Best AI Coding Agent (2026): Ranked by Terminal-Bench, Price, and Source

Terminal-Bench 2.1 leaderboard (the agent benchmark that matters)

Terminal-Bench measures an agent driving a real terminal to complete development tasks: editing files, running commands, fixing failures. It tests the agent and model together, which is the right unit, because the same model scores differently inside different agents. The chart below is the harness-level run of the 89 v2.1 tasks that Artificial Analysis executes independently on the Terminus 2 agent (pass@1, 3 repeats per task), read August 5, 2026. The vendor-submitted agent-plus-model entries on tbench.ai follow in the table.

Ranks 1 and 2 carry the Artificial Analysis harness score for the agent's default model; rank 3 is the vendor-submitted tbench.ai pair. Agents below rank 3 are model-agnostic (BYOK), IDE-bundled, or pay-as-you-go and are not submitted as single pairs, so "n/a" means no public score for that pair, not a low score. Aider's 88.0% is gpt-5 (high) on Aider's own polyglot leaderboard. On SWE-bench Pro, Fable 5 reports 80.3% and Opus 4.8 reports 69.2%.

Two leaderboards disagree on the top model, and that is fine because they test different things. Terminal-Bench rewards driving a terminal end to end. SWE-bench Pro rewards fixing real GitHub issues. On the self-reported SWE-bench Pro aggregate at llm-stats.com, Claude Fable 5 leads at 80.3% and Opus 4.8 scores 69.2% (up 4.9 points from Opus 4.7's 64.3%), ahead of GPT-5.5 (58.6%) and Gemini 3.1 Pro (54.2%). Read benchmarks as the agent-plus-model pair, not the model alone.

The frontier models behind the agents

Every BYOK agent inherits the model you give it. As of August 5, 2026, Claude Fable 5 leads agentic coding at 95.0% SWE-bench Verified and 80.3% SWE-bench Pro, and it is generally available again after the June 12 export order was lifted on July 1. Claude Opus 5 is the Claude Code default at $5 in / $25 out per 1M tokens with a 1M-token context window. GPT-5.6 went generally available on July 9 in three tiers, Sol, Terra, and Luna, and Codex defaults to gpt-5.6-sol at medium effort. DeepSeek V4, GLM-5.2, Qwen 3.7, MiniMax M3, and Kimi K2.6 are the open-weight options you can self-host or buy by the token. One frontier model is still gated: Claude Mythos 5 is invitation-only inside Anthropic's Project Glasswing.

Pricing, side by side

Open source agents are free as tools; you pay for model tokens. Subscription agents bundle model access into a plan with usage windows or credits. Prices verified from vendor pages on August 5, 2026.

Claude Code

Anthropic's terminal-native agent, also available in VS Code and JetBrains, a desktop app, and the web. Its default model has been Claude Opus 5 since July 24, 2026 ($5 in / $25 out per 1M tokens, 1M-token context), which scores 89.1% on Terminal-Bench 2.1 at max effort; the older Opus 4.8 pairing scores 78.9%. Fable 5 is selectable again and leads SWE-bench Pro at 80.3%. The repo anthropics/claude-code has 140,331 stars and shipped v2.1.222 on August 4, 2026, but is proprietary (the repo is for issues and docs, with no open-source license). It supports MCP, sub-agents, background and cloud sessions, CLAUDE.md memory, hooks, and skills.

Install

# Native install (recommended)
curl -fsSL https://claude.ai/install.sh | bash      # macOS / Linux / WSL
# Windows PowerShell:
#   irm https://claude.ai/install.ps1 | iex

# Alternatives
brew install --cask claude-code
npm install -g @anthropic-ai/claude-code           # Node 18+

# Add an MCP server
claude mcp add --transport http notion https://mcp.notion.com/mcp

Pricing and limits

Claude Pro is $20/mo monthly or $17/mo billed annually and includes Claude Code; Max starts from $100/mo, with Max 20x at $200/mo. Usage runs on a 5-hour rolling session window plus a weekly cap, shared across claude.ai, Claude Desktop, and Claude Code on the same subscription. The free Claude.ai plan does not include Claude Code. It also runs via Amazon Bedrock, Google Vertex AI, and Microsoft Foundry, and the terminal CLI and VS Code extension support third-party model providers.

