained, used for training, or processable only within a private/self-hosted/air-gapped deployment. This matters most for regulated industries and is reviewed as a hard requirement, not a nice-to-have.

@@ -6,1 +6,1 @@ benchmark-performance

Benchmark Performance — Where a vendor publishes independent or reproducible benchmark results (task-completion rates, cost-per-task), we cite the source. Where no third-party benchmark exists, we say so instead of inventing a score.

Why Methodology Matters Right Now

@@ -1,2 +1,2 @@ trust-gap.diff

Developer sentiment toward AI coding tools is shifting under the surface of rising adoption. The Stack Overflow 2025 Developer Survey found that 84% of developers are using or planning to use AI tools in development, up from 76% in 2024. But positive sentiment declined to 60% in 2025, down from over 70% in 2023-2024, and only 29% of developers trust AI output accuracy — an 11-point drop from the previous year. Roughly half of developers still avoid agentic workflows entirely or stick to simpler tools. That gap between adoption and trust is exactly why we source every claim on this site instead of repeating vendor marketing.

Full Tool Reviews

Every tool is diffed against the same four lines: what it is, what it costs, what makes it stand out, and where it falls short.

@@ -1,4 +1,5 @@ GitHub Copilot

a3f9c1e

The ubiquitous, deeply GitHub-integrated AI pair programmer.

Free tier available; Pro $10/mo, Pro+ $39/mo, and Max $100/mo for individuals; Business $19/user/mo and Enterprise $39/user/mo for organizations. Copilot moved to usage-based GitHub AI Credits billing on June 1, 2026 (1 credit = $0.01), though code completions remain free and unlimited on paid plans.

Native integration across VS Code, Visual Studio, JetBrains IDEs, Vim/Neovim, and Azure Data Studio — the widest editor reach of any tool reviewed here.

Agentic sessions can burn through included credits faster than the old flat-rate model did, making cost less predictable for heavy agent use.

$ Try GitHub Copilot →

@@ -2,4 +2,5 @@ Cursor

7e2b9d4

An AI-native code editor built as a VS Code fork.

Free/Hobby tier; Pro at roughly $20/mo, Pro+/Ultra up to $200/mo, Teams around $40/user/mo, and custom Enterprise pricing.

Composer/Agent handles multi-file editing with full codebase-wide indexing and multi-model support across Claude, GPT, and Gemini — backed by scale, with parent company Anysphere crossing $2B in annualized revenue and 1M+ paying subscribers by February 2026.

Requires switching away from your existing IDE, and its entry-level paid tier is pricier than Copilot's.

$ Try Cursor →

@@ -3,4 +3,5 @@ Claude Code

c48e112

A terminal-based, agentic coding tool built for autonomous, multi-step engineering work.

No standalone plan or free tier — access is bundled into Claude Pro (roughly $17-20/mo), Max ($100 or $200/mo), Team, Enterprise, or pay-per-token via the API (Sonnet priced around $3/$15 per million input/output tokens).

Deep CLI integration with plan mode, subagents, and a large context window; usage limits were doubled on May 6, 2026.

Rolling 5-hour session and weekly usage caps are shared with regular Claude chat use, which can create capacity constraints for heavy users, and there is no free tier.

$ Try Claude Code →

@@ -4,4 +4,5 @@ Windsurf

5d6a0f3

An AI-first code editor (formerly Codeium, now part of Cognition/Devin) built around the Cascade agent.

Free tier with unlimited Tab autocomplete; Pro $20/mo (raised from $15), Max $200/mo, Teams $40/user/mo, and custom Enterprise pricing. Windsurf moved from a credit-pool model to daily/weekly quotas in March 2026.

The Cascade agent is built specifically for agentic, multi-step editing rather than being bolted onto an existing chat interface.

Pricing has risen to match Cursor, eroding the lower-cost positioning that originally set it apart.

$ Try Windsurf →

@@ -5,4 +5,5 @@ Amazon Q Developer

9b1c7a6

AWS's AI coding assistant for cloud-native and legacy-modernization workloads — currently being retired.

Previously Free plus Pro at $19/user/mo, with a 4,000 lines-of-code monthly allocation and $0.003/LOC overage. New signups have been blocked since May 15, 2026, with full end-of-support on April 30, 2027.

Purpose-built for large-scale legacy code modernization on AWS.

The product is being discontinued; AWS is pivoting customers toward a new agentic IDE, Kiro, and migration is required.

$ Visit Amazon Q Developer →

@@ -6,4 +6,5 @@ Cody (Sourcegraph)

2f7d4e8

An enterprise-only AI coding assistant built on Sourcegraph's code-intelligence platform.

Self-serve Free and Pro plans were discontinued on July 23, 2025 (new signups stopped June 25, 2025); Cody is now sold exclusively bundled into Sourcegraph Enterprise contracts.

Multi-repo code-graph context designed for navigating large, complex codebases.

