Best AI Model for Coding? What 1,200 Developers Actually Launch

Most answers to this question start from a benchmark or from someone's taste. We have a different starting point: a record of which model people pick when they have work to do.

AgentsRoom is a desktop app that runs coding agents side by side. When someone starts an agent, the app notes the CLI and the model, and nothing about the work itself. This article reads 28 days of those records, from 9 October to 5 November 2026: 205,366 launches and 1,194 signed-in accounts.

The short answer. If you want one name, it is Opus on Claude Code: it comes first whichever way we count. GPT-6.1 Sol on Codex comes second. But a good part of that lead is a default we set ourselves, and the numbers say more about how people choose than about which model writes better code.

The ranking, counted three ways

Launches are very unevenly spread. The median account started 10 agents in 28 days, and ten accounts made 35% of all launches. One account with a few scheduled jobs can move a model by several points, so here is the same ranking counted three ways: all launches, launches without the ten heaviest accounts, and one account, one vote (the model each account launched most).

Three things to read in that table.

The first two lines of Opus are one model. The opus alias follows the latest release, and Claude Code has resolved it to Opus 5.5 since version 2.1.280. Added together, Opus 5.5 is 37% of launches and the main model of a little over half of accounts.

Sonnet is a second model. It makes 13% of launches but it is the main model of only 6% of accounts: people use it next to something else.

Counting people changes the picture. Gemini 3.8 Flash goes from 1.3% of launches to 4.6% of accounts. At the CLI level, the gap is wider:

Our own default is in this ranking

Here is the part a vendor benchmark would never have to admit. In AgentsRoom, each built-in role arrives with a suggested model: Opus for fullstack, frontend, backend, architect, security and mobile, Sonnet for QA, PM, DevOps and git, Haiku for marketing and translation. Create a fullstack agent and the picker is already on Opus.

You can see it in the data. On Claude Code, the most launched model for DevOps, PM and QA agents is Sonnet, and for fullstack, backend and architect agents it is Opus. That is our table of defaults, almost line for line. So we won't tell you that developers prefer Sonnet for QA. Mostly they kept what we suggested.

Inertia does the rest: for 70% of accounts, the first model launched in the period is also the one they launched most.

The more honest signal is in the choices nobody gets by accident. 198 accounts pinned Opus 5.5 by hand although the alias gave them the same model. 116 launched Fable and 121 pinned Sonnet 5.5, and neither is a default for any role. These groups are smaller, and they are the people who looked at the list and decided.

What people hand over to automation

Without the ten heaviest accounts, 28% of launches come from a backlog ticket, a team run or a trigger, and the rest mostly from someone opening or resuming an agent. The ten heaviest are a different world: 68% of their launches start from automation. Those accounts have largely stopped starting agents by hand.

Share of each model's launches that start from automation, without those ten accounts:

On Claude Code, the tiers split as you would expect: about 20% for Opus, about 40% for Sonnet and Haiku. Our default inflates the Sonnet figure, since QA and PM seats in a team run start on it, and pinned Sonnet 5.5 sits at 18%. On Codex there is no such pattern, and GPT-6 Astra, the frontier model of the list, is at 40%. So we would not turn this into a law. The nearest thing to a rule we can defend: on Claude Code, people keep the strongest tier for the work they watch, and hand more of the routine to cheaper ones.

Effort follows the same logic. Among launches where one was set, high leads on Opus (48%) and on GPT-6.1 Sol (59%), and medium leads on Sonnet (61%). Reasoning effort is a setting of its own, and people who pin Opus 5.5 push it further: xhigh or max on 25% of their launches, against 20% on the alias.

Few people stop at one

30% of accounts launched a single model in 28 days, and 21% launched six or more. Role defaults inflate that, since they can put Opus and Sonnet in the same account without anyone deciding. The CLI count is cleaner, because a second CLI means an install and a login: 45% of accounts launched at least two.

The overlap is lopsided. More than eight Codex accounts in ten also launched Claude. Fewer than three Claude accounts in ten launched Codex. For a good share of these developers, the answer to "which model is best" is a pair, and that is a reasonable outcome.

One small detail, on a small sample: 6 of the 34 accounts on GitHub Copilot launched anything on a weekend, against 47% of Claude accounts and 52% of Codex accounts. That would fit a tool an employer pays for, though we can't prove it.

