Google spends billions building Gemini, markets it as one of the industry's handful of true frontier model families, and wires it into everything from Search to Workspace. So it's a little jarring that, starting this month, the company is handing its own engineers a personal quota to use a competitor's model instead. According to a September 15, 2026 report by Business Insider's Hugh Langley, Google has opened access to Anthropic's Claude Opus 5 to every engineer at the company through its internal development platform, Antigravity — a reversal of a policy that, until now, kept most employees off outside coding tools entirely.
Before this change, Claude access inside Google was tightly rationed: only select Google DeepMind teams and engineers on a short list of priority projects could reach for it. Everyone else was generally barred from external coding assistants such as Anthropic's own Claude Code or OpenAI's Codex. Google still isn't handing out Claude Code itself — engineers reach Claude Opus 5 by selecting it as a model option inside Antigravity, Google's own tool, under an individual usage cap.
Quick summary: Google has opened Claude Opus 5 to all of its engineers through its internal Antigravity platform, each with a personal usage quota — a reversal from a policy that limited Claude access to select DeepMind teams and priority projects. A Google spokesperson says Gemini "remains our primary and foundational model for internal development," with third-party models like Opus 5 available on quota for "specialized use cases." The move follows reported internal frustration with Gemini's coding performance, and echoes a similar reversal at Amazon in May 2026.
What Business Insider actually reported
Here's what's confirmed across the Business Insider report and the trade coverage that followed it on September 15, 2026:
- Scope: every engineer at Google can now request access to Claude Opus 5, not just DeepMind staff or people on priority projects.
- How it's delivered: access runs only through Antigravity, Google's own internal development platform — not through Anthropic's standalone Claude Code tool.
- Quota, not unlimited use: each employee gets an individual usage allowance for the third-party model, rather than open-ended access.
- Gemini stays the default: Google says Gemini remains its "primary and foundational model" for internal engineering work; Claude is positioned as a supplementary option for specific use cases.
- Reported driver: multiple outlets describe internal frustration among Google engineers with Gemini's coding performance on hard tasks, even after Gemini gained a refreshed Flash model.
What is Google Antigravity?
Antigravity is Google's own answer to the new wave of "agentic" coding tools — a platform where, instead of typing every line yourself, you describe a goal and an AI agent plans the work, writes code, runs the terminal, and even clicks through a built-in Chrome browser to test what it built. Google took it to public preview on November 18, 2025, and notably launched it multi-model from day one, powered by Gemini 3, Claude Sonnet, and an open GPT-OSS model rather than Gemini alone. Sources: [Build with Google Antigravity](https://developers.googleblog.com/build-with-google-antigravity-our-new-agentic-development-platform/), [Google Antigravity — Wikipedia](https://en.wikipedia.org/wiki/Google_Antigravity).
Google expanded the platform considerably at I/O in May 2026, unveiling "Antigravity 2.0" as a standalone desktop app with its own CLI, SDK, managed agent execution, and enterprise support through the Gemini Enterprise Agent Platform. Sources: [Google Launches Antigravity 2.0 at I/O 2026 — MarkTechPost](https://www.marktechpost.com/2026/05/19/google-launches-antigravity-2-0-at-i-o-2026-a-standalone-agent-first-platform-with-cli-sdk-managed-execution-and-enterprise-support/). That multi-model design is exactly why Google could flip a policy switch and widen Opus 5 access without shipping new infrastructure — the wiring to reach Claude was already inside the tool it built for its own developers.
The quote — and what it really signals
"Engineers have access to select third-party models in Antigravity, which is aligned with our external Antigravity enterprise offering. Gemini remains our primary and foundational model for internal development, with third-party models available on a quota to support specialized use cases." — Google spokesperson, quoted by Business Insider
Read literally, it's a careful non-admission: Gemini is still described as the foundation, and Claude is filed under "specialized use cases," the same bureaucratic language companies reach for whenever they don't want a policy change to read as a verdict on their own product. But the subtext is hard to miss. Google restricts external AI tools more tightly than most of the industry — it kept Claude Code and Codex largely off-limits even as rivals loosened up — and a carve-out broad enough to cover every engineer, company-wide, is a bigger deal than the quote's calm tone suggests.
Why now: a coding gap and a $40 billion insurance policy
Two things make the timing make sense. First, Google's own Gemini line has kept shipping through 2026 — Gemini 3.1 Pro arrived February 19, and Gemini 3.8 Flash followed on September 2, just 13 days before this story broke — yet several reports describe engineers still reaching for Claude and OpenAI's Codex on tougher coding tasks despite those updates. Sources: [Gemini 3.8 Flash launch — 9to5Google](https://9to5google.com/2026/09/02/gemini-3-8-flash-launch/), [Dataconomy](https://dataconomy.com/2026/09/15/google-opens-anthropic-claude-access-for-engineers/).
Second, Google has real money riding on Anthropic doing well regardless of which chatbot wins the internal popularity contest. In April 2026, Google agreed to invest up to $40 billion in Anthropic, a deal reported at the time to combine an initial cash investment with additional tranches tied to future milestones. Sources: [Google Plans to Invest Up to $40 Billion in Anthropic — Bloomberg](https://www.bloomberg.com/news/articles/2026-04-24/google-plans-to-invest-up-to-40-billion-in-anthropic), [Google to invest up to $40 billion in Anthropic — CNBC](https://www.cnbc.com/2026/04/24/google-to-invest-up-to-40-billion-in-anthropic-as-search-giant-spreads-its-ai-bets.html). Anthropic also runs a large share of its training on Google's TPUs and Cloud. Letting engineers use Claude inside a Google-controlled sandbox isn't really Google conceding a rivalry — it's a company hedging across a portfolio it partly owns.
