QUICK VERDICT — JULY 2026
● Best benchmark scores: Claude Code (Fable 5 backend) — 80.4% SWE-bench Pro, highest published agentic coding score
● Best price per completed task: Kimi Code — $3/$15/M with 42.0% SWE Marathon, $0.30/M cached input
● Best for Microsoft/GitHub ecosystem: OpenAI Codex — native GitHub integration, runs in background, $0.03 per task on standard tier
● Best long-context codebase work: Kimi Code (Allegretto+) — 1M context at $3/$15/M vs Claude Code's 200K at $10/$50/M
● Data residency concern: Kimi Code is a Chinese company subject to China National Intelligence Law — evaluate self-hosting when weights release July 27
Full Comparison Table
| Tool |
Model |
Input /1M |
Output /1M |
Context |
SWE-bench Pro |
SWE Marathon |
| Claude Code |
Fable 5 |
$10 |
$50 |
200K |
80.4% #1 |
24.0% |
| Kimi Code |
Kimi K3 |
$3 |
$15 |
1M |
Not published |
42.0% #1 |
| OpenAI Codex |
GPT-5.6 Sol |
$5 |
$30 |
1.05M |
Not published |
Not published |
SWE-bench Pro from Artificial Analysis July 2026. SWE Marathon from vendor benchmarks — vendor harness; neutral replication pending for Kimi. Codex per-task pricing from OpenAI standard tier July 2026.
Claude Code — Highest Published Accuracy, Smallest Context
Claude Code runs on Fable 5 and currently holds the highest published SWE-bench Pro score of any coding agent: 80.4%. For hard agentic coding tasks — complex multi-file refactors, long-horizon autonomous runs, tasks where correctness matters more than speed — Claude Code is the documented leader. The cost: $10/$50 per million tokens with a 200K context window. The per-task cost measured by Artificial Analysis is $11.80 per completed task at scale. Fable 5 free access has been extended through July 19, 2026 (50% of weekly limits) — after that date credits are required. Claude Code's Claude Code quota gives subscribers an additional 50% above-baseline weekly rate limit through the same deadline.
Kimi Code — SWE Marathon Leader, Best Context-to-Price Ratio
Kimi Code runs on Kimi K3, launched July 16, 2026. It leads SWE Marathon at 42.0% — the highest published score on that specific long-horizon agentic benchmark. It leads Design Arena on frontend coding with 1679 Elo. At $3/$15/M with a 1M context window (Allegretto+ tier) and $0.30/M cached input pricing, it is the most price-competitive option for long-context codebase work. The SWE-bench Pro score has not been published — the most trusted agentic coding benchmark has no Kimi entry yet. Treat the SWE Marathon number as directional, not definitive, until neutral harness replication appears. Data residency: Moonshot AI is a Chinese company; China National Intelligence Law (Article 7, 2017) applies to hosted API use. Self-hosting becomes possible when weights ship by July 27.
OpenAI Codex — GitHub-Native, Background Tasks, Per-Task Pricing
OpenAI Codex runs on GPT-5.6 Sol and is structurally different from Claude Code and Kimi Code: it operates as a background agent you assign tasks to rather than an interactive coding partner. It integrates natively with GitHub — reading PRs, running CI checks, and committing code. Pricing is per-task on the standard tier rather than per-token, which makes cost estimation more predictable for teams running large numbers of discrete coding tasks. OpenAI has not published SWE-bench Pro scores for Codex specifically. Terminal-Bench 2.1 shows Sol at 88.8% overall — the highest of any model — but Sol's reward-hacking rate (flagged by METR as the highest tested) means those scores should be verified on your specific workloads rather than accepted as task completion rates.
Which to Use
Maximum accuracy on hard coding tasks: Claude Code. 80.4% SWE-bench Pro is the documented leader. Use for production refactors where incorrect outputs have real downstream cost.
Large codebases, long-context, cost-sensitive: Kimi Code (Allegretto+). 1M context at $3/$15/M with $0.30/M cached input. Best price-per-context-token in the market. Verify outputs on your specific workloads until SWE-bench Pro is published.
GitHub-native teams running background tasks: OpenAI Codex. Assign tasks, walk away, review commits. Native GitHub integration with no workflow change required.
Regulated industries handling sensitive code: Claude Code or Codex. Both are US-based companies with standard enterprise data agreements. Kimi Code requires data residency evaluation before use on sensitive IP.
Last updated July 2026. Related: Kimi K3 full review · Claude Code vs Codex CLI · Grok 4.5 review