A dated snapshot, deliberately: this domain decays faster than any other in this wiki. Everything below was verified against public sources in August 2026. If you are reading this more than ~6 months later, treat specific model names, scores, and prices as historical and re-verify.
The defining fact of 2026: frontier models have largely converged for mainstream coding tasks, and the harness — the agent's scaffolding of context management, tool access, permissions, verification, and orchestration — now does most of the differentiating work. Reviewers consistently reach the same conclusion: there is no single best coding agent, only best-per-dimension. Claude Code has the deepest programmable harness; Codex spans the most surfaces at the lowest entry price; Cursor is the strongest in-editor agent; Devin is the most autonomous.
The practical consequence: switching models inside a good harness moves results less than switching workflows inside the same harness. That is why most of this cluster is about workflow, not tools.
CLI agents live where the build, test, and git tooling lives, which makes them the natural home for serious multi-step work.
| Agent | Vendor | Distinguishing strength |
|---|---|---|
| Claude Code | Anthropic | Deepest programmable harness: subagents, hooks, skills, MCP, deterministic workflow scripts; 1M-token context on Claude 5-generation models |
| Codex CLI | OpenAI | GPT-5.x models, AGENTS.md-based context convention, multi-agent worktrees, low entry price |
| Gemini CLI | Generous free tier, reads the same skills standard, tight Google Cloud integration | |
| OpenCode | open source | Model-agnostic, free, terminal-UI-first; popular self-host choice |
| Aider | open source | The original git-native pair programmer; still excellent for surgical, diff-oriented edits |
Asynchronous delegation — hand the agent an issue, get back a pull request — matured into its own category around 2025–2026. Devin 2.0 dropped its entry price from $500/mo to ~$20/mo, moving autonomous cloud agents from enterprise-only to individually trialable. Google Jules runs an issue-driven async workflow tied to your repository; Codex Cloud, Cursor Cloud Agents, GitHub Copilot's coding agent (works inside a GitHub Actions sandbox), and Claude Code on the web all run async and open PRs. Details and the local-vs-cloud decision live in BackgroundAgentsAndAsyncDelegation.
AI review became a distinct product category rather than a feature. As of 2026: Claude Code Review (launched March 2026) runs a multi-agent architecture — parallel finder sub-agents plus a verification pass per finding — producing the highest-precision comments at token-based pricing; CodeRabbit crossed 13 million PRs reviewed and integrates the most platforms, with a CLI that reviews code as agents generate it, before the PR exists; Copilot code review wins on frictionless activation inside GitHub. Greptile and Qodo compete on codebase-context depth. How review fits the workflow — including pre-PR review of agent output — is covered in VerificationLoopsForAgenticCoding.
| Model | Vendor | SWE-bench Verified | Notes |
|---|---|---|---|
| Claude Fable 5 / Mythos 5 | Anthropic | ~95% | Claude 5 family (July 2026); also leads SWE-bench Pro (~80%); Mythos is the restricted-availability variant |
| GPT-5.5 | OpenAI | ~88.7% | Leads Terminal-Bench 2.0 (~82.7%); GPT-5.6 family (Luna/Terra/Sol tiers) rolling out |
| Claude Opus 4.8 | Anthropic | ~88.6% | The workhorse tier below the Claude 5 family |
| Gemini 3.1 Pro | ~80.6% | Strong multimodal + long-context | |
| DeepSeek V4 Pro | DeepSeek (open weights) | ~80.6% | Open-weights parity with Gemini 3.1 on this benchmark |
| Kimi K3, MiniMax M3 | Moonshot / MiniMax (open weights) | ~80% | Open-weights models now routinely inside 15 points of the frontier |
Two caveats. First, SWE-bench Verified is saturating and contamination-suspect; the field is moving to SWE-bench Pro (multi-language, standardized scaffold) and Terminal-Bench 2.0, which still discriminate. Second, benchmark deltas overstate felt differences: in a good harness with verification loops, the gap between an 88% model and a 95% model shows up mainly on the hardest long-horizon tasks. See AiEvaluationAndBenchmarks for how to build task-specific evals instead of trusting leaderboards.
Cost discipline in 2026 means routing by task difficulty, not defaulting to the flagship: a frontier model (Fable 5, GPT-5.5-class) for architecture, planning, and the hardest debugging; a mid-tier model (Sonnet-class, GPT-5.6 Terra-class) for multi-file implementation against an existing pattern; a small fast model (Haiku-class, GPT-5.6 Luna-class) for mechanical single-file work with a complete spec. Subagent-capable harnesses let you set the tier per delegated task — see SubagentOrchestrationPatterns.
The common professional pattern is two or three specialized agents at once: a terminal agent for deep multi-step work and refactors, an IDE agent for interactive daily flow, and a cloud delegate for asynchronous background tasks. Pick each dimension deliberately: terminal depth (Claude Code), cross-surface continuity (Codex), in-editor speed (Cursor), or hands-off delegation (Devin, Jules, Copilot coding agent). Open-source/self-host requirements point to OpenHands or OpenCode.