From Three-Horse Race to Duel
In January 2026 you could still have placed a bet on how the terminal-agent race would end. Claude Code was in front, Codex was a niche tool, Google had just launched Antigravity and had Gemini CLI – the darling of the open-source scene – in its portfolio. Everything looked open.
Nine months later it isn't open anymore. The JetBrains Developer Ecosystem Survey 2026 – published in August, fieldwork May to July, more than 15,000 professional developers worldwide – paints a picture that is hard to read as anything other than: two standards, lots of spectators.
- Claude Code: used by 39% of developers worldwide, up from 18% in January. In the US it's 47%. 31% name it their most-used tool, and around 80% of regular users make it their primary one.
- Codex: 16%, up from 3% in January. Awareness jumped from 27% to 65%.
- GitHub Copilot: 21%, down from 29% a year earlier.
- Cursor: 12%, down from 18% in January.
- Google Antigravity: 6%. Flat – even though awareness rose from 29% to 47%.
- OpenCode: 7%.
The backdrop: 90% use AI coding agents at work at least weekly, 68% daily. The "everyone experiments with everything" phase is over. What remains are two defaults – and a company with the best ingredients in the field that is nonetheless standing still.
At nh labs we use Claude Code and Codex in client work every day. This piece is not a takedown of Google. It's an attempt to explain why a company with the strongest models, its own hardware and the most aggressive pricing is losing the developer-tool race right now.
How Claude Code Became the Default
Claude Code got two things right early that sound trivial in hindsight.
First: terminal first. While the competition was refining autocomplete inside IDEs, Claude Code was an agent living in the terminal from day one – where Git, CI, Docker and shell scripts also live. The tool didn't sit next to the workflow; it sat inside it.
Second: agentic, not assistive. Not "suggest my next line", but "do this task, read the codebase, run the tests, make the commit". In 2025 that was still a risky product promise; in 2026 it's what everyone means when they say "coding agent".
The enterprise consequence was clear. Menlo Ventures put Anthropic's share of enterprise AI coding at 54% in December 2025, against OpenAI's 21%. Across total enterprise LLM spend, Anthropic sat at 40%, OpenAI at 27%, Google at 21% – and coding made up 51% of all enterprise gen-AI usage. Claude Code's annualised revenue run-rate was reported above US$2.5 billion by late 2025.
In February, in "From 4% to 20%: Why Claude Code Is Taking the GitHub World by Storm", we picked up the SemiAnalysis projection: roughly 4% of all public GitHub commits at the time, doubling within a month, 20% by year-end as the forecast. Where are we? Independent estimates from mid-March put it near 10%. Whether the 20% falls by December is open – plausible, but not confirmed. We're writing it that way deliberately, because nobody should argue with a number that hasn't been measured yet.
How Codex Caught Up
The real surprise of the year isn't Claude Code's dominance but the fact that Codex quintupled its share in six months. From 3% to 16% – that isn't organic growth, that's distribution.
OpenAI didn't sell Codex as a developer tool; it put it inside the ChatGPT subscriptions. If you're paying anyway, you have Codex. The desktop app arrived in February 2026 and from there it went vertical: in early June OpenAI reported more than 5 million weekly active users, up more than six-fold since the desktop launch. As recently as April, Sam Altman had cited 3 million and month-over-month token growth above 70%.
The most interesting figure is a different one: roughly 20% of Codex users aren't developers. OpenAI has long stopped positioning Codex as a CLI for programmers and instead frames it as a general work agent that also tidies up data, builds documents or writes a script for the finance team. That's a different bet from Anthropic's – broader, less sharp – and so far it's paying off.
When we wrote "Claude Code vs. Codex CLI: Honest User Feedback from Real-World Practice" in January, Codex was the less mature tool in a head-to-head. That's still partly true. But maturity was never the bottleneck. Reach was, and OpenAI delivered it.
What Happened at Google
Now for the difficult part, because Google actually had the best starting position of anyone.
In November 2025 Google launched Antigravity, its agentic desktop IDE. Alongside it, Gemini CLI was what you'd call an open-source darling: Apache 2.0 licence, more than 100,000 GitHub stars, more than 6,000 merged external pull requests. Developers were actively building it with Google. That's a form of trust you can't buy.
At I/O on 19 May 2026 Google announced consolidation under Antigravity. On 18 June, Gemini CLI stopped serving requests for free accounts, Google AI Pro, Google AI Ultra and individual Code Assist accounts. Enterprise licences for Gemini Code Assist Standard and Enterprise, as well as paid API keys, keep working – that's fair to say. But for the large majority of users who had made the tool big, that was it. Four weeks' notice.
The replacement is called Antigravity CLI: a closed-source Go binary, command agy. Google itself said there would be no 1:1 feature parity at launch. So you take away from developers an open tool they helped build, and hand them a closed one that does less.
What that means in practice: CI pipelines that called gemini broke. Shell scripts, Git hooks, team documentation – all of it had to be touched. Not because the model had got worse, but because the vendor changed the rules. The "Google Graveyard" reflex every developer knows got a new exhibit – and this time not a consumer product, but something in the middle of the toolchain.
