Does the harness matter more than the model?

- Below the frontier, no: a weak model fails regardless of harness. - At the frontier, yes: context strategy, tool design and failure feedback decide completion rate. - The same model in Claude Code and in a bare chat window are not the same product.

Which coding agents can open a pull request on their own?

- Rung 4: Claude Code (hosted cloud sessions and the Claude GitHub App via /install-github-app). - Also rung 4: opencode (the opencode-agent GitHub App), Codex CLI cloud tasks, Cursor background agents, GitHub Copilot, Devin. - Gemini CLI has an official Action but no App and leans on review, so it grades partial. - Everything else (Cline, Aider, Goose, Zed, Qwen Code, Kilo Code) stops at editing and running tests locally.

Can any of them merge and deploy without a human?

- Not out of the box; treat any claim otherwise as a description of your own CI. - Copilot's agent is explicitly forbidden from approving its own pull request, and Devin stops at review. - Claude Code and Codex will merge only with a privileged token and approvals disabled, which is you removing the gate.

Are Terminal-Bench scores comparable between harnesses?

- Only when harness version, model and benchmark version are all stated; the column is empty because nothing traceable met that bar. - Terminal-Bench 2.0 results are not comparable to 1.0, and a harness can move several points between releases. - A figure from the benchmark's own reference agent is a model result, not a harness result. - Half a point on an eighty-task benchmark is a fraction of one task, not a ranking.

Is it safe to let an agent run shell commands unattended?

- Only inside a sandbox with a real filesystem and network boundary. - Codex CLI (Seatbelt on macOS, Landlock on Linux) and Claude Code ship OS-level sandboxing. - Gemini CLI and Goose ship one too, but off by default, so check the cell note. - Devin and Copilot run in vendor-hosted VMs, which we rank level with a local kernel sandbox. - Most IDE extensions offer only an auto-approve allowlist, which is convenience, not a security boundary.

Coding agents compared: Claude Code vs Codex CLI vs Cursor vs the rest Pricing Compared

Q: What is the cheapest way to run a coding agent all day?

- A flat subscription with soft limits: Claude Code on a Max plan, Codex on a ChatGPT Pro plan. - Or a free harness with your own key: opencode, Aider, Goose, Cline, Zed. - Gemini CLI and Qwen Code both have usable free tiers tied to a personal account. - Avoid credit-based pricing if you need to forecast; the credit-to-token rate is the vendor's to change.

Q: Should my team standardise on one harness?

- Standardise the repo-level configuration, not the harness. - AGENTS.md, MCP server definitions and a documented test command are portable across nearly everything here. - Skills, hooks and rules files mostly are not portable. - Then swapping harnesses next quarter costs a config change rather than a migration.

- The flat subscription is the most honest deal here, but read the limits. - Every subscription harness enforces a rolling usage window in vendor-specific units, not tokens. - Heavy agentic use burns those windows far faster than chat; then you wait or move to metered billing. - Credits are worse for planning: Windsurf, Devin and Amp price in units the vendor defines. - The vendor can re-price the credit-to-token conversion when model prices change, and you find out afterwards. - Copilot's premium-request multipliers have the same shape. - For a defensible quarterly number, use a per-seat subscription or your own provider key. - BYOK looks cheapest but moves the variance onto your API bill rather than removing it. - A harness that re-reads a large file every turn costs multiples of one that caches aggressively. - At team scale, meter BYOK per repo for a fortnight and check for prompt caching before extrapolating.

Providers: 17
Fields compared: 3
Source confidence: 35%
Last verified: 2026-01-15 (8mo ago)
Re-verified: every 14 days

