Claude vs GPT for Coding: Same Mid-Tier Price, a 2.5× Flagship Gap, Benchmarks Measured Differently

The mid-tier models most people actually code with are priced identically: Claude Sonnet 5.5 and GPT-6 Sol (plus GPT-6.1 Sol, released in October) both cost $2 per million input tokens and $10 per million output tokens. Anthropic also says Sonnet 5.5 finishes the same work with fewer tokens, cutting per-task cost by roughly 30% versus the previous generation — that is the vendor’s claim, and real savings depend on your workload.

The gaps are at the ends. At the flagship tier Claude is cheaper: Opus 5.5 costs $4 / $20 versus $10 / $50 for GPT-6 Astra — about 2.5×. At the entry tier it flips: GPT-6 Luna is $0.10 / $0.50, while Claude’s cheapest model, Haiku 4.5, is $1 / $5 — a tenfold difference.

Both labs ship their own coding agent — Anthropic’s Claude Code and OpenAI’s Codex — and both work directly in your repository from the terminal and the IDE. The entry points differ: ChatGPT Free includes limited Codex use and the $8 Go plan covers lightweight tasks, while Claude Free does not include Claude Code; you start at the $20/month Pro plan. At $20 both plans include the coding agent, with usage limits that vary by model and task size.

Anthropic reports 70.6% for Sonnet 5.5 on Terminal-Bench 4.0. OpenAI has not published a verifiable score for the GPT-6 family on the same benchmark and version, and comparison tables circulating online often mix Terminal-Bench 2.x with 4.0 and different test harnesses. Self-reported scores from different versions and harnesses cannot be ranked directly, so this article's table records that row as a fact and declares no winner.

• Everyday coding and bug fixes: Sonnet 5.5 or GPT-6 Sol — same price, so pick the one your team and toolchain already use.
• Large refactors, hard problems, long tasks: if you need a flagship, Opus 5.5 costs about 2.5× less than GPT-6 Astra — a real difference at volume.
• Tight budget, students, or just trying it out: GPT offers free Codex access and the very cheap Luna tier — the lowest cost to start.
• Testing it yourself: run the same batch of tasks on your own repo with each for a week, and compare how many changes pass your tests and what they cost.