Long sessions stay coherent with built-in auto-compaction. Compare directly at Claude Code vs Codex and Claude Code vs Cursor.

OpenAI Codex CLI

OpenAI's open-source agent (openai/codex, 104,106 stars, Apache-2.0, stable 0.146.0 as of July 29, 2026). Its default Power setting runs gpt-5.6-sol at medium effort, and GPT-5.6 Sol at extra-high effort tops Terminal-Bench 2.1 at 89.5%. GPT-5.6 ships three tiers: Sol for hard, open-ended work ($5 in / $30 out per 1M tokens), Terra as the everyday workhorse ($2 / $12), and Luna for high-volume repeatable tasks ($0.20 / $1.20). gpt-5.2 and gpt-5.3-codex are deprecated for ChatGPT sign-in. Surfaces include the CLI, an IDE extension, the Codex Web cloud agent at chatgpt.com/codex, a desktop app, and iOS, with automatic code review in the cloud.

Install

curl -fsSL https://chatgpt.com/codex/install.sh | sh   # macOS / Linux
npm install -g @openai/codex
brew install --cask codex

codex            # run, then "Sign in with ChatGPT"
# /model         # switch model (gpt-5.6-sol default, gpt-5.6-terra, gpt-5.6-luna)

Pricing and limits

Codex ships with the ChatGPT plan ladder: Free $0, Go $8/mo, Plus $20/mo, Pro from $100/mo (5x and 20x), Business $20/user/mo billed annually, and Enterprise custom. Usage is metered in messages per shared 5-hour window. You can also auth with an OpenAI API key and pay per-token rates (Sol $5 / $30, Terra $2 / $12, Luna $0.20 / $1.20 per 1M tokens after OpenAI's July 30, 2026 cut to the Terra and Luna tiers), with no cloud features in API-key mode.

Cursor

A VS Code fork built around an agent loop, now from Anysphere, which also acquired Continue.dev. Individual plans: Pro $20/mo includes $20 of API-rate usage; Pro+ $60/mo includes $70; Ultra $200/mo includes $400; Teams is $40/user/mo. Cursor's in-house Composer line draws from a separate, more generous Auto and Composer pool designed for everyday agentic coding at lower cost than frontier API models. The Hobby tier is free with limited Agent requests and Tab completions, no card required. Paid plans add frontier models, MCPs, cloud agents, and Bugbot reviews on usage-based billing.

See Cursor alternatives and Cursor vs Windsurf vs Copilot.

GitHub Copilot

Copilot now meters everything in GitHub AI Credits, the replacement for the old premium requests, where 1 credit equals $0.01 of model usage. Free $0 gives 2,000 code completions per month with models reached through auto model selection only. Pro $10/user/mo adds unlimited completions, the Copilot cloud agent, and 1,500 credits a month (1,000 base plus a 500 flex allotment). Pro+ $39/user/mo includes 7,000 credits (3,900 plus 3,100) and premium models; Max $100/user/mo includes 20,000 (10,000 plus 10,000). Organizations pay $19 per granted seat for Business and $39 for Enterprise, and new self-serve Business sign-ups for orgs on GitHub Free or Team have been paused since April 22, 2026. Credits consume on token usage at published per-model rates, and paid plans get a 10% discount on model costs when they let auto model selection pick. Code completions and next edit suggestions are not billed in credits and stay unlimited on paid plans.

Install the CLI

npm install -g @github/copilot     # Node 22+
brew install copilot-cli
# supports MCP servers and a /model switch

Compare at Copilot vs Claude Code and Cline vs Copilot.