Differences only

17 tools · verified 8mo ago

Coding agents and harnesses, compared, 17 tools compared across 3 of 3 fields.
Tool	Weighted
		w6	w6	w5
AiderMinimal terminal pair-programmer with a tree-sitter repo map and automatic git commits.	91.2	BYOK tokens only	$0/seat/mo	No
ClineOpen-source VS Code and JetBrains agent built around an explicit plan/act split.	91.2	BYOK tokens only	$0/seat/mo	No
DIY harnessbaseline · Your own loop over a model API with an agent SDK, the build-it-yourself reference point.	91.2	BYOK tokens only	$0/seat/mo	No
GooseBlock's Apache-2.0 agent, MCP-native from the start, as both a CLI and a desktop app.	91.2	BYOK tokens only	$0/seat/mo	No
Kilo CodeOpen-source VS Code agent in the Cline/Roo lineage, with an orchestrator mode and a CLI.	91.2	BYOK tokens only	$0/seat/mo	No
opencodeProvider-agnostic open-source TUI with a client/server split and LSP awareness.	91.2	BYOK tokens only	$0/seat/mo	No
Qwen CodeApache-2.0 Gemini CLI fork tuned for Qwen3-Coder, with a large free daily quota.	91.2	BYOK tokens only	$0/seat/mo	No
ZedGPL-licensed Rust editor whose agent panel can host Claude Code, Codex and Gemini CLI.	91.2	BYOK tokens only	$0/seat/mo	No
Gemini CLIApache-2.0 terminal agent from Google with an unusually generous free tier.	82.4	subscription + token overage	$0/seat/mo	No
Claude CodeAnthropic's terminal-first agent, with IDE, web and GitHub Action surfaces.	75.3	subscription + token overage	$20/seat/mo	No
Codex CLIOpenAI's open-source Rust agent, with the strongest local sandboxing on this page.	75.3	subscription + token overage	$20/seat/mo	No
Grok BuildxAI's coding harness, launched July 2026, on this page but not yet verified.	73.4	·	·	Unknown
GitHub Copilot (Agent HQ)Copilot's coding agent plus a control plane for running third-party agents on your repos.	61.2	subscription + token overage	$10/seat/mo	Partial
CursorVS Code fork with a real codebase index, background cloud agents and a CLI.	57.6	subscription + token overage	$20/seat/mo	Partial
AmpSourcegraph's opinionated agent with no model picker and an ad-supported free tier.	35.3	opaque credits	$0/seat/mo	Yes
WindsurfAgentic IDE built around Cascade, now owned by Cognition after a turbulent 2025.	30.0	opaque credits	$15/seat/mo	Yes
DevinFully delegated cloud engineer: file a task in Slack or the web app, get a PR back.	28.2	opaque credits	$20/seat/mo	Yes

AiderMinimal terminal pair-programmer with a tree-sitter repo map and automatic git commits.91.2

Pricing

Pricing model: BYOK tokens only
Entry price (/seat/mo): $0/seat/mo
Token markup: No

ClineOpen-source VS Code and JetBrains agent built around an explicit plan/act split.91.2

Pricing

Pricing model: BYOK tokens only
Entry price (/seat/mo): $0/seat/mo
Token markup: No

DIY harnessYour own loop over a model API with an agent SDK, the build-it-yourself reference point.91.2

Pricing

Pricing model: BYOK tokens only
Entry price (/seat/mo): $0/seat/mo
Token markup: No

GooseBlock's Apache-2.0 agent, MCP-native from the start, as both a CLI and a desktop app.91.2

Pricing

Pricing model: BYOK tokens only
Entry price (/seat/mo): $0/seat/mo
Token markup: No

Kilo CodeOpen-source VS Code agent in the Cline/Roo lineage, with an orchestrator mode and a CLI.91.2

Pricing

Pricing model: BYOK tokens only
Entry price (/seat/mo): $0/seat/mo
Token markup: No

opencodeProvider-agnostic open-source TUI with a client/server split and LSP awareness.91.2

Pricing

Pricing model: BYOK tokens only
Entry price (/seat/mo): $0/seat/mo
Token markup: No

Qwen CodeApache-2.0 Gemini CLI fork tuned for Qwen3-Coder, with a large free daily quota.91.2

Pricing

Pricing model: BYOK tokens only
Entry price (/seat/mo): $0/seat/mo
Token markup: No

ZedGPL-licensed Rust editor whose agent panel can host Claude Code, Codex and Gemini CLI.91.2

Pricing

Pricing model: BYOK tokens only
Entry price (/seat/mo): $0/seat/mo
Token markup: No

Gemini CLIApache-2.0 terminal agent from Google with an unusually generous free tier.82.4

Pricing

Pricing model: subscription + token overage
Entry price (/seat/mo): $0/seat/mo
Token markup: No

Claude CodeAnthropic's terminal-first agent, with IDE, web and GitHub Action surfaces.75.3

Pricing

Pricing model: subscription + token overage
Entry price (/seat/mo): $20/seat/mo
Token markup: No

Codex CLIOpenAI's open-source Rust agent, with the strongest local sandboxing on this page.75.3

Pricing

Pricing model: subscription + token overage
Entry price (/seat/mo): $20/seat/mo
Token markup: No

Grok BuildxAI's coding harness, launched July 2026, on this page but not yet verified.73.4

Pricing

Pricing model: ·
Entry price (/seat/mo): ·
Token markup: Unknown

GitHub Copilot (Agent HQ)Copilot's coding agent plus a control plane for running third-party agents on your repos.61.2

Pricing

Pricing model: subscription + token overage
Entry price (/seat/mo): $10/seat/mo
Token markup: Partial

