You commented MODELS, so here's the breakdown. Anthropic released Claude Opus 5.5 and OpenAI released GPT-6 Sol and GPT-6 Luna on the same day, 22 September 2026. Both companies cut prices and both pitched their model at coding agents.
Every number below comes from the official model pages. I checked them on 24 September 2026.
The short version: Sol costs half as much as Opus 5.5 per token. Luna is the one most students should look at first, because Free and Go users get it in the Codex desktop app. Don't pick between Opus and Sol from launch-day benchmarks. Give both the same task on your own code.
The numbers
| Claude Opus 5.5 | GPT-6 Sol | GPT-6 Luna | |
|---|---|---|---|
| Input, per 1M tokens | $4 | $2 | $0.10 |
| Output, per 1M tokens | $20 | $10 | $0.50 |
| Cached input, per 1M tokens | $0.20 | $0.20 | |
| Context window | 1M tokens | 1.05M tokens | |
| Max output | 128K tokens | 128K tokens | |
| Model ID | claude-opus-5-5 | gpt-6-sol |
Sources: Claude models overview, GPT-6 Sol model page, OpenAI's launch post. The blank cells are numbers I couldn't find on an official page, so I didn't fill them in.
1. Claude Opus 5.5
Anthropic says Opus 5.5 is 30% faster than Opus 5 and costs about 40% less to run on typical workloads. Per-token prices dropped 20% and cached reads dropped 60%. The numbers they highlight for coding and agents: Terminal-Bench 4.0 at 66.4%, CursorBench 4.0 at 57.8% and OSWorld 2.0 at 81.8%.
If you use Claude Code you already have it. Opus 5.5 became the default model in Claude Code 2.1.280 (release notes).
2. GPT-6 Sol
openai.com/index/introducing-gpt-6-sol-and-luna
OpenAI's model for coding and agent work. It costs 50% less than the GPT-5.6 generation it replaces. It's in Codex for Plus, Pro, Business, Enterprise and Edu plans, and in GitHub Copilot on Pro+ and higher plans (GitHub changelog).
OpenAI's headline claim is that Sol comes within 1.1 points of Claude Fable 5's best score on DeepSWE v1.1 at roughly 80% lower cost.
3. GPT-6 Luna, the cheap one people are missing
Luna costs $0.10 per million input tokens. Free and Go users get it in the Codex desktop app (OpenAI's announcement), and it's on Copilot Pro. It's good enough for small scripts, tests, boilerplate and explaining code, which covers most of what you do in a college project.
Why the benchmarks don't settle it
Look at the numbers above again. Anthropic reports Terminal-Bench, OpenAI reports DeepSWE and AutomationBench. Neither launch post reports the other's benchmarks, so you can't put them side by side. Each company picked the tests it wins. The only comparison that tells you anything is one you run on your own code.
How to run your own test in 20 minutes
- Pick one real bug or small feature from a project you've already built.
- Make two branches from the same commit, one per model.
- Give both the exact same prompt, once in Claude Code and once in Codex.
- Compare three things: the diff, how many times you had to correct it, and what it cost.
- Keep the branch you'd actually merge. That's your answer.
How I'd use them on a student budget
- Luna for the small everyday jobs: tests, renames, docs, quick scripts.
- Opus 5.5 or Sol only for the hard parts: a migration, a gnarly bug, a feature that touches many files.
- Whichever wins your test as the default. Switching costs you a model menu, not a rewrite.
If you want the harness side of this, read 8 GitHub repos that make Claude Code 10x better. A better model helps less than you'd think if the setup around it is weak.
Follow @avi_vashishta29 for one practical AI dev tool a day.