Best AI Coding Agents (June 2026): The Scored Leaderboard, Updated After Fable 5

The best AI coding agent in June 2026 depends on the benchmark that matches your work. Codex CLI on GPT-5.5 leads Terminal-Bench 2.1 at 83.4%; Claude Code on Fable 5 follows at 83.1% and Claude Code on Opus 4.8 at 78.9%. On SWE-bench Verified, Fable 5 leads at 95.0% and Opus 4.8 at 88.6%. On SWE-bench Pro, Fable 5 leads at 80.3% and Opus 4.8 at 69.2%. Fable 5 and Mythos 5 are export-suspended as of June 12.

This is a scored leaderboard, dated 2026-06-18. The agent column pairs each tool with the model it runs; benchmark columns use the agent-plus-model entry where one is published (Terminal-Bench 2.1) and the underlying model score otherwise (SWE-bench Verified and Pro). Price is the entry paid tier plus the per-token or per-credit rate after included usage runs out. Nothing here is a paid placement.

The Scored Leaderboard: AI Coding Agents (June 2026)

One row per agent paired with its strongest available model. SWE-bench Verified and SWE-bench Pro are the underlying model scores; Terminal-Bench 2.1 (Terminal-Bench v2) is the agent-plus-model entry. Open-source agents run any model, so their model row shows the frontier model they most commonly drive. All scores are dated 2026-06-18 and sourced below.

BYOK = Bring Your Own Key: the agent is free, you pay the model provider directly or run a local model. "model-set" means the agent runs whatever model you point it at, so its benchmark score equals the model row you choose (top rows). "No entry" means the agent has no published Terminal-Bench 2.1 leaderboard row. GPT-5.5 SWE-bench Verified is OpenAI-reported (88.7%); it does not yet appear on the independent llm-stats leaderboard. Codex + GPT-5.5 Terminal-Bench score dated 2026-05-01; Fable 5 entries dated 2026-06-17. SWE-bench Pro figures are vendor-reported on the Anthropic/OpenAI scaffolds (Scale's standardized public set ranks GPT-5.4 xHigh at 59.1%). Fable 5 and Mythos 5 are export-suspended.

The Benchmarks, Explained

Three benchmarks separate the agents. SWE-bench Verified measures whether a model resolves a real GitHub issue in Python (human-validated, 500 tasks). SWE-bench Pro is the harder, contamination-resistant set on larger commercial-style codebases. Terminal-Bench v2 (the 2.1 leaderboard) measures whether an agent completes a terminal-driven task end to end, scoring the agent-plus-model pair, not the model alone.

ProjDevBench (arXiv 2602.01655, Feb 2026) adds an end-to-end project-development view: agents average 138 turns and 4.81M tokens per problem, and Codex on GPT-5 leads at 77.85% overall with a 27.38% mean acceptance rate. It runs less often than Terminal-Bench, so treat its numbers as a snapshot.

What Changed in June 2026 (Post-Fable-5)

Updated June 18, 2026. Claude Fable 5 entries landed on the Terminal-Bench 2.1 leaderboard on June 17: Claude Code + Fable 5 at 83.1% and Terminus 2 + Fable 5 at 80.4%, both just behind Codex + GPT-5.5 at 83.4%. Fable 5 leads SWE-bench Verified at 95.0% and SWE-bench Pro at 80.3%, but Fable 5 and Mythos 5 are export-suspended as of June 12, so most users cannot run them today.

Best Open-Source AI Coding Agents (2026)

The open-source agents are free to install and run on any model you point them at. Ranked by GitHub stars, which track real adoption better than marketing:

anthropics/claude-code has 131,380 stars but the tool is proprietary (license field empty, the repo is for issues and docs), so it is not in this open-source list. OpenCode moved to anomalyco/opencode (redirect from sst/opencode). Goose moved to aaif-goose/goose under the Linux Foundation Agentic AI Foundation. Aider's last repo push was 2026-05-22, a visibly slower cadence than OpenCode and Cline which push daily.

