AI Coding Assistants in 2026

An authorization patch can pass its tests and still grant access after an admin permission expires. That failure is easy to miss when a timestamp uses the wrong unit.

The backend still has to verify a JSON Web Token (JWT), load the user's role in the requested workspace, and reject expired permissions. If an assistant helps, decide first how much of that loop it may own and what evidence must come back before merge. This comparison uses an illustrative authorization task, not a measured head-to-head product benchmark.

Pick an operating mode before you pick a vendor:

• Autocomplete or next edit: You drive the change. The tool predicts the next local edit.
• Live agent: You stay in an interactive loop. The tool searches, edits, and runs permitted commands while you can steer.
• Delegated agent: You hand off a bounded task and review the result later. Execution might be local or remote; delegation alone says nothing about its security isolation.

One product can offer all three. Code Completion System explains why inline completion and task agents need different latency, context, and evaluation designs.

As of September 2, 2026, Cursor, Codex, GitHub Copilot, and Claude Code each cover more than one surface. Their labels won't tell you where you can steer, what runs remotely, or which evidence returns with the patch. Model menus, usage meters, and plan names move faster than the habits that keep a security-sensitive change reviewable.^{[1]Reference 1Cursor Pricinghttps://cursor.com/pricing[2]Reference 2Worktrees - Codex apphttps://developers.openai.com/codex/app/worktrees[3]Reference 3Pricing - Codexhttps://developers.openai.com/codex/pricing[4]Reference 4GitHub Copilot Planshttps://docs.github.com/en/copilot/get-started/plans[5]Reference 5GitHub Copilot cloud agenthttps://docs.github.com/en/copilot/concepts/agents/cloud-agent/about-cloud-agent[6]Reference 6Claude Code overviewhttps://code.claude.com/docs/en/overview[7]Reference 7Plans & Pricing | Claudehttps://claude.com/pricing}

Run one task through the agent loop

A weak prompt says, "add admin settings auth." A useful brief gives the agent a bounded change and a way to prove it:

• Reuse the JWT middleware pattern in app/middleware/
• Read db/schema.ts before adding queries
• Require active admin membership in the workspace named by the route
• Reject missing or expired admin permission, including the exact expiry boundary
• Leave token verification, password hashing, dependencies, and other routes unchanged
• Add tests for a valid token, an expired token, missing membership, wrong-workspace membership, and expired permission

The brief names scope, constraints, security rules, and expected evidence. It also gives the agent a reason to reuse the auth system you already have instead of inventing a second one.

Assemble the smallest useful context

Before it edits, ask which files can answer this task. The agent shouldn't dump the whole repository into context. Search for the smallest working set:

• db/schema.ts for membership state and timestamp representation
• app/routes/admin-settings.ts for the workspace identifier and response contract
• app/middleware/ for existing token verification and auth conventions
• package.json for test commands and installed libraries
• tests/ for fixtures and runner style
• AGENTS.md, CLAUDE.md, or other project rules

Claude Code reads CLAUDE.md across terminal, IDE, desktop, and web surfaces.^{[6]Reference 6Claude Code overviewhttps://code.claude.com/docs/en/overview} Cursor documents project, team, user, and AGENTS.md rules.^{[8]Reference 8Ruleshttps://cursor.com/docs/rules} Instruction files keep constraints visible, but relevant source and tests still beat a stale rule.

Patch, test, and consume the failure

Once that working set exists, the agent can propose a small plan: reuse the JWT helper, query workspace membership, check admin_expires_at, wire the route, and add failure-mode tests.

Suppose the first run fails:

For this example, the database stores expiry in milliseconds. A broken check compares it with current time in seconds: expiresAtMs > nowMs / 1000. An already-expired timestamp near 1.8 trillion still exceeds a current timestamp near 1.8 billion, so the request is allowed. The correct comparison uses milliseconds on both sides and rejects equality.

This checks one pure predicate, not the route or JWT verifier. The agent must also rerun the request-level test and existing auth checks. A green run establishes those tested behaviors, not complete security coverage. Save exact commands, exit codes, and skipped checks rather than accepting a summary that says only "tests passed."

Keep policy with the human reviewer

Now inspect the green diff against the brief. It can still remove signature verification, weaken an assertion, or query admin membership without filtering by the requested workspace. A user who administers workspace A must not gain access to workspace B. Check that query and its negative test yourself. Reject unrelated dependency or password-hashing changes; this task doesn't authorize them.

⚠️ Review boundary: Passing tests don't prove policy compliance. You still own security posture, dependency choices, permission grants, architecture, and release judgment.

The loop closes on evidence: a task contract, a selected working set, a patch, test output, and a reviewable diff. That evidence packet is what we can compare across products.

Compare surfaces, not logos

Current products cross local and remote boundaries. Compare the handoff inside the mode you actually plan to use.

Surface details change the result. GitHub's cloud agent works on one repository, one branch, and one pull request per task, uses Actions minutes plus GitHub AI Credits, and has a 59-minute hard limit. It's available on paid Copilot plans, not Copilot Free.^{[5]Reference 5GitHub Copilot cloud agenthttps://docs.github.com/en/copilot/concepts/agents/cloud-agent/about-cloud-agent[4]Reference 4GitHub Copilot Planshttps://docs.github.com/en/copilot/get-started/plans} An open-ended migration needs smaller handoffs or another surface. Codex worktrees create separate working directories on the computer or remote development environment holding the project.^{[2]Reference 2Worktrees - Codex apphttps://developers.openai.com/codex/app/worktrees} A worktree is not a security sandbox: it separates edits, not operating-system permissions, network access, or secrets.