And it wasn't the first restructuring. Code Assist, Jules, Gemini CLI, Antigravity, Antigravity 2.0 – four rebrands in 18 months. Over the same period Claude Code was one product with one command. So was Codex. If you have to relearn what the tool is called every few months, you don't build a habit. The JetBrains numbers show exactly that: awareness from 29% to 47%, usage flat at 6%. People know Antigravity. They don't trust it.
The Real Reason: Developer Tools Are Habit Markets
This is the core of the argument, and it has little to do with model quality.
A coding agent isn't a chat window you switch like a browser tab. It sits in the terminal, in CI, in Git hooks, in the conventions of a CLAUDE.md or AGENTS.md, in the MCP servers a team has configured over months. The switching cost is workflow, not model. A team that has spent a year building prompts, scripts and rules around one tool doesn't switch over three benchmark points.
Which is precisely why different things count day to day than on the leaderboard:
- Instruction-following over long sessions. Does the agent stick to what's in the config file after two hours and 40 tool calls? Independent reviewers reported in 2026 that Gemini CLI drifted from instructions in longer sessions and modified code when only an analysis had been requested. That's not a benchmark problem, it's a trust problem – and it destroys habits faster than any ranking builds them.
- Reliability. Will the nightly pipeline still run tomorrow with the same command?
- Pricing predictability. A team of eight at £200 per head per month wants to know in advance whether the agent will send the same bill next quarter.
- The vendor doesn't pull the rug. Since 18 June that's no longer an abstract point.
Yes, reviews also cited gaps on SWE-bench Verified, in the order of ten or more points between Gemini and the leading Claude model. But that's the less important half of the story. A model that scores a few points lower and behaves exactly the same way for two years beats one that scores higher and carries a new name every six months.
The Counter-Case: Google Isn't Out
If you're reading this as "Google has lost", you're drawing the wrong conclusion. The raw ingredients are still first-rate, possibly the best in the field.
The Gemini models are top-tier, cheap, and come with context windows that are a genuine advantage on large monorepos. Google runs its own TPUs and is less dependent on anyone else's hardware than any other player. Antigravity 2.0, announced at I/O in May, is a serious multi-agent desktop with parallel subagents, custom workflows and scheduled background tasks. Jules V2 pursues outcome-based goals rather than individual tasks. And Google owns channels nobody else has: Android Studio, Firebase, Cloud, the enterprise sales motion through Workspace.
Nor are the numbers flat everywhere. In India, Antigravity sits at 15%, level with Cursor, up from 10%. And enterprise Code Assist customers were never cut off – Google didn't alienate the company paying the bill. It alienated the community that would have convinced the company.
A duel in 2026 can be a three-horse race again in 2027. The condition is banal and, for Google, apparently hard: one product, one command, stable for 24 months. If agy is still called agy in autumn 2028 and does what it does today, the survey will look different.
And the counter-case has a second side. A two-vendor market is comfortable for the two vendors. Lock-in and pricing power aren't theory – we went through this at length in "Model Lock-in: The Dependency No One Plans For". Anyone betting on Claude Code or Codex today should know that these two will also raise prices and rebuild features. Google stumbling doesn't make the two winners charities.
What This Means for Teams Choosing Now
Four rules follow from our practice.
- Choose by workflow fit and vendor continuity, not by benchmark. Trial the agent for two weeks in the real repo, with real CI, with real reviews. Whatever sticks in the routine afterwards is the right tool – regardless of what the leaderboard says.
- Keep your conventions portable. AGENTS.md and CLAUDE.md are text files; write them so a rename makes them work for the other tool. MCP servers are an open standard – use it instead of proprietary plugins. Anything that works with plain Git survives any vendor change.
- Run both leaders if the budget allows. We do. Not out of indecision, but because it's the cheapest insurance against point four – and because the tools have different strengths that complement each other in review.
- Treat the agent CLI as infrastructure. With the same vendor-risk discipline as a database: who is the vendor, how often have they rebuilt the product in the last 18 months, what happens to our pipeline if they deprecate it tomorrow? Those questions would have saved every team four frantic weeks in May 2026.
Conclusion
Nine months ago the terminal-agent market looked like an open race. Today two have pulled away, and not because their models are unbeatable. Claude Code was the first to understand that the agent belongs in the terminal and should complete tasks rather than suggest lines. Codex understood that distribution counts for more than feature maturity. Both understood that a product needs a name and has to keep it.
Google has the best ingredients – models, hardware, context, price, sales channels – and still, in June 2026, squandered the thing that grows slowest and dies fastest in the developer market: trust. Not because Gemini is bad. But because developers don't build their pipelines on something that will be called something else next year.
Developer tools aren't won with benchmark points but with habit, continuity and workflow. Whoever understands that can be back in the race in 2027. Whoever doesn't has the best models in the world – and nobody calling them from the terminal.