1. OpenAI Codex

Codex CLI on GPT-5.5 holds the top of Terminal-Bench 2.1 at 83.4% (entry dated 2026-05-01). It spans five surfaces from one account: the CLI, an IDE extension for VS Code, Cursor, and Windsurf, the Codex Web cloud agent at chatgpt.com/codex, a desktop app (codex app), and iOS. Cloud integrations include automatic code review and Slack.

Install the CLI with curl -fsSL https://chatgpt.com/codex/install.sh | sh, npm install -g @openai/codex, or brew install --cask codex, then run codex and sign in with ChatGPT. Switch models in-session with /model (GPT-5.4, GPT-5.3-Codex, and others, with adjustable reasoning levels). You can also auth with an OpenAI API key and pay per-token API rates, though API-key mode drops the cloud features.

Best for: Developers who want the highest Terminal-Bench score and fire-and-forget cloud tasks. Plus and Pro share one combined 5-hour window across local and cloud usage. See the Codex vs Claude Code comparison.

2. Claude Code (Anthropic)

Claude Code runs in your shell and connects to any editor via VS Code or JetBrains, with desktop apps for macOS and Windows. On Terminal-Bench 2.1 it scores 83.1% with Fable 5 (entry 2026-06-17) and 78.9% (±2.5) with Opus 4.8. Fable 5 leads SWE-bench Verified at 95.0% and SWE-bench Pro at 80.3%, but it is export-suspended as of June 12, so most users run Opus 4.8, which scores 88.6% Verified and 69.2% Pro. Anthropic says Opus 4.8 is around four times less likely than its predecessor to let flaws in its own code pass unremarked.

Native install (recommended): curl -fsSL https://claude.ai/install.sh | bash on macOS, Linux, or WSL; irm https://claude.ai/install.ps1 | iex on Windows PowerShell. Also brew install --cask claude-code, winget install Anthropic.ClaudeCode, npm install -g @anthropic-ai/claude-code (Node 18+), and signed apt/dnf/apk repos. Add MCP servers with claude mcp add --transport http notion https://mcp.notion.com/mcp.

Best for: Developers who want the strongest underlying model on SWE-bench and a tool that runs in both terminal and IDE. System requirements: macOS 13.0+, Windows 10 1809+, Ubuntu 20.04+, 4GB+ RAM. Compare with Cursor.

3. Gemini CLI

Gemini CLI is the most generous free tier on this list: 60 requests per minute and 1,000 requests per day with a personal Google account via OAuth, served by a managed Gemini 3 mix of flash and pro. On Terminal-Bench 2.1, Gemini CLI on Gemini 3.1 Pro scores 70.7% (±2.9).

Install with npx @google/gemini-cli, npm install -g @google/gemini-cli, or brew install gemini-cli. MCP servers are configured in ~/.gemini/settings.json. With an API key you can pin a specific model (for example gemini-2.5-flash) instead of the OAuth-managed mix.

Best for: Developers who want a capable agent at zero cost and high daily limits. Compare with Codex and Claude Code.

4. GitHub Copilot

On June 1, 2026 Copilot switched to usage-based billing with GitHub AI Credits. One credit equals $0.01. Premium request units are gone; credits are consumed on token usage (input, cached, output) at published per-model rates. Basic code completions and next-edit suggestions are never billed in credits and stay unlimited on paid plans.

Install the CLI with npm install -g @github/copilot (Node 22+), brew install copilot-cli, or winget install GitHub.Copilot. It supports MCP servers and a /model slash command. The model menu spans Claude Opus 4.5 through 4.8 ($5 in / $25 out per 1M tokens), Sonnet 4 through 4.6 ($3 / $15), GPT-5.5 ($5 / $30 at 272K context or less), GPT-5.4 ($2.50 / $15), Gemini 3.1 Pro ($2 / $12), and Claude Fable 5 (currently suspended, see note) ($10 / $50).