Google isn't alone — Amazon did this first
Google is following a script Amazon already ran. After an internal November 2025 memo telling staff the company would not "support additional third party, AI development tools" beyond its in-house Kiro assistant, Amazon reversed course in May 2026 and rolled out Claude Code and OpenAI's Codex company-wide on Amazon Bedrock, with no special approval required. An Amazon spokesperson downplayed the shift at the time, saying engineering teams were still "primarily using" Kiro — putting the figure at 83 percent of the company's engineers — even as Claude Code and Codex became available company-wide. Sources: [Amazon rolls out Claude Code and Codex to its employees — The Daily Star](https://www.thedailystar.net/news/tech-startup/news/amazon-rolls-out-claude-code-and-codex-its-employees-report-4169101), [Amazon relents, lets its programmers use Codex and Claude — Slashdot](https://developers.slashdot.org/story/26/05/10/0618225/amazon-relents-lets-its-programmers-use-openais-codex-and-anthropics-claude). Two of the handful of companies racing to build the best foundation model have now quietly told their own engineers that, for writing code, outside tools are worth the quota.
Opus 5 vs. Google's own models, side by side
Here's how the model Google just opened up compares with its two most recent in-house releases on the specs Anthropic and Google have each published themselves:
| Model | Maker | Released | Context window | Price per 1M tokens (in / out) |
|---|---|---|---|---|
| Claude Opus 5 | Anthropic | July 24, 2026 | 200K tokens | $5 / $25 |
| Gemini 3.1 Pro | February 19, 2026 | 1M tokens | $2 / $12 (≤200K prompts) | |
| Gemini 3.8 Flash | September 2, 2026 | ~1M tokens | $0.75 / $3.75 (through Dec 2026) |
Opus 5 costs roughly two to six times more per token than Google's own models on the published list prices, which suggests the personal quota system is less about cost-cutting and more about rationing an expensive option for when it matters. Worth noting: Anthropic doesn't publish an Opus 5 SWE-bench Verified score in its own launch materials — it leans on newer benchmarks like Frontier-Bench and CursorBench instead — so a clean, apples-to-apples coding score against Gemini isn't something we can verify yet. Source: [Introducing Claude Opus 5 — Anthropic](https://www.anthropic.com/news/claude-opus-5).
What this signals for chatbot and coding-tool watchers
For readers who track these models to decide what to use themselves, the interesting signal here isn't the press-release language — it's the fact that Google, which has every incentive to keep engineers on Gemini, decided the productivity cost of not doing so was too high. That's a data point benchmarks can't fully capture: engineers who write code for a living, at the company best positioned to build its own tool, choosing to spend part of a rival's quota on it anyway.
Curious whether your own team would notice a similar gap? The cleanest test is a blind one: run the same coding prompt through Claude Opus 5, Gemini 3.1 Pro, and whatever in-house or default tool you're using, then compare the actual output — not the benchmark slides.
01Can Google employees use Anthropic's Claude Code directly now?
Not according to the reporting. Engineers reach Claude only by selecting Opus 5 as a model option inside Google's own Antigravity platform, under an individual usage quota, rather than getting Anthropic's standalone Claude Code tool.
02Does this mean Google thinks Gemini is worse at coding than Claude?
Google's official line is no: a spokesperson told Business Insider that "Gemini remains our primary and foundational model for internal development," with Claude available on quota for "specialized use cases." But several reports point to internal engineer frustration with Gemini's coding performance as the practical reason for the policy change, even after Gemini 3.8 Flash shipped on September 2, 2026.
03What is Google Antigravity?
It's Google's free agentic development platform — you describe a goal and an AI agent plans, writes, tests, and iterates on code across an editor, terminal, and built-in browser. It launched in public preview on November 18, 2025, running on Gemini 3, Claude Sonnet, and GPT-OSS, and was expanded into a standalone app with its own CLI and SDK, called Antigravity 2.0, at Google I/O in May 2026.
04Has any other big AI company done something similar?
Yes. Amazon reversed a similar internal restriction in May 2026, rolling out Anthropic's Claude Code and OpenAI's Codex to its entire workforce on Amazon Bedrock after employees kept asking for Claude over its in-house tool, Kiro.
05How does Claude Opus 5's price compare to Gemini's models?
Opus 5 is priced at $5 per million input tokens and $25 per million output tokens. Gemini 3.1 Pro runs $2/$12 for prompts under 200K tokens, and Gemini 3.8 Flash is $0.75/$3.75 through the end of 2026 — meaning Opus 5 costs several times more per token than either in-house model, which is likely one reason Google capped access with a personal quota rather than opening it up freely.
If a story like this makes you want to see Claude, Gemini, and the rest actually run side by side rather than take a spokesperson's word for it, that's the idea behind LumiChats — one place to compare how different AI chatbots and models answer the same question, coding or otherwise, before you decide which one earns a spot in your own workflow.
