Claude Code vs Grok Build vs Codex CLI: August 2026
The Short Answer
| Claude Code | Grok Build | Codex CLI | |
|---|---|---|---|
| Vendor | Anthropic | xAI (SpaceXAI) | OpenAI |
| Default model | Claude Opus 5 | Grok 4.6 (since Aug 12) | GPT-5.6 |
| API rate /1M | $5 / $25 | $2 / $6 | $5 / $30 (Sol) |
| Reference task | $0.275 | $0.09 | $0.30 |
| Context window | 1M | 500K | Large |
| Parallel agents | Subagents | Up to 8 | Sandboxed cloud tasks |
| Best at | Hard reasoning, refactors | Cost-efficient bulk work | Autonomous delegation |
Rates verified August 17–21, 2026. Reference task = 30K input + 5K output.
What Changed Since June
Three things moved this comparison in the last ten weeks:
- Claude Opus 5 shipped July 24, 2026 at $5/$25 — same price as Opus 4.8, with a 1M context window and 128K max output. It is now the default Opus in Claude Code.
- Grok 4.6 shipped August 12, 2026 and became Grok Build’s default. 500K context, configurable reasoning levels, explicit support for long-running agents, and it landed on Amazon Bedrock shortly after.
- Codex has run GPT-5.6 since July 9, 2026, bundled into ChatGPT plans on token-based credits, with the cheaper dedicated Codex model available as a lower-cost path.
The Token Efficiency Problem
Headline rates are the wrong unit for terminal agents, because the agent decides how many tokens to spend. Independent August 2026 comparisons found Claude Code consuming far more tokens than Codex on identical tasks — one measurement reported roughly 4×, with the net cost gap landing near 23% higher per task after model rate differences.
That is not straightforwardly a defect. Claude Code re-reads files, verifies its own work and explores more before committing to an approach. On a subtle refactor that behaviour is exactly what you are paying for. On a mechanical rename across forty files, you are funding an expensive tour of your codebase.
The practical rule: token-hungry agents are worth it in proportion to how much ambiguity the task contains.
Claude Code — Best For Hard Problems
Where it wins: genuinely difficult work. Multi-file refactors with unclear boundaries, debugging where the cause is not where the symptom is, architecture decisions with real trade-offs. Opus 5’s 1M context means large codebases fit without aggressive pruning, and Claude Code’s willingness to verify its own output catches errors the cheaper agents ship.
Where it hurts: cost and limits. At $0.275 per reference task on Opus 5 it is roughly 3× Grok Build, and heavy users hit weekly usage caps on subscription plans. Anthropic extended those limits into late August 2026, but limit anxiety remains a real part of the experience on Max tiers.
Cheap fix: run Sonnet 5 as the Claude Code default at $2/$10 ($0.11 per task) and reserve Opus 5 for the problems that genuinely need it. Most teams over-provision here by habit.
Grok Build — Best Cost Per Unit Of Work
Where it wins: economics and parallelism. Grok 4.6 at $2/$6 is the cheapest frontier-tier option among these three by a wide margin, and Grok Build runs up to 8 parallel agents locally — well suited to embarrassingly parallel work like test generation, dependency upgrades or mechanical migrations across many files.
Third-party agent leaderboards in August 2026 placed Grok 4.6 at the top of at least one coding board at a reported $2.81 per task, in the same score band as Claude Fable 5 and Opus 5 at dramatically lower cost. Treat single-board results with caution, but the price advantage is not in dispute.
Where it hurts: the 500K context window is half of Opus 5’s, which matters on genuinely large monorepo work. The ecosystem is younger, and the vendor relationship carries more governance friction in some organisations than Anthropic or OpenAI.
Codex CLI — Best For Delegation
Where it wins: the “hand it off and walk away” model. Codex runs tasks autonomously in a sandboxed cloud environment, and surfaces span the ChatGPT web app, the CLI, VS Code and a macOS desktop app. If your team already pays for ChatGPT seats, Codex’s marginal cost is effectively zero until credits run out — which is a stronger economic argument than any per-token table.
Its token efficiency is the best of the three, so the effective cost gap versus its list rate is smaller than it appears.
Where it hurts: the sandbox is a real constraint when work needs local services, and credit-based metering makes forecasting harder than a flat per-token rate.
How To Choose
Default to Codex if you already pay for ChatGPT seats and your work is well-specified. Best marginal economics for most teams.
Default to Claude Code if your work is hard rather than voluminous, and switch the model to Sonnet 5 unless a task actually needs Opus 5.
Add Grok Build for bulk parallel work. At roughly a third of Claude’s per-task cost with 8 concurrent agents, it is the right tool for migrations, test backfill and dependency sweeps.
Run all three if you can. They install independently, they operate on the same working directory, and the switching cost is a shell alias. The only real overhead is keeping each agent’s instructions file in sync — which is a good discipline regardless.