Best for: Teams on GitHub who want completions, chat, agent mode, cloud agent, and code review on one bill across VS Code, JetBrains, Neovim, and Xcode. As of June 2026 new sign-ups for paid Copilot plans were paused during the billing rollout, reopening in the coming weeks. Legacy request-based annual subscribers keep 300 (Pro) or 1,500 (Pro+) premium requests/mo at $0.04 each.

5. Cursor

Cursor is a VS Code fork rebuilt around AI: Tab predicts multi-line edits, Composer handles multi-file changes, and codebase indexing feeds project context. Paid plans include frontier models, MCPs, skills and hooks, cloud agents, and Bugbot on usage-based billing.

Pricing is included-usage plus burn rate. Pro at $20/mo includes about $20 of API-rate usage; Pro+ at $60/mo includes $70; Ultra at $200/mo includes $400. Selecting Auto or Cursor's in-house Composer 2.5 model draws from a separate, more generous pool designed for everyday agentic coding at lower cost than frontier API models. Model choice sets the burn rate.

Best for: Developers who want inline AI editing without leaving the editor. See Cursor alternatives for the free options.

6. OpenCode

OpenCode (anomalyco/opencode, redirect from sst/opencode) is the most-starred open-source coding agent at 172,198 stars, ahead of Gemini CLI and Codex. It supports 75+ LLM providers through the AI SDK and the Models.dev catalog, plus local models via Ollama, LM Studio, and llama.cpp. OpenCode Zen is the team's curated list of models tested for agentic coding.

Install with curl -fsSL https://opencode.ai/install | bash, npm install -g opencode-ai, or brew install anomalyco/tap/opencode. Add a custom OpenAI-compatible provider in JSON:

{
 "provider": {
 "myprovider": {
 "npm": "@ai-sdk/openai-compatible",
 "options": { "baseURL": "https://api.myprovider.com/v1" },
 "models": { }
 }
 }
}

Best for: Developers who want a Claude Code-style CLI without lock-in to one model vendor. Compare with Claude Code and Codex.

7. Cline

Cline is an Apache-2.0 agent that runs in VS Code, JetBrains (Early Access), Cursor, and Windsurf, plus a CLI installed with npm i -g cline for macOS, Windows, and Linux. Its homepage pitch is "every model, your choice": Claude, GPT, Gemini, any OpenAI-compatible endpoint, BYOK, or local via Ollama and LM Studio. It supports MCP servers and custom tools, with an enterprise tier at cline.bot/enterprise.

For local runs, Cline's docs suggest 16-32GB RAM for small or quantized models, 32-64GB for mid-size coding models, and 64GB+ for larger models, and recommend enabling the Use Compact Prompt setting.

Best for: VS Code or JetBrains users who want agentic AI without a subscription. Compare with Claude Code and Cursor.

8. Goose (Linux Foundation AAIF)

Goose moved from block/goose to the Linux Foundation's Agentic AI Foundation (aaif-goose/goose). It is built in Rust and ships as a desktop app for macOS, Linux, and Windows, plus a CLI and an API. It works with 15+ providers (Anthropic, OpenAI, Google, Ollama, OpenRouter, Azure, Bedrock, and more), can reuse existing Claude, ChatGPT, or Gemini subscriptions via ACP, and connects to 70+ extensions over MCP.

Install the CLI with curl -fsSL https://github.com/aaif-goose/goose/releases/download/stable/download_cli.sh | bash. Goose positions itself as general-purpose, not code-only: research, writing, automation, and data analysis alongside coding. Governance runs under the AAIF with a public GOVERNANCE.md and custom-distribution support.

Best for: Developers who want a foundation-governed, provider-agnostic agent that also handles non-code automation. Compare with Claude Code.

9. Aider

Aider is the Git-native terminal agent: every change is staged with a descriptive commit, no copy-paste or manual staging. Install with python -m pip install aider-install && aider-install, the one-liner curl -LsSf https://aider.chat/install.sh | sh, or uv, pipx, and pip. It connects per-run via flags, for example aider --model sonnet --api-key anthropic=<key> or aider --model deepseek --api-key deepseek=<key>, and runs local models via Ollama and any OpenAI-compatible API.