CursorVS Code fork with a real codebase index, background cloud agents and a CLI.57.6

Pricing

Pricing model: subscription + token overage
Entry price (/seat/mo): $20/seat/mo
Token markup: Partial

AmpSourcegraph's opinionated agent with no model picker and an ad-supported free tier.35.3

Pricing

Pricing model: opaque credits
Entry price (/seat/mo): $0/seat/mo
Token markup: Yes

WindsurfAgentic IDE built around Cascade, now owned by Cognition after a turbulent 2025.30.0

Pricing

Pricing model: opaque credits
Entry price (/seat/mo): $15/seat/mo
Token markup: Yes

DevinFully delegated cloud engineer: file a task in Slack or the web app, get a PR back.28.2

Pricing

Pricing model: opaque credits
Entry price (/seat/mo): $20/seat/mo
Token markup: Yes

yespartialnounknown

Sources shown beside each value · Learn how sourcing works

measured vendor-claimed community inferredExpand any row for the source, verification date, and caveat behind every cell.

Is the cheapest option the right one?

Price is one group of 3 fields out of 26. Everything else, features, limits, portability, how well any of it works for agents, is on the full coding agents compared: claude code vs codex cli vs cursor vs the rest comparison.

Pricing traps

The flat subscription is the most honest deal here, but read the limits.
Every subscription harness enforces a rolling usage window in vendor-specific units, not tokens.
Heavy agentic use burns those windows far faster than chat; then you wait or move to metered billing.
Credits are worse for planning: Windsurf, Devin and Amp price in units the vendor defines.
The vendor can re-price the credit-to-token conversion when model prices change, and you find out afterwards.
Copilot's premium-request multipliers have the same shape.
For a defensible quarterly number, use a per-seat subscription or your own provider key.
BYOK looks cheapest but moves the variance onto your API bill rather than removing it.
A harness that re-reads a large file every turn costs multiples of one that caches aggressively.
At team scale, meter BYOK per repo for a fortnight and check for prompt caching before extrapolating.

How is this pricing data collected?

Every cell comes from vendor docs, the tool's repo, or hands-on use, each carrying its provenance.
Pricing-page facts are marked vendor-claimed with the URL.
Inferences from adjacent behaviour, and judgement calls like scores, are marked inferred.
Widely reported but undocumented facts are marked community.
Unknown facts are null and render as a dash rather than a guess.
Closed-source harnesses have null GitHub stars because no comparable repo exists.
Verify cadence is 14 days; the last full pass was 15 January 2026, well outside it.
We publish dates rather than refresh them; a moved date without a redone check is a lie.
Read every date as the last time a human actually looked.
The killer field is scored on a fixed five-rung ladder, at the highest rung reached out of the box.
Rung 4 needs a first-class PR path: built-in command, vendor-published GitHub App, or hosted runner.
Invoking gh because you asked it to is not rung 4; any shell can be scripted up a rung.
A vendor Action with no App identity behind it is graded partial, for first- and third-party alike.
The Terminal-Bench column ships empty because no traceable figure met the bar.
Traceable figures named a model and version without a harness, or came from the benchmark's own reference agent.
Terminal-Bench 1.0 results are superseded by 2.0 and not comparable.
Gaps between top entries were smaller than one task on an eighty-task benchmark, so not a ranking.
An empty column is the honest state until we run the benchmark against pinned harness versions.

Pricing questions

What is the cheapest way to run a coding agent all day?

A flat subscription with soft limits: Claude Code on a Max plan, Codex on a ChatGPT Pro plan.
Or a free harness with your own key: opencode, Aider, Goose, Cline, Zed.
Gemini CLI and Qwen Code both have usable free tiers tied to a personal account.
Avoid credit-based pricing if you need to forecast; the credit-to-token rate is the vendor's to change.

Should my team standardise on one harness?

Standardise the repo-level configuration, not the harness.
AGENTS.md, MCP server definitions and a documented test command are portable across nearly everything here.
Skills, hooks and rules files mostly are not portable.
Then swapping harnesses next quarter costs a config change rather than a migration.

Coding agents compared: Claude Code vs Codex CLI vs Cursor vs the rest Pricing Compared

Pricing

Pricing

Pricing

Pricing

Pricing

Pricing

Pricing

Pricing

Pricing

Pricing

Pricing

Pricing

Pricing

Pricing

Pricing

Pricing

Pricing

Is the cheapest option the right one?

Pricing traps

How is this pricing data collected?

Pricing questions

What is the cheapest way to run a coding agent all day?

Should my team standardise on one harness?

Every tool compared on this page

Related comparisons