Pichai Admits Google Trails on Coding Agents

Coding tools are a bit like coffee shops: the one with the comfiest chairs wins, even if the espresso down the road is technically better. For two years Anthropic has owned the comfy chair, and now the CEO of the largest AI lab on earth has publicly admitted it. Sundar Pichai sat down with the New York Times and said the quiet part out loud.

What Happened

In an interview published on May 28, as The Times of India reported, Pichai conceded that Google is trailing rivals on AI coding agents. Not on text. Not on multimodality. Not on reasoning. On coding. The bit, as he put it, that "ends up being very foundational in everything we do."

The wording was careful but unmistakable. "When it comes to agentic coding with tool use, instruction following and long-horizon tasks, I think we are a bit behind at this moment," Pichai said. He named names too, or near enough. Claude Code. Cursor. Tools where Anthropic, directly or through partnership, captured the developer's screen before Google did.

The diagnosis Pichai offered was not about model architecture. It was about data flows. "Coding was an area where getting access to the data flows was important. We maybe didn't quite have the surface, like Claude Code as an example, or what Anthropic maybe had with Cursor, too." Translation: the other lot are reading over the developer's shoulder, and we aren't.

The counter-punch has two names. Gemini 3.5 Flash, which has shipped, and Antigravity 2.0, an internal tool Googlers have been using for a while. At Google I/O, Pichai shared the token usage on Antigravity 2.0 and said it is doubling every week internally. "I have never seen anything like it internally," he told the Times. He framed the gap as temporary. He used the word "confident" twice.

Technical Anatomy

The interesting bit is what Pichai is admitting about the shape of the moat. For years the assumption inside the big labs was that frontier capability flows from pretraining: more tokens, more compute, more parameters, better model. The boring bit nobody wanted to discuss was that agentic coding does not actually fall out of pretraining alone. It falls out of post-training on the exact traces a developer produces while wrestling with a real codebase.

And those traces only exist where the developer is already typing. Cursor sees every keystroke, every accept, every reject, every Cmd+K rewrite. Claude Code sees every tool call, every failed test, every retry loop. That is reinforcement signal you cannot synthesise. You can scrape GitHub until your eyes bleed and you still won't get the moment a senior engineer rejects a suggestion because it touched a file they happen to know is load-bearing.

This is the gap Pichai is describing when he talks about "long-running tasks where serious developers are working on complicated code bases." Long-horizon agent work is fundamentally a sequence of tool calls. The Anthropic tool use docs describe a pattern Claude Code has been refining in production for months: structured calls, observation parsing, retry behaviour, context window management. Anyone who has shipped an agent against a real repo knows the first ninety percent of the work is making the loop not collapse under its own weight.

Gemini 3.5 Flash, per Pichai, addresses the areas Google was behind in. The Gemini API docs already expose the surface for the agent patterns, but a model and a product are different beasts. Antigravity 2.0 is the product attempt: an internal IDE-shaped surface where Google engineers generate the very data flows Pichai admits the company lacks externally. Doubling weekly token usage is the leading indicator. The lagging indicator is whether that internal signal generalises when the model meets a stranger's monorepo.

Who Gets Burned

Three groups should be reading this interview carefully.

The first is Anthropic. Pichai just told the market that Claude Code is the reference product for agentic coding. That is excellent for the next funding round and terrible for complacency. Google has more capital, more distribution, and Workspace as a Trojan horse. Anthropic's lead is real but it is a lead in a race Google has now publicly entered. The window where Claude Code can charge a premium without serious competition is shorter than it looked six months ago.

The second is Cursor and every other independent coding-agent startup. Pichai name-checked Cursor as a surface Google did not have. That is flattering and ominous in equal measure. The structural question for every IDE-layer startup is whether they are a feature or a company. If Google ships an Antigravity-flavoured product into the VS Code ecosystem at zero marginal cost to Workspace customers, "we have better UX" stops being a defensible answer.

The third group is enterprise platform leads who bet their internal developer platform on a single coding-agent vendor. The bet looked safe in late 2025. It looks twitchier now. Pichai signalled that the leaderboard is going to reshuffle on a cycle that, in his own words, "may not exactly match up" across labs. If your golden path forces every engineer through one provider's API, you are running a portfolio of one in a market where the leader changes every two quarters.

The next 90 days for these groups look like accelerated benchmark theatre, aggressive pricing, and a quiet scramble inside platform teams to abstract the model layer behind an internal interface that does not care which lab is winning this week.

Playbook for AI Development

If you are building product on top of coding agents, three things to do this week.

One: treat the coding-agent provider as swappable infrastructure, not a partner. Wrap whichever API you depend on behind a thin internal interface that exposes the agent loop primitives (tool call, observation, retry, context trim) rather than vendor-specific SDK calls. Pichai just told you the frontier moves on uneven cadences. Your architecture should assume that.

Two: instrument your own data flows now. The lesson buried in Pichai's confession is that whoever sees the developer's screen wins. If you are an iGaming platform, a fintech, an ad-tech shop with a real engineering org, you have a developer surface inside your own walls. The acceptance and rejection signals from your engineers using AI tools are valuable training data for any future fine-tune. Capture them with consent and structure, even if you have no immediate plan to use them.

Three: stop benchmarking on toy tasks. Pichai was explicit that Google is fine on short coding work and behind on long-horizon work over complicated codebases. That is exactly the split most internal benchmarks fail to capture. Build an evaluation harness over your actual repos, with multi-step tasks that span files, before you commit to a vendor for the next budget cycle. Anyone who has watched a coding agent confidently rewrite the wrong service at 4pm on a Friday knows why this matters.

Key Takeaways

• Pichai publicly conceded Google is behind on agentic coding while claiming parity or lead on general intelligence, text, multimodality, voice, audio, and reasoning.
• The root cause he named is data flows, not model capability. Anthropic's developer surface advantage through Claude Code and association with Cursor is the moat being described.
• Gemini 3.5 Flash and the internal Antigravity 2.0 tool are Google's counter. Antigravity 2.0 internal token usage is doubling weekly per Pichai's Google I/O remarks.
• Pichai framed the gap as temporary and said leading labs' pretraining cycles "may not exactly match up," signalling expected leaderboard churn.
• Back to the coffee shop: the chairs matter as much as the espresso. Whoever wins agentic coding will be whoever the developer is already sitting inside when the next model drops.