GPT-6.1 Sol Coding: Codex Setup, Effort Levels and the COBOL Result
GPT-6.1 Sol coding results are the reason this model got so much attention. OpenAI tested it on moving old COBOL code into modern languages, and it went from 5.4% on the previous Sol to 86.6% at low effort. In the video below I show what that means and how to use it. This guide covers the coding scores, the exact Codex commands, which effort level to pick and why you still need a human to check the output.
Key takeaways
86.6% on OpenAI’s COBOL migration test at low effort, and 97.9% at higher effort.
Most coding models are judged on Python and JavaScript. OpenAI tried something harder: old languages that very few people still write. On 19 October 2026 it shared results on COBOL, the language still behind a huge amount of banking.
OpenAI’s own figures show why that matters. COBOL supports over 40% of online banking, 80% of in-person card payments and 95% of ATM transactions. Moving that code to something modern is slow and risky, and the people who know COBOL are retiring.
If you want the full model overview first — pricing, availability and non-coding benchmarks — read my GPT-6.1 Sol guide.
The COBOL migration test
Here’s how the test worked. The model had to move a COBOL program to C++, Java, Kotlin, Python or Rust. Hidden tests then checked that the new program gave exactly the same outputs, byte for byte. A task only counted if every test passed.
Model
Setting
Success rate
GPT-5.6 Sol
Low effort
5.4%
GPT-6.1 Sol
Low effort
86.6%
GPT-6.1 Sol
Higher effort
97.9%
And at low effort GPT-6.1 Sol did it at 61% of the older model’s API cost, so it was 39% cheaper as well as far more accurate.
OpenAI also tested building a SQLite-compatible database engine from scratch in 20 languages. GPT-6 Astra averaged 87.4% there vs 26.8% for GPT-5.6 Sol. OpenAI didn’t publish a GPT-6.1 Sol score for that test, so I won’t quote one.
Want AI coding agents shipping work for your business? Inside the AI Profit Boardroom you get the Agent OS — Hermes, OpenClaw, OpenCode, Claude and Codex in one place — plus step-by-step Hermes training and four live coaching calls a week.
On DeepSWE v1.1, OpenAI’s main agentic coding test, here’s what each effort level gets you, per the launch post:
Setting
DeepSWE v1.1
Cost per task
GPT-6.1 Sol low
64.4%
$0.17
GPT-6.1 Sol medium
73.0%
$0.42
GPT-6.1 Sol high
75.2%
$0.65
GPT-6 Sol max
68.8%
$2.74
GPT-6 Astra high
73.2%
$3.92
GPT-6 Astra xhigh
74.1%
$4.43
Read that table again. GPT-6.1 Sol on medium already beats Astra on high, at about a ninth of the cost per task. That’s why I now use Sol as my default in Codex.
OpenAI’s Codex model guide says to use GPT-6.1 Sol for complex coding and agent work when your account has it. Here are the commands:
# start Codex on GPT-6.1 Sol
codex --model gpt-6.1-sol
# one-off job without opening a session
codex exec -m gpt-6.1-sol "Review the current changes"
To make it your default, add one line to your Codex config.toml:
model = "gpt-6.1-sol"
Inside a session, type /model to switch models or change the effort level. It works in the Codex desktop app and the CLI.
Which effort level should you pick?
Codex names the effort levels from Light to Ultra. Light is the same as Low in the CLI.
Effort
Use it for
Light / Low
Small edits, quick fixes, renaming
Medium
Most everyday coding
High
Hard bugs, new features across many files
Max
One big task where it should take its time
Ultra
Big jobs split across subagents running in parallel
OpenAI’s own advice is that most tasks don’t need Max or Ultra. I agree. Start on Medium and move up only if it gets stuck.
Ultrafast mode
Ultrafast generates tokens up to 8× faster than standard speed in Codex. It’s listed for Pro ($500) and eligible Enterprise and Edu plans. In the API, Ultrafast costs 6× the standard price and Fast costs 2×. It’s worth it when you’re sitting and waiting, not for overnight jobs.
This is important. OpenAI itself says production code still needs expert review and testing against your business rules and dependencies. An 86.6% pass rate means some migrations still fail. Passing hidden tests also doesn’t prove the new code handles every edge case your business relies on.
Keep a human who knows the system in the loop.
Run your own test suite, not just the model’s.
Ship in small pieces you can roll back.
Watch out for API tool calling: use the Responses API, since Chat Completions doesn’t support tools for this model.
A simple GPT-6.1 Sol coding workflow
Write the brief. Say what the code should do, which files it can touch and how you’ll test it.
Start on Medium. It already beats Astra on high in DeepSWE.
Ask for a plan first. Read it before any files change.
Run your tests. If something fails, paste the error back and let it fix it.
Review the diff. Use codex exec -m gpt-6.1-sol “Review the current changes” as a second pass, then read it yourself.
This loop keeps costs low and catches most mistakes before they reach your users.
For old-code migrations, go one module at a time. Give it the original COBOL, the target language and the expected outputs, and have it write the tests before the new code. That mirrors how OpenAI scored its own test, where a task only passed if every output matched exactly.
My verdict on GPT-6.1 Sol coding
For GPT-6.1 Sol coding work, this is the best value OpenAI model right now. It beats Astra on DeepSWE at medium effort for a fraction of the cost, and the COBOL result shows real progress on old code. Use it as your Codex default, keep effort on Medium, and always review before you ship GPT-6.1 Sol coding output to production.
About Julian Goldie
I’m Julian Goldie — founder of Goldie Agency and the AI Profit Boardroom. I test new AI agents almost every day for 432K+ YouTube subscribers and run Hermes, OpenClaw, OpenCode, Claude and Codex side by side in my own Agent OS. Claim your free strategy session or join the AI Profit Boardroom.
FAQ: gpt-6.1 sol coding
Is GPT-6.1 Sol good for coding?
Yes. It scores 75.2% on DeepSWE v1.1 at high effort, ahead of GPT-6 Astra at xhigh, for $0.65 per task.
How do I use GPT-6.1 Sol in Codex?
Run codex –model gpt-6.1-sol, or set model = “gpt-6.1-sol” in config.toml. Use /model to switch models or effort.
What was the GPT-6.1 Sol COBOL result?
On OpenAI’s COBOL migration test it passed 86.6% of tasks at low effort, vs 5.4% for GPT-5.6 Sol, and 97.9% at higher effort.
Which effort level should I use for coding?
Medium for most work. Move to High for hard bugs. OpenAI says most tasks don’t need Max or Ultra.
Can GPT-6.1 Sol code go straight to production?
No. OpenAI says it still needs expert review and testing against your business rules.
Ready to put Hermes to work in your business? Join the AI Profit Boardroom for the Agent OS, setup training and daily updates.
