Codex vs Claude Code vs Kimi Code

Choose the coding tool for its workflow, controls, and billing model. Choose the underlying model separately. Tool subscriptions, API token prices, and public benchmark scores are not interchangeable.

Quick Decision

Current recommendation: use the tool that fits your repository controls, then run the same real task through the model lanes you can actually access. Do not select a tool from an old SWE-bench row alone.

Current Model Map

The model exposed by a subscription or tool-facing alias can differ from the public API model discussed in a benchmark. Confirm the exact model ID or account UI instead of inferring it from the product name.

Model Evidence

Useful evidence has three levels:

1. Independent: a named third party publishes a methodology and model-specific result.
2. Vendor: the provider publishes a result or relative improvement; useful for deciding what to test, not for adopting the claim as an AIHackers ranking.
3. Site-owned: the same repository task, harness, completion rules, and artifacts are available for review.

AIHackers does not yet have a controlled cross-tool repository evaluation for current Codex, Claude Code, and Kimi Code. Claims such as “95% of another model’s capability,” “best quality,” or “8x better value” are therefore not supported here.

Pricing: Keep Three Lanes Separate

Tool subscriptions

Codex, Claude Code, and Kimi Code expose plan limits, quotas, credits, or checkout-controlled offers. Those entitlements can change independently of API list prices. Verify the live plan page and account before purchase.

API token prices

Prices are per 1 million tokens. They do not include tool subscription fees, retries, cache writes, failed patches, or human review.

Cost per successful task

This is the metric that matters. Record:

• input, cache writes/reads, and output tokens;
• wall-clock time and retries;
• whether tests pass;
• whether the patch is accepted without repair;
• human review and cleanup time.

A cheaper token price can lose if a model produces longer outputs, retries more often, or requires expensive review. A premium model can be rational when it prevents a costly mistake.

Workflow Comparison

No tool makes an autonomous agent safe by default. Apply least-privilege credentials, scoped worktrees, confirmation gates for destructive actions, and independent test/diff review.

Evaluation Protocol

Run the same four tasks:

Pin the tool version, model ID, reasoning setting, repository commit, prompt, and acceptance rules. If a tool hides the exact model, record that as an evaluation limitation.

Current Verdict

• Choose Codex for OpenAI-native cloud-agent workflows after confirming the current model and plan limits.
• Choose Claude Code when Claude-native terminal work and an Opus 5 premium review lane matter.
• Choose Kimi Code/K3 when you want the newest Kimi flagship and can measure CAR; choose K2.7 Code when low Kimi API pricing fits.
• Route GPT-5.6 by workload and plan; use Trusted Access only for eligible cyber work. Keep Mythos 5 behind its access checks and test restored Fable only as a measured escalation after Opus.
• Keep Opus 4.5, Kimi K2.5/K2.6, and older GPT rows only as historical or product-specific context.

Sources

• OpenAI: GPT-5.6 general availability and system card
• OpenAI: Codex
• Anthropic: Claude model overview and pricing
• Kimi: K3 launch blog, K3 quickstart, K2.7 Code quickstart, and K2.7 pricing
• Artificial Analysis: Claude Opus 5, Claude Opus 4.8, and GLM-5.2

Related links

• /models/gpt-5-6/
• /models/claude-opus-5/
• /models/claude-opus-4-8/
• /models/kimi-k3/
• /models/kimi-k2.7-code/
• /tools/kimi-code/
• /value/smart-spend/
• /risks/codex/cloud-dependency-risks/

Last verified: July 26, 2026. Tool plans, model pickers, access, quotas, open-weight status, and API prices can change independently.