One caveat: Aider's last repo push was 2026-05-22, slower than OpenCode and Cline which push daily, and its model guidance still recommends Gemini 2.5 Pro, DeepSeek R1/V3, Claude 3.7 Sonnet, o3/o4-mini, and GPT-4.1 rather than the 2026 frontier models. You can still point it at any current model by flag.

Best for: Terminal-native developers who want Git-integrated editing and full control over which model they pay for. Compare with Claude Code and Cline.

10. Kilo Code

Kilo Code (Kilo-Org/kilocode) is a free, open-source extension. The Kilo Gateway is $0/mo plus usage at exact provider rates with no markup. Kilo Pass subscriptions run $19, $49, and $199/mo with up to 50% bonus credits; Teams is $15/user/mo. BYOK works for Anthropic, OpenAI, Google, Azure, and Bedrock keys with no Kilo plan required. The kilocode.ai domain now 308-redirects to kilo.ai.

Best for: Developers who want no-markup gateway pricing or BYOK with optional bonus-credit passes. Compare with Claude Code.

11. Kiro

Kiro is a credit-based IDE. The free tier gives 50 credits/mo with open-weight models and Claude Sonnet 4.5. Pro is $20/mo for 1,000 credits, Pro+ is $40/mo for 2,000, and Power is $200/mo for 10,000. Overage is $0.04/credit billed month-end, and unused credits do not roll over. Team plans mirror the individual credits and add centralized billing, usage analytics, and SSO via AWS IAM Identity Center. AWS GovCloud pricing is about 20% higher with no free tier. New users get $20 credited toward a first upgrade.

Best for: Developers who want a predictable credit budget and a free tier that includes Claude Sonnet 4.5. Compare with Cursor and Claude Code.

12. Google Antigravity

Antigravity 2.0, announced at Google I/O on May 19, 2026, split into a unified harness with two surfaces: a redesigned desktop app and a new standalone CLI. It adds specialized subagents for parallel tasks, cross-platform terminal sandboxing, credential masking, and hardened Git policies. It runs Gemini 3.1 Pro, which scores 80.6% on SWE-bench Verified and 54.2% on SWE-bench Pro; on Terminal-Bench 2.1, Gemini-3.1-Pro harness entries land at 70.3% to 70.7%. Antigravity 2.0 also ships a Managed Agents API and an Antigravity SDK for self-hosted deployments.

Google AI Pro is $19.99/mo with higher rate limits to the agent model and Gemini 3.1 Pro at 4x free limits. Google AI Ultra starts at $99.99/mo with the highest rate limits and up to 20x Pro usage. In early June 2026 Google wiped quota counters to zero for all free and paid Gemini users to fix post-launch issues.

Best for: Developers in the Gemini ecosystem who want visual parallel-agent management and a Managed Agents API. Compare with Cursor and Claude Code.

How to Choose: Decision Framework

Most developers settle on two or three agents. A common stack: Codex or Claude Code for heavy agent work, Copilot or Cursor for inline completions, and one free open-source agent (OpenCode, Cline, or Aider) for model flexibility. The agents are increasingly interoperable through MCP and ACP, so the model layer underneath them is where most of the cost and quality actually lives.

The Model Layer: Where Cost and Quality Actually Live

Every agent above is a harness around a model. The harness sets the workflow; the model sets the cost and the output quality. Morph's model router picks the cheapest model that passes each request, so requests resolve cheaper and faster without a per-request model choice. When you run open-source models like DeepSeek, where you serve them matters as much as which agent calls them.

