Nine AI CLI coding agents now compete for your terminal in July 2026: Claude Code, Grok Build, Gemini CLI, Codex CLI, GLM 5.2, OpenCode, Qwen Code, Kimi CLI, and DeepSeek TUI. Grok Build (Grok 4.5, 83.3% Terminal-Bench) and GLM 5.2 (81.0% Terminal-Bench, MIT license) are the two entrants that actually change the rankings.
I have spent the last week running builds through most of these. The field has genuinely changed since spring. Two new entrants -- Grok Build and GLM 5.2 -- post benchmark numbers that belong in the top tier. Here is what the data actually shows, and where each tool wins on real work.
What are the 9 AI CLI coding agents competing in 2026?
The nine active AI CLI coding agents in July 2026 are Claude Code (Anthropic), Grok Build (xAI), Gemini CLI (Google), Codex CLI (OpenAI), GLM 5.2 (Zhipu / Z.ai), OpenCode (open-source TUI), Qwen Code (Alibaba), Kimi CLI (Moonshot AI), and DeepSeek TUI. Each ships as a terminal-native tool you install locally and connects to a model backend via API or local weights.
The split into three categories makes this easier to navigate. Vendor-native CLIs tie you to one model provider: Claude Code to Anthropic, Codex CLI to OpenAI, Gemini CLI to Google, Grok Build to xAI, Qwen Code to Alibaba, Kimi CLI to Moonshot. Open-source model-agnostic tools -- OpenCode and DeepSeek TUI -- let you swap backends mid-session. GLM 5.2 is the outlier: it is an open-weight model (MIT license, full weights downloadable) that any CLI can call.
Chinese open-weight models including GLM 5.2, Kimi K2.5, and Qwen 3.6 accounted for 41% of HuggingFace downloads in spring 2026, according to TechCrunch reporting on the open-model surge. That share is not going down.
How do they rank on real coding benchmarks?
On Terminal-Bench 2.1, the current ranking is: Claude Opus 4.7 Adaptive 87.6% (SWE-Bench leaderboard), Grok 4.5 83.3%, GLM 5.2 81.0%, Codex 83.4%, Claude Code (Sonnet 5 default) 78.9%, Gemini CLI 70.7%. On SWE-Bench Pro, Grok 4.5 posts 64.7%, GLM 5.2 62.1%, GPT-5.5 58.6%.
The Codex CLI number (83.4% Terminal-Bench) is real, but worth contextualizing. That score runs on Codex o3, OpenAI's highest-tier reasoning model. If you are on ChatGPT Plus at 0/month and using Codex CLI without an API key, you are running a throttled tier -- not the benchmark model. Claude Code at 78.9% uses Sonnet 5 as the default, which is also not Anthropic's top model. Benchmark numbers and default-tier numbers are two different things.
For Grok Build specifically: xAI published 70.8% on SWE-Bench Verified for grok-code-fast-1, the fast variant. The full Grok 4.5 model achieves 64.7% on SWE-Bench Pro, a harder benchmark. Both numbers place Grok 4.5 solidly in tier 2 on pure coding, behind Opus 4.7 Adaptive but ahead of GPT-5.5's 58.6% on the same test.
What does Grok Build actually bring to the field?
Grok Build is xAI's terminal-native coding agent, launched May 2026 and upgraded to Grok 4.5 on July 8. It runs parallel subagents by default, ships with a Plan Mode that blocks file edits until you approve the plan, and supports MCP out of the box. API pricing for Grok 4.5 is per million input tokens and per million output tokens -- cheaper than Sonnet 5 input, comparable output.
What Grok Build gets right: the Plan Mode default is genuinely useful. Instead of watching the agent make changes while you read a plan you have not approved, Grok Build stops and waits. Claude Code has Plan Mode too, but it is not the default. For new users who do not know to set it, Grok Build's conservative default prevents a lot of unintended file mutations.
What it gets wrong: the Grok 4.5 context window is 500K tokens, half of Gemini CLI's 1M and significantly smaller than GLM 5.2's 1M. For large monorepos, that ceiling matters. The MCP support is also version-pinned to the pre-July-28 spec for now -- the stateless RC is not yet supported, which will require an update when the final July 28 spec drops.
Get the AI Agent Briefing
One email per week. The best AI agent news, tutorials, and tools -- written by someone who actually builds with them.
Subscribe Free
How does Claude Code still differentiate itself?
Claude Code's primary differentiator is the Agent Skills system (SKILL.md), an open standard Anthropic released in December 2025 that lets you install reusable agent playbooks across Claude Code, Cursor, Codex CLI, Gemini CLI, and GitHub Copilot. No other vendor has shipped a comparable portable skill layer. You install a SKILL.md file once; every compatible agent can invoke it.
I use this daily. The practical effect is that Claude Code knows your team's review checklist, your deployment sequence, your repo conventions -- without you re-explaining them in every session. That institutional memory compounds. OpenCode and Grok Build have no equivalent mechanism yet. Codex CLI adopted the SKILL.md spec, but the library of skills is still much smaller than the Claude Code ecosystem.