There is no individual or small-team plan anymore; former self-serve users have been pointed toward Sourcegraph's newer product, Amp.

$ Visit Sourcegraph Cody →

@@ -7,4 +7,5 @@ Tabnine

e63a2c1

A privacy-and-compliance-focused AI coding platform built for regulated enterprises.

No free tier since April 2025. Code Assistant Platform is $39/user/mo and Agentic Platform is $59/user/mo, both billed annually per seat.

Fully private, self-hosted, or air-gapped deployment options with zero code retention and IP indemnification.

Annual billing only, with no individual or monthly plan, at a price point above many per-seat competitors.

$ Try Tabnine →

@@ -8,4 +8,5 @@ Replit Agent

018f9b5

A browser-based, full-stack "vibe coding" platform combining an AI agent with hosted development, database, and deployment.

Free Starter tier; Core around $20-25/mo; Pro around $100/mo (supporting up to 15 builders, since February 20, 2026), billed on effort-based credits tied to task complexity.

Handles the entire build-to-deploy loop in one browser-based environment, not just code generation.

Effort-based billing can be unpredictable, with reported overages reaching several times the base subscription for heavy use.

$ Try Replit Agent →

@@ -9,4 +9,5 @@ JetBrains AI Assistant / Junie

b4d2e07

JetBrains' native AI layer, pairing IDE-aware completions and chat with the Junie autonomous agent.

Free tier plus AI Pro $10/mo and AI Ultimate $30/mo for individuals, with custom Enterprise pricing.

Junie scores 62.8% on the SWE-Rebench benchmark at an average cost of about $1.14 per task — one of the few tools here with a published third-party benchmark figure.

Tied to the JetBrains ecosystem, requires a separate paid subscription on top of the IDE license, and Junie is newer and less mature than Cursor's Composer.

$ Try JetBrains AI Assistant →

@@ -10,4 +10,5 @@ OpenAI Codex

6c9f3a2

OpenAI's agentic, cloud-sandboxed coding tool, bundled entirely into ChatGPT subscriptions.

Free, Go $8/mo, Plus $20/mo, and Pro from $100/mo (with 5x/20x rate-limit tiers), plus Business and Enterprise/Edu plans, on token-based credit pricing since April 2, 2026.

Delegated, sandboxed task execution — hand off a task and Codex works autonomously, then reports back with a diff, logs, and tests.

A shared 5-hour rolling usage window across CLI and cloud tasks can exhaust limits quickly under heavy parallel use; real-world active-user cost is estimated at $100-200/mo.

$ Try OpenAI Codex →

@@ -11,4 +11,5 @@ Gemini Code Assist

f10d8b4

Google's Cloud-billed coding assistant, cited here as evidence of the industry-wide shift to metered, enterprise-only pricing.

Individual and free tiers ended June 18, 2026; the product is now Standard $19/user/mo or Enterprise $45/user/mo, billed through Google Cloud.

Tight integration with Google Cloud's broader development and deployment tooling.

No individual or free option remains, matching a broader 2026 pattern of vendors retreating from self-serve pricing.

$ Visit Gemini Code Assist →

Frequently Asked Questions

@@ -1,4 +1,4 @@ faq.diff

Is there a free AI coding assistant that's actually usable?

Yes. GitHub Copilot has a free tier, Windsurf offers a free tier with unlimited Tab autocomplete, JetBrains AI Assistant has a free tier, and Replit Agent has a free Starter tier. Tabnine and Cody, by contrast, have dropped their free tiers entirely.

Which tool is best for enterprise compliance?

Tabnine is built specifically for regulated enterprises, offering fully private, self-hosted, or air-gapped deployment with zero code retention and IP indemnification. Cody is also enterprise-only, bundled into Sourcegraph Enterprise contracts, and built for multi-repo context at scale.

Why did AI coding tool pricing change in 2026?

Multiple vendors shifted from flat-rate or credit-pool pricing to usage-based or quota-based billing in 2026 — GitHub Copilot moved to AI Credits on June 1, 2026, Windsurf moved to daily/weekly quotas in March 2026, OpenAI Codex moved to token-based credits on April 2, 2026, and Gemini Code Assist ended its individual tier on June 18, 2026. Agentic workflows consume far more compute per task than simple autocomplete, and vendors have been re-pricing to reflect that cost.

Which tool works best inside JetBrains IDEs?

JetBrains AI Assistant with the Junie agent is the native option, built directly into the IDE and scoring 62.8% on the SWE-Rebench benchmark at roughly $1.14 per task. GitHub Copilot also supports JetBrains IDEs directly if you want a single assistant across multiple editors.

## Last Diffed

This page reflects each vendor's publicly listed pricing and feature set at the time of review. Pricing for agentic tools has changed multiple times across 2025-2026 and can change again without notice — check each vendor's official pricing page before you commit. Questions or corrections: hello@aicodingagents.com.