Windsurf (now Devin Desktop)

Cognition, the maker of Devin, folded Windsurf into Devin Desktop; windsurf.com/pricing now redirects to devin.ai/pricing. The former free Windsurf editor is the Devin Free tier ($0/mo, unlimited Tab completions and inline edits, a light agent quota and limited model availability). Devin Pro is $20/mo with full model availability, free use of SWE 1.6 and leading open-source models, and Devin Cloud agents. Devin Max is $200/mo with much higher quotas, and Devin Teams is $80/mo plus $40/mo per full dev seat.

opencode

anomalyco/opencode (moved from sst/opencode) has 193,678 stars under MIT and shipped v1.18.13 on August 4, 2026, ahead of Claude Code (140,331), Gemini CLI (106,374), and OpenAI Codex (104,106). Terminal-native, with a desktop app and IDE extension, it is model-agnostic: configure any LLM provider with your own keys, plus local models through Ollama and LM Studio. opencode Zen is the team's curated, tested model list for agentic coding. It supports MCP servers, LSP servers, and sub-agents.

Install and add a custom provider

curl -fsSL https://opencode.ai/install | bash
npm install -g opencode-ai
brew install anomalyco/tap/opencode

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

Cline

cline/cline: 65,680 stars, Apache-2.0, free, with usage-based inference only and no subscription. It is model-agnostic via your own API key across Anthropic, OpenAI, Google, OpenRouter, AWS Bedrock, GCP Vertex, Groq, Cerebras, and DeepSeek, or local models via Ollama and LM Studio. It runs as a VS Code and IDE extension plus an SDK and CLI, with explicit Plan and Act modes that require approval before each change.

See Cline alternatives, Cline vs Cursor, and the head-to-head below.

Aider

Aider-AI/aider: 47,956 stars, Apache-2.0, free, BYOK and model-agnostic. The terminal pair-programming pioneer that thinks in git, every edit a commit. Aider publishes the polyglot leaderboard (225 Exercism exercises across C++, Go, Java, JavaScript, Python, Rust), where gpt-5 (high) leads at 88.0%, o3-pro (high) at 84.9%, and gemini-2.5-pro at 83.1%. Its last repo push was May 22, 2026, a slower cadence than opencode or Cline, and the leaderboard has not been refreshed with the newest 2026 frontier models.

python -m pip install aider-install && aider-install
# or one-liner:
curl -LsSf https://aider.chat/install.sh | sh

aider --model sonnet --api-key anthropic=<key>
aider --model deepseek --api-key deepseek=<key>

Related: Aider vs Cline, opencode vs Aider, Morph vs Aider diff.

Kilo Code

Kilo-Org/kilocode: 26,720 stars, MIT (the domain kilocode.ai now redirects to kilo.ai). The extension is free and open source for VS Code, JetBrains, and the CLI; AI usage is billed separately. The Kilo Gateway is pay-as-you-go at exact provider rates with 0% markup, routing across 500-plus models from 60-plus providers with an Auto Model selector (Frontier, Balanced, and Free tiers). BYOK works for Anthropic, OpenAI, Google, Azure, and Bedrock keys, and local models run via Ollama or LM Studio. See Kilo Code vs Claude Code.

Gemini CLI and Google Antigravity

google-gemini/gemini-cli: 106,374 stars, Apache-2.0. On June 18, 2026 Gemini CLI and the Gemini Code Assist IDE extensions stopped serving requests for free, Google AI Pro, and Ultra users, replaced by Antigravity CLI, which is available to everyone. Enterprise access through Code Assist Standard and Enterprise is unchanged. The former free quota was 60 requests per minute and 1,000 per day with a personal Google account.

Gemini CLI and Antigravity run Gemini 3 models with auto-routing (Gemini 3 Pro and Gemini 3 Flash), a 1M token context, MCP, Google Search grounding, and shell and file tools. With Gemini 3.1 Pro the pairing scores 70.7% on Terminal-Bench 2.1.

npx @google/gemini-cli
npm install -g @google/gemini-cli
brew install gemini-cli
# MCP servers configured in ~/.gemini/settings.json

Compare at Gemini CLI vs Claude Code, Gemini CLI vs Codex, and Antigravity vs Claude Code.

Zed and Amp

Zed

zed-industries/zed: 88,064 stars, a Rust-based multiplayer editor with agentic AI. The Personal (Free) plan gives 2,000 accepted edit predictions per month and supports external agents (Claude Agent, Codex CLI). Pro is $10/mo with unlimited edit predictions plus $5 of included tokens; Business is $30/seat/mo. BYOK is unlimited across Anthropic, OpenAI, Google AI, Ollama, OpenRouter, and Bedrock.