Most serverless providers quantize activations to fp8 to cut cost, which degrades output quality. Morph serves open-source coding models with 16-bit (bf16) activations, no fp8 or int8 quantization, so output matches the reference weights. morph-dsv4flash (DeepSeek V4 Flash) costs $0.139 per 1M input tokens and $0.278 per 1M output tokens. For coding specifically, Morph runs codegen-tuned speculative decoding (draft and ngram tuned on code) plus custom low-level inference kernels. See pricing for full rates.

On the search side, every agent spends tokens building context before it writes code. Cognition measured coding agents spending 60% of their time on search. WarpGrep runs as an MCP server inside Codex, Claude Code, Cursor, or any MCP-compatible agent, executing 8 parallel searches per turn across 4 turns in under 6 seconds. It is free for 100,000 requests, then $1 per 1M requests. Fast Apply merges generated diffs into your codebase at 10,500 tokens per second.

Frequently Asked Questions

What is the best AI coding agent in June 2026?

On Terminal-Bench 2.1, Codex CLI on GPT-5.5 leads at 83.4%, Claude Code on Fable 5 follows at 83.1%, and Claude Code on Opus 4.8 at 78.9%. On SWE-bench Verified the models rank Fable 5 at 95.0%, GPT-5.5 at 88.7% (OpenAI-reported), Opus 4.8 at 88.6%, and Opus 4.7 at 87.6%. On SWE-bench Pro, Fable 5 leads at 80.3% and Opus 4.8 at 69.2%. Fable 5 and Mythos 5 are export-suspended. The right agent depends on your surface, not one score: terminal-first (Codex, Claude Code, Aider, OpenCode), IDE-native (Cursor, Copilot, Cline), or free and open-source (OpenCode, Cline, Goose, Aider, Kilo Code).

What are the best open source AI coding agents in 2026?

By GitHub stars: OpenCode (172,198, MIT) is the most-starred, ahead of Gemini CLI (105,104, Apache-2.0), OpenAI Codex CLI (89,991, Apache-2.0), Cline (62,996, Apache-2.0), Goose (48,542, Apache-2.0), Aider (45,945, Apache-2.0), and Kilo Code (19,968, MIT). All are free to install and run on your own API key or local models via Ollama and LM Studio. Claude Code's repo has 131,380 stars but the tool itself is proprietary.

Which AI coding agent has the most GitHub stars in 2026?

OpenCode leads open-source agents at 172,198 stars. anthropics/claude-code has 131,380 but is proprietary. Gemini CLI has 105,104, Codex 89,991, Cline 62,996, Goose 48,542, and Aider 45,945. Aider's last repo push was 2026-05-22, slower than OpenCode and Cline which push daily.

Which AI coding tools are free in 2026?

Free to install: OpenCode, Cline, Goose, Aider, Kilo Code, and Gemini CLI (60 requests/min and 1,000/day on a personal Google account). Free paid-product tiers: GitHub Copilot Free (2,000 completions/mo), Cursor Hobby (no card), and Kiro Free (50 credits/mo with Claude Sonnet 4.5). Open-source agents are free to run but you pay per token for the model unless you run a local model.

What changed for AI coding agents in June 2026?

Claude Fable 5 entered Terminal-Bench 2.1 on June 17: Claude Code + Fable 5 at 83.1% and Terminus 2 + Fable 5 at 80.4%. Fable 5 leads SWE-bench Verified at 95.0% and SWE-bench Pro at 80.3%, but Fable 5 and Mythos 5 are export-suspended as of June 12. GitHub Copilot moved to usage-based AI credits on June 1 (1 credit = $0.01), retired premium requests, and paused new paid sign-ups during the rollout. Codex CLI + GPT-5.5 still leads Terminal-Bench at 83.4%. Goose moved to the Linux Foundation, OpenCode to anomalyco/opencode.

Should I use multiple AI coding agents?

Most developers run two or three. A common stack: Codex or Claude Code for heavy agent work, Copilot or Cursor for inline completions, and one free open-source agent (OpenCode, Cline, or Aider) for model flexibility. Because the agents interoperate over MCP and ACP, the model layer underneath them drives most of the cost and quality.