💡 Key insight: The interesting product difference isn't "editor vs terminal." It's when you can still steer, and what evidence you get when you can't.

Treat those as routing rules, not a leaderboard. Steer when feedback is cheap, delegate when work can be bounded, use a pull request when team visibility matters, and choose terminal access when inspection and tool composition matter. Pilot the boundary with the same task before you make a broader choice.

Read price as a meter, not a task quota

Vendor units aren't interchangeable. A "5x usage" multiplier, GitHub AI Credits, included model usage, and API tokens measure different things. Context size, model choice, tool calls, local versus cloud execution, and retries can change consumption for the same task.

These are USD list prices checked on September 2, 2026, before tax, negotiated contracts, or temporary promotions. The individual column uses monthly billing; annual discounts are not mixed into it. For example, Claude Pro is also sold for $200 upfront per year, displayed as about $17/month. Availability and checkout prices can depend on region and organization type.^{[7]Reference 7Plans & Pricing | Claudehttps://claude.com/pricing}

GitHub's plans page currently says new Business and Enterprise self-serve purchases were paused on April 22, with reopening planned; verify checkout rather than treating the advertised price as proof you can buy today.^{[4]Reference 4GitHub Copilot Planshttps://docs.github.com/en/copilot/get-started/plans} Cursor's team pricing also adds a Cursor Token Rate to third-party model usage, so a model's API price alone is not the full team usage rate.^{[10]Reference 10Models & Pricinghttps://cursor.com/docs/models-and-pricing} Codex API-key access is a separate billing path and doesn't unlock every cloud feature.^{[3]Reference 3Pricing - Codexhttps://developers.openai.com/codex/pricing}

Don't convert a plan multiplier into a promised task count. Run the same repository tasks for one billing cycle, record usage after each task, and compare cost per accepted change. Record the exact model and mode too. Vendors offer multiple models or routing profiles, and a cheaper or faster choice can change both quality and consumption.

Privacy and permissions belong in the purchase decision

Price tells you about budget. Security posture depends on the plan and the settings, not the product name alone.

"Not used for training" doesn't mean zero retention, offline execution, or no third-party processing. Check retention, subprocessors, regional requirements, and whether the selected surface follows the policy you purchased. Separately inspect repository read scope, command execution, network access, secrets, branch protection, and the ability to push or open a pull request. Code Generation & Sandboxing covers isolation, allowlists, and execution boundaries.

🎯 Rollout tip: Use a disposable checkout for edits and a separately configured sandbox for untrusted execution. Don't mount production credentials. Grant only the commands, network destinations, and write actions the task needs.

Pilot with repository evidence

Once the boundary is clear, run a pilot. Give each candidate a fresh checkout of the same starting commit, the same brief, allowed tools, checks, and time budget. Preinstall equivalent dependencies so setup differences don't dominate. Repeat tasks, vary candidate order, and give equal human guidance; log every intervention.

Choose the question before the model. A product-default comparison tests each product as sold, including its model routing. A matched-model comparison holds model and reasoning settings fixed where supported, so differences more closely reflect tools and context assembly. They answer different questions. Record product version, plan, model identifier or Auto routing, mode, and date. Don't label an Auto-routed run as a test of one named model.

Score the accepted outcome, diff size, review time, tests actually run, risky-command approvals, policy preservation, task duration, and measured cost. Evaluating AI Agents shows how to grade outcome, process, safety, cost, and repeatability as one episode.

Count failures in the cost calculation

Define an accepted change before the run: required checks pass, no scope violation remains, and a reviewer accepts the patch. Include the cost of retries and rejected attempts in the numerator, but only accepted changes in the denominator. With zero accepted changes, cost per accepted change is undefined, not zero.

For a small illustrative pilot, setup A spends $20 on ten attempts and produces five accepted changes: $4 per accepted change. Setup B spends $30 on the same ten attempts and produces eight: $3.75 per accepted change. A is cheaper per attempt, but B is cheaper per accepted result. These invented numbers demonstrate the calculation; they are not vendor results.

For subscription plans, distinguish incremental cash spend from an allocated plan cost. Included usage is not a per-task invoice. Declare how you allocate the seat fee and shared chat usage; keep estimated API-equivalent costs separate from actual charges. Report reviewer minutes and wall-clock time alongside dollars, or use an explicit hourly rate if combining labor with tool cost. Anthropic's cost documentation makes the same distinction between subscription limits and dollar-metered usage.^{[11]Reference 11Manage costs effectively - Claude Code Docshttps://code.claude.com/docs/en/costs}

SWE-bench originally turned real GitHub issues into repository-level test tasks.^{[14]Reference 14SWE-bench: Can Language Models Resolve Real-World GitHub Issues?.https://arxiv.org/abs/2310.06770} Such a score is useful evidence for that setup, not a direct measurement of your team's review time or cost. Record the benchmark variant, scaffold, model, tool access, and evaluation date. Don't compare scores produced under incompatible setups.

The purchase decision should come from repeatable work in your repositories. Save the prompt, starting commit, expected checks, final diff, usage, and review notes. Rerun that pack after major model, product, or pricing changes.

Tool menus will change. Operating discipline survives: define the task, bound the environment, keep evidence visible, and keep a human on policy and release. AI Coding Workflow with Agents turns that discipline into a repeatable repository workflow.