Amp

Sourcegraph's coding agent: a CLI plus IDE integrations for VS Code, JetBrains, Neovim, and Zed. Pricing is pay-as-you-go credits with no subscription, a $5 minimum, and no markup for individuals or teams; Enterprise costs 50% more and requires a one-time $1,000 purchase. Its modes are deep (GPT-5.5 extended thinking), smart, and rush, with an Oracle tool (GPT-5.5 high-reasoning) for a second opinion. Amp spawns parallel sub-agents and supports MCP. See Cursor alternatives for where these fit.

Roo Code and Continue.dev

Roo Code (24,294 stars, Apache-2.0) is a free, BYOK Cline fork for VS Code; roocode.com now redirects to roomote.dev, the team's separate hosted cloud product (from $99/mo). Continue.dev (34,559 stars, Apache-2.0) was acquired by Cursor, and its open source extension remains available.

Cline vs opencode

Both are free and BYOK. opencode wins on community size; Cline wins if you want the agent embedded in VS Code or another IDE with explicit per-change approval. Full breakdown: opencode vs Cline.

Kilo Code vs opencode

Both are MIT-licensed and free. Kilo Code is a VS Code, JetBrains, and CLI extension with a 0% markup gateway and an Auto Model router; opencode is the terminal-native agent with the largest community and any-provider BYOK.

Kilo Code routes across 500-plus models with no gateway markup; opencode has a far larger community (193,678 vs 26,720 stars). Full breakdown: opencode vs Kilo Code.

Aider vs opencode

opencode pushes code daily and is model-agnostic across any provider. Aider is the git-native pioneer, but its last repo push was May 22, 2026 and its polyglot leaderboard has not been refreshed for 2026 frontier models.

Full comparison: opencode vs Aider.

The model backend matters as much as the agent

Most of these agents are BYOK: opencode, Cline, Aider, Kilo Code, Zed, and Gemini CLI all let you point at any OpenAI-compatible endpoint. The model and the inference provider behind it set both your cost and your output quality, independent of the agent.

If you run DeepSeek, GLM, Qwen, or MiniMax, where you serve them matters. Morph Open Source Models serve these with 16-bit (bf16) activations and no fp8 or int8 quantization. Most serverless providers quantize activations to fp8 to cut cost; keeping full 16-bit means responses match the reference weights rather than an 8-bit approximation.

For codegen, Morph runs codegen-tuned speculative decoding plus custom low-level inference kernels. morph-v3-fast applies edits at ~10,500 tokens per second; morph-dsv4flash (DeepSeek V4 Flash) is $0.139 per 1M input tokens and $0.278 per 1M output, and morph-glm52-744b (GLM-5.2) is $1.1 per 1M input and $4.1 per 1M output, both at 16-bit bf16 against the fp8 activations typical of serverless hosts. See pricing.

Morph also runs a model router that classifies each request in ~430ms at $0.001 per request and sends it to the cheapest model that passes, which lowers both cost and latency without you pinning a model per call. Pairing any BYOK agent with WarpGrep ($0 for 100k requests, $1 per 1M Pro) lifts retrieval quality independent of the model.

Sources

Primary sources behind the numbers on this page, verified August 5, 2026:

• Terminal-Bench 2.1, independent harness runs: Artificial Analysis (Terminus 2, pass@1 over 3 repeats)
• Terminal-Bench 2.1 agent-plus-model submissions: tbench.ai
• Model availability, context windows, and Claude pricing: Anthropic models overview
• GPT-5.6 tiers, defaults, and per-token rates: Codex models, gpt-5.6-sol
• GitHub AI Credits allowances: Copilot usage-based billing, Copilot plans
• SWE-bench Verified and Pro (vendor self-reported aggregate): llm-stats.com Verified, llm-stats.com Pro
• SWE-bench Pro public leaderboard (standardized): Scale Labs
• GitHub stars: GitHub API for opencode, Cline, Aider, Kilo Code, Codex, Claude Code
• Vendor pricing: Claude, Codex, Cursor, Copilot, Devin, Kilo, Zed, Amp
• 2026 shifts: Gemini CLI to Antigravity CLI, Anthropic Fable and Mythos access, Continue.dev acquired by Cursor
• CLI versions: GitHub releases for Claude Code v2.1.222 and Codex 0.146.0

Frequently Asked Questions

What is the best AI coding agent in 2026?

