For most of 2026, developers argued about whether Claude Code or OpenAI’s Codex was the better coding agent. A chart posted by YouTuber and T3 Code founder Theo Browne on November 10 suggests that, at least among his users, the argument has tilted hard: Claude Code now receives 75% of user-started turns in T3 Code, up from 52% in the week before Anthropic released Claude Opus 5.5. “The complete and utter domination of Opus 5.5 is hard to put into words,” Browne wrote.
Quick Take
- The data: Claude Code’s share of user-started turns in T3 Code: 52% in the 7 days before Opus 5.5, 64% in the 7 days after, 75% in the last 7 days through November 9. Codex has the rest.
- The source: T3 Code is an open-source “control plane for coding agents” that runs Claude Code, Codex and others with users’ own subscriptions — so the chart measures which harness users choose, not which company sells more.
- Why Claude wins right now: Developers praise Opus 5.5 on complex, multi-file work and say Anthropic’s $200 plan stretches far.
- Why Codex isn’t dead: OpenAI’s GPT-6.1 Sol costs $2/$10 per million tokens and OpenAI says it “nearly matches” GPT-6 Astra; some developers find Codex faster and terser.
- The pragmatic answer: Many developers report paying for both and routing tasks by type and by remaining quota.
What happened
Browne shared a chart titled “Claude Code vs Codex,” covering “which harness T3 Code users send their turns to” from March 7 to November 9, 2026. The chart marks the day each model became usable in T3 Code — Opus 5.5 on September 22, GPT-6.1 Sol on September 29 — and shows Claude Code’s daily share climbing after Opus 5.5 and ending at 78% on the last day plotted. Before Opus 5.5, the lead had swung back and forth: Codex led most days from April through June, and launches like GPT-6 Astra on September 4 briefly pulled the line back below 50%.
One day earlier, Browne had posted that the “$200 Claude sub is the best money you can spend as an engineer right now,” adding that he had run “between 8 and 25 threads” most of the day on Opus 5.5 while using “less than 20% of a $200 Claude sub.”
Background: what T3 Code measures
T3 Code describes itself as “the open-source control plane for coding agents” that can “orchestrate Claude Code, Codex, Antigravity, OpenCode, Cursor, and Grok from one surface.” It “doesn’t resell tokens”: users plug in subscriptions they already have. The code is on GitHub. Because it runs the official agent harnesses, its usage split is a rare, if narrow, look at what developers actually choose when switching is one click away. The chart compares only Claude Code and Codex, not the other harnesses T3 Code supports.
How the two sides compare
On the OpenAI side, the main Codex model is GPT-6.1 Sol. OpenAI’s launch post calls it “near-Astra intelligence for a fifth of the price,” at $2 per million input tokens, $0.10 cached and $10 output. OpenAI says it matches GPT-6 Astra on DeepSWE v1.1 “at roughly one-fifth of the cost” and scores “2.2 percentage points above Opus 5.5 at medium reasoning effort, at roughly a third of the cost” on AutomationBench. A faster “Ultrafast” variant with “up to 8x faster token generation” in Codex was announced as coming.
On the Anthropic side, developers mostly talk about Opus 5.5 rather than the newer, pricier Fable 5.2. Browne wrote that the $200 Claude plan “gets you $15,000 of Opus 5.5 in a month,” but far less of Fable, and that “even ignoring that Fable is restricted to 50% of your weekly limit, it burns faster per $ than Opus does.” Anthropic’s own model guide also tells developers to start on Opus 5.5.
| Claude Code (Anthropic) | Codex (OpenAI) | |
|---|---|---|
| Main model in the debate | Claude Opus 5.5 (Fable 5.2 for hardest work) | GPT-6.1 Sol (GPT-6 Astra for hardest work) |
| Share of T3 Code turns, 7 days before Opus 5.5 | 52% | 48% |
| Share of T3 Code turns, 7 days after Opus 5.5 | 64% | 36% |
| Share of T3 Code turns, last 7 days (to Nov 9) | 75% | 25% |
| Price point developers cite | $200/month Claude plan (per Theo Browne) | GPT-6.1 Sol API: $2 input / $10 output per 1M tokens |
| Praised for | Complex multi-file work, planning, plan limits | Price, speed, terse output |
| Common complaint | Session/weekly limits, Fable burns limits fast | Usage limits, resets, behaviour vs. Opus |
Share figures are from Browne’s chart (Claude Code share; Codex has the rest). Prices are as cited by OpenAI and Browne; subscription limits vary by plan.