The second differentiator is nested multi-agent architecture. Claude Code supports 5-level-deep subagent chains where each agent spawns its own subagents. Grok Build runs parallel subagents but caps at one level of nesting. For complex workflows -- parallel test writing while refactoring runs -- Claude Code's architecture has more room to operate.
The third is the Claude Code Max tier economics. At 00/month for Max 5x or 00/month for Max 20x, you get high-volume API access bundled. Compared to paying per million Sonnet 5 input tokens on the API, the Max tier makes sense above roughly 15-20 hours of active use per month. No other vendor bundles comparable compute at that price point yet.
Which open-source CLI tools are actually viable for production?
GLM 5.2 and OpenCode are the two open-source options ready for serious production use. GLM 5.2 is a 744B-parameter Mixture-of-Experts model (approximately 40B active parameters per token) with a 1M-token context window, MIT license, and full downloadable weights. OpenCode supports 75+ LLM endpoints including Anthropic, OpenAI, Google, AWS Bedrock, Azure OpenAI, and local Ollama servers, with mid-session model switching and context carryover.
GLM 5.2's cost case is strong. API access through Z.ai runs .40 per million input tokens and .40 per million output tokens. Opus 4.8 runs approximately /5 per million. For teams that want near-frontier coding performance at one-sixth the API cost, GLM 5.2 is the current best answer. Semgrep published an independent benchmark showing GLM 5.2 beating Claude on their internal cyber engineering benchmarks -- not a claim to take at face value without reading their methodology, but worth noting as independent third-party data.
Qwen Code (Alibaba, Apache 2.0) is worth watching but not ready to replace a primary workflow yet. It is a fork of the Gemini CLI codebase optimized for the Qwen 3.6 480B MoE model. Kimi CLI from Moonshot AI runs on K2.5 with 256K context and approximately 100 tokens per second output -- fast, but the context ceiling limits it for large-repo work.
What does the cost breakdown actually look like across all 9 tools?
Pricing in July 2026 ranges from free-tier access to 00/month all-in. Gemini CLI offers a generous free tier (though Google cut it from 1,000 to 20 requests per day for heavy users in late 2025). Codex CLI comes bundled in ChatGPT Plus at 0/month, but the benchmark model (Codex o3) is API-only at /0 per million tokens input/output. Claude Code Pro starts at 0/month; Max 5x is 00/month; Max 20x is 00/month.
Grok Build API (Grok 4.5) costs / per million tokens. GLM 5.2 via Z.ai costs .40/.40. OpenCode is free software; you pay whatever model backend you connect. DeepSeek TUI runs on DeepSeek-V4 API, which remains one of the cheapest frontier options at prices DeepSeek publishes on their API pricing page.
For most solo builders and small teams, the practical decision is: if you want the best skills ecosystem and are already in the Anthropic stack, Claude Code Max 5x at 00/month is the value tier. If you want the best open-weight cost curve, GLM 5.2 via OpenCode gives you MIT-licensed weights at .40 input per million. If you want the newest serious entrant with a good default UX, Grok Build is worth a week on the free trial before committing to API costs.
FAQ
Which AI CLI coding agent has the best benchmark scores in July 2026?
On Terminal-Bench 2.1, Grok 4.5 scores 83.3% and GLM 5.2 scores 81.0% as of July 2026 -- both above Claude Code's default Sonnet 5 tier at 78.9%. Claude Opus 4.7 Adaptive leads the SWE-Bench leaderboard at 87.6%, but that is not the default Claude Code model tier. Benchmark scores vary significantly by which model tier each CLI uses.
Is Grok Build a serious alternative to Claude Code?
Yes, for most coding workflows. Grok Build's Plan Mode default, parallel subagents, and Grok 4.5 benchmark scores (83.3% Terminal-Bench, 64.7% SWE-Bench Pro) place it firmly in tier 2. Its weaknesses are a 500K-token context ceiling versus Claude Code's higher tiers, no equivalent to the Claude Code Skills (SKILL.md) ecosystem, and one level of subagent nesting versus Claude Code's 5-level chains.
What makes GLM 5.2 different from other open-weight coding models?
GLM 5.2 from Zhipu AI ships 744B parameters as a Mixture-of-Experts model with 1M-token context, MIT license, full downloadable weights, and a public API at .40/.40 per million tokens. It scores 62.1% on SWE-Bench Pro, beating GPT-5.5's 58.6%, at roughly one-sixth the cost. The MIT license is the key detail -- you can fine-tune and self-host it without restrictions.
Can I use Claude Code Skills (SKILL.md) with other CLI agents?
Yes. Anthropic open-sourced the Agent Skills specification in December 2025. The same SKILL.md format now works across Claude Code, Cursor, Codex CLI, Gemini CLI, and GitHub Copilot. OpenAI adopted the spec within weeks of the open-source release. You install a SKILL.md file once and any compatible agent can invoke it -- the skill library is currently largest on the Claude Code side but growing across tools.
Get the AI Agent Briefing
One email per week. The best AI agent news, tutorials, and tools -- written by someone who actually builds with them.
Subscribe Free