A new model takes over in about a week, if it is the default

GPT-6.1 Sol first shows up in our records on 29 September, and on 1 October it became the model a new Codex agent gets unless you pick another. Among the main Codex models, its share of launches went like this, week by week: 19%, 50%, 53%, 61%, 68%, 65%. Over the same six weeks, GPT-5.6 Sol went from 82 accounts to 24.

Those 24 are interesting. So are the 17 accounts still launching Opus 5 in the week of 2 November, which was still running when we pulled these figures. Some stay by choice, and some because a plan does not offer the newer model. Our data can't separate them.

On Claude Code, you can barely see a release at all. Everyone on the alias moves without doing anything, which is why pinned Opus 5.5 understates the model.

There is also a trap in reading launches as popularity. Without the ten heaviest accounts, Codex fell from 26% of launches in the week of 21 September to 16% in the week of 2 November. Yet the share of active accounts launching Codex was 36% at both ends, and stayed between 33% and 38% in between. About the same share of people, launching a smaller share of their agents on it. If you only had the first number, you would write the wrong headline.

Free and local models

Many people search for the best free model, or the best one to run locally. This data can't answer that. Freebuff, free and ad-funded, makes 0.3% of launches. Local models through Ollama make 0.2%, across 47 accounts. Our users mostly already pay for a subscription, so a low share here is no verdict. If local is what you are after, start with how we run Ollama next to cloud agents.

What these numbers can't tell you

• A launch is not a quality score. Nothing here measures whether the code worked or got merged.
• A launch is not a volume of work. A five-minute fix and an overnight run count the same, and a resumed session counts as a launch (20% of the total).
• The panel is not all developers. These are people who already drive several agents from a desktop app that began around Claude Code, so Claude is over-represented, and only accounts that keep usage sharing on are counted.
• Our defaults are in the numbers, as described above.

How to choose on your own code

Since no ranking knows your repository, here is the test we would run:

1. Take a real task whose result you can judge, ideally one you have already solved.
2. Give the exact same prompt to two or three models, each in its own git worktree so they can't touch each other's files.
3. Compare the diffs, the time and the cost.
4. Do it three times on different tasks. A single try proves nothing.

Compare agents does steps 2 and 3 from one launcher, and that page also gives the method by hand with plain git commands, if you would rather not install anything. To find the same tier across CLIs, use the model equivalents table. For the choice of tool rather than model, see how to choose by the way you work. The live version of these figures, over 7, 30 or 90 days or 12 months, is on the usage page.

Frequently asked questions

What is the best AI model for coding right now?

No measurement can name one for your code. What we can say is what gets used: over 28 days ending 5 November 2026, Opus on Claude Code was the most launched model, 37% of launches once the alias and the pinned Opus 5.5 are added together, and the main model of a little over half of accounts. GPT-6.1 Sol on Codex came second with 13%. Part of the Opus lead comes from our own default, and a launch says nothing about the quality of the result.

Which AI is best for coding for free?

These numbers cannot answer that. The people in this panel mostly pay for a subscription: Freebuff, which is free and ad-funded, makes 0.3% of launches, and local models through Ollama make 0.2%. A low share here is not a verdict on those tools. Run the same test described above on a task of yours, with the free options you can actually use.

Is Opus better than GPT-6.1 Sol for coding?

We don't know, and this data can't tell. It shows that more accounts launch Opus, not that Opus writes better code. More than eight Codex accounts in ten also launched Claude over the period, which suggests many developers keep both rather than settling the question. The honest test is the same task on both, on your repository, repeated a few times.

Should I pin a model version or use the alias?

An alias such as opus follows the latest release, so you get new models without doing anything, and your results can change without warning. A pinned version such as Opus 5.5 stays put until you move it, which is useful for scheduled jobs and for comparing runs. 198 accounts pinned Opus 5.5 over the period even though the alias already resolved to it.

What reasoning effort do developers use?

Among launches where an effort was set, high leads on Opus (48%) and on GPT-6.1 Sol (59%), and medium leads on Sonnet (61%). Accounts that pin Opus 5.5 go further than the alias: xhigh or max on 25% of launches against 20%.

How were these numbers collected?

When an agent starts, the AgentsRoom desktop app records the CLI, the model, the reasoning effort, the agent role and how the agent was started, and nothing else: no prompt, no project, no file, no name. Only users who keep Share usage data on contribute. The live version, updated every ten minutes, is on the public usage page.