Independent tests point both ways
Small developer tests on X show how much results depend on the task. Stanislav Sorokin of Tectomight reported that on eight Swift tasks “Claude Code built 4/4, Codex 0/4.” In another test, he found that on the same coding tasks “Codex read 168K tokens, Claude Code 53K” — the opposite of the common claim that Codex is the token-efficient option. These are tiny samples, not benchmarks. Earlier i10x reporting showed that the harness around a model can move scores as much as the model itself: Coding agent harness vs. model.
Pros and cons
For users (individual developers)
- Pro: Tools like T3 Code make switching between agents cheap, so you can follow whichever model is best this month.
- Con: Juggling two subscriptions and separate usage limits adds cost and friction; one developer complained that “Claude Code lasts me a week, while Codex runs out in just 2 days.”
For developers building tools
- Pro: Real usage data like Browne’s is more telling than vendor benchmarks for deciding which harness to support first.
- Con: Share can swing with every launch; the chart shows Codex briefly regaining ground after GPT-6 Astra arrived.
For companies
- Pro: Competition is pushing prices down — GPT-6.1 Sol’s pricing is a direct answer to Opus 5.5.
- Con: Standardizing on one agent is risky when developer preference can flip within weeks.
What X is saying
The complete and utter domination of Opus 5.5 is hard to put into words. I charted out the split of Codex vs Claude in T3 Code over time. Never seen anything like this before.
— Theo – t3.gg (@theo) November 10, 2026 · View on X
The $200 Claude sub is the best money you can spend as an engineer right now. Truly unbelievable value. I don’t care how “cheap” other models are. I don’t care that you can host your own inference with shitty models on a $10,000 GPU. There is no place you can put $200 and get more value out the other side.
— Theo – t3.gg (@theo) November 9, 2026 · View on X
I feel like in mid 2026 people were still using a lot of different AI coding agents but now everything has converged around Claude. Or am I in a Claude bubble? FWIW I have only ever used Claude, never tried anything else.
— Jon Yongfook (@yongfook) November 10, 2026 · View on X
Not everyone accepts the chart at face value. @jatingargiitk cautioned: “75% of turns could come from a much smaller share of users. i’d want the user split beside this before treating it as a broad preference shift.” Others blamed OpenAI’s policies: @jvivas_official wrote that “openai lost the plot with the resets and the price increases,” while adding, “I like how they behave much better than Claude models, but Opus5.5 is just too good.” Developer @SSShken offered the pragmatic view: “for multi-file changes in a real repo I’d pick claude code, for cheaper small tasks codex.”
What Grok says
In a reply on X, Grok said: “In coding agent traces Claude Code usually reads more files first (often ~8 vs Codex’s ~4) before writing, so it leans harder into exploration. Codex tends to stay tighter to the exact task and finish faster with fewer tokens.” That matches the common perception, though Sorokin’s token count above shows it does not hold in every test.
Critical analysis
Browne’s chart is the most concrete public data point in this debate, but it has limits. T3 Code users are a self-selected group of power users, many of them Browne’s own audience, and the chart counts turns, not users or completed tasks. A handful of heavy users running dozens of parallel threads — as Browne himself does — can move the share. He has also been very publicly enthusiastic about Anthropic’s plans, which shapes what his audience tries.
Still, the direction is hard to dismiss. The jump from 52% to 64% within a week of Opus 5.5 is a sharp, launch-linked change, and the share has kept rising since. The bigger lesson is about pricing structure: developers are comparing subscription value (“work done per week,” in Browne’s words) more than per-token API prices. OpenAI’s cheaper Sol model wins on list price, but if users hit Codex limits faster, list price does not decide the choice.
What’s still open
- Will GPT-6.1 Sol Ultrafast or a new Codex model reverse the trend, as Astra briefly did in September?
- Will Anthropic keep its generous plan limits as usage grows, or tighten them?
- Does the T3 Code split match broader data from other tools and IDEs?
- How will Google’s Gemini 4 Carbon, which tops CursorBench, shift the balance?
About the analysis
Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated illustration. Read our AI Use Policy. Spotted an error? Request a correction.
This independent, research-based i10x analysis draws on Theo Browne’s published T3 Code chart, OpenAI’s GPT-6.1 Sol announcement, the T3 Code website and repository, and public posts on X through November 11, 2026. i10x did not run its own comparison of Claude Code and Codex.
i10x Perspective
The coding-agent market now moves on weekly cycles: a model launch shifts real usage within days. For teams, the safest strategy is to keep agents swappable — through tools like T3 Code or internal routing — and measure work completed per dollar and per week instead of chasing leaderboards. Today that favors Opus 5.5 for the hard work; the chart’s own history shows it may not stay that way.