Codex with GPT-5.6 Sol and Claude Code with Opus 5 lead, and they are half a point apart: 89.5% and 89.1% on Terminal-Bench 2.1 as run by Artificial Analysis on the Terminus 2 harness, with GPT-5.6 Terra at 88.0%. Among the vendor-submitted pairs on tbench.ai, Claude Code with Fable 5 is 83.8% and Codex CLI with GPT-5.5 is 83.1%. Fable 5 is runnable again since July 1, 2026 and leads SWE-bench Pro at 80.3%. For open source, opencode (193,678 stars, MIT) is the most-starred agent. Pick Codex for the benchmark ceiling, Claude Code for reasoning depth on long tasks, and opencode or Cline for a free, model-agnostic agent.

Are AI coding agents actually good?

Yes, by 2026 the top agents complete a majority of real tasks. GPT-5.6 Sol completes 89.5% of the 89 Terminal-Bench 2.1 tasks and Claude Opus 5 completes 89.1%, both on the Terminus 2 harness, while Claude Fable 5 fixes 80.3% of SWE-bench Pro GitHub issues; two years earlier those numbers were under 40%. They still fail on long, underspecified, or unfamiliar tasks, so they work best with review on each change rather than fully unattended.

Cline vs opencode: which is better?

opencode has 193,678 stars (MIT) and is model-agnostic across any provider; Cline has 65,680 stars (Apache-2.0) and runs as a VS Code and IDE extension plus a CLI with Plan and Act approval modes. Choose opencode for a terminal agent and the largest community; choose Cline for an in-IDE agent with step-by-step approval. See opencode vs Cline.

Kilo Code vs opencode: which is better?

Both are MIT and free. Kilo Code (26,720 stars) is an IDE and CLI extension with BYOK or local models and a 0% markup gateway routing across 500-plus models. opencode (193,678 stars) is terminal-native with any provider and a larger community. Pick Kilo Code for in-IDE BYOK with no markup and a model router, opencode for a CLI workflow.

Aider vs opencode: which is better?

opencode (193,678 stars) ships daily and is model-agnostic across any provider. Aider (47,956 stars) is the git-native pioneer but its last repo push was May 22, 2026 and its polyglot leaderboard is not refreshed for 2026 frontier models. Use opencode for active development and provider choice; use Aider for its auto-commit-per-edit git workflow.

How much does an AI coding agent cost?

Open source agents (opencode, Cline, Aider, Kilo Code, Zed) are free as tools; you pay only for model tokens. GitHub Copilot Free gives 2,000 code completions per month. Copilot Pro is $10/mo with 1,500 AI credits, Cursor Pro and Claude Code Pro are $20/mo ($17/mo annual for Claude), and Codex ships from Free $0 to ChatGPT Plus at $20/mo. To cut token cost on open-weight models, run DeepSeek V4 Flash on Morph at $0.139/1M input and $0.278/1M output.

Is Windsurf still a separate coding agent?

No. Cognition folded Windsurf into Devin Desktop, and windsurf.com/pricing now redirects to devin.ai/pricing. The former free Windsurf editor is the Devin Free tier ($0/mo); Devin Pro is $20/mo with SWE 1.6 and open-source models, and Devin Max is $200/mo. Continue.dev was acquired by Cursor, with its open source codebase still available.

SWE-bench vs Terminal-Bench: what is the difference?

Terminal-Bench 2.1 tests an agent driving a terminal end to end (GPT-5.6 Sol leads the independently run harness at 89.5%). SWE-bench Pro and Verified test fixing real GitHub issues (Fable 5 leads the self-reported SWE-bench Pro aggregate at 80.3%, with Opus 4.8 at 69.2%). Read scores as the agent-plus-model pair, because the same model scores differently in different agents. See context engineering for why scaffolding changes outcomes.