Google gave every engineer a rival's model and kept its own harness. The config it gave them is not the one you can buy.

What changed

Until this week, Google engineers were directed to build with Gemini. Outside coding tools — Claude Code, OpenAI’s Codex — were off-limits for general use, with carve-outs for some Google DeepMind teams and a short list of high-priority engineering projects. That was a coherent position for a company that ships a frontier model and a harness built on it.

As of 15 September, every Google engineer can select Claude Opus 5 inside Antigravity, subject to an individual usage quota. Gemini stays the default. Claude Code stays blocked.

Read that carefully, because three separate decisions are packed into it:

1. The model restriction was dropped. Engineers can use a competitor’s weights on Google’s own code.
2. The tool restriction was kept. They cannot use the competitor’s agent product to do it.
3. The quota was added. Access is rationed per person, not granted flat.

Google conceded exactly one of the three things its engineers were asking for, and it happens to be the one that costs Google the least amount of control.

The third reversal, and the narrowest one

This is not the first time a hyperscaler has lost this argument with its own engineers.

Amazon ran the same play earlier in 2026 and lost it more comprehensively. Internal guidance steered engineers toward Kiro, its in-house assistant, and imposed stricter approval requirements on third-party AI coding tools for production work. Roughly 1,500 employees backed internal threads pushing for Claude Code, several of them making the uncomfortable point that they were being asked to sell a tool through Bedrock that they were not permitted to use themselves. By May 2026, Amazon had relented and opened access to Claude Code and Codex.

Microsoft sits at the other end: it made its own first-party model, Project Polaris, the default engine for paid GitHub Copilot subscribers in August — roughly 20 million seats — while Copilot itself remains multi-model by design. It never had to reverse a ban because it never ran one.

Google’s version is the most controlled of the three. Amazon gave up the tool. Google gave up only the weights, and routed them through infrastructure it owns, meters and logs. If you are building a policy for your own engineering organisation, that is the template worth studying — not because it is generous, but because it is the minimum concession that actually answers the complaint.

The tier Google named is the tier where the models are switched off

Here is the part a buyer can check, and should.

Google’s statement grounds the internal arrangement in the commercial one: third-party model access is “consistent with our external Antigravity enterprise offering.”

Antigravity’s public model documentation tells a different story. The third-party entries — Claude Sonnet 4.6 (thinking), Claude Opus 4.6 (thinking) and GPT-OSS-120b — each carry the same availability row across the four plan columns: available on Free & Google AI Plus, Google AI Pro and Google AI Ultra; not available on Enterprise. Every Gemini model, including Gemini 3.1 Pro, is available on all four.

Two things follow.

The direction is backwards from what procurement expects. Enterprise tiers are normally supersets — everything the consumer plan has, plus controls and a contract. Here, Enterprise is the only tier that loses models. A team that pilots on individual Pro seats with Claude selected, likes what it sees, and then procures Enterprise discovers the loss at the exact point the rollout stops being reversible. That is not a hypothetical failure mode; it is the default path a sensible evaluation takes.

The published list is three Claude releases stale. Antigravity documents Opus 4.6 as its newest Claude. Anthropic has since shipped Opus 4.7, Opus 4.8 and, in July, Opus 5 — the model Google’s engineers now have.

There are innocent explanations for both. Docs pages lag products constantly. “Enterprise offering” may refer to a Gemini Enterprise Agent Platform arrangement with terms the public table does not describe. An internal deployment at a company that has invested $40 billion in Anthropic may simply be bespoke. Google has not said which, and we are not asserting bad faith.

But the buyer’s position is unchanged across all three: the configuration Google gave its own engineers is not one you can currently select from the published menu. If Antigravity is on your shortlist because it can run Claude, get that confirmed for your tier and your contract, in writing, before the pilot rather than after.

”Gemini remains our primary model” is the credible part

It is tempting to read the Gemini sentence as face-saving. It probably is not.

A default is a different thing from a ranking. Google runs Gemini as the internal default for reasons that have nothing to do with which model wins a head-to-head: cost at Google’s volume, latency against Google’s serving infrastructure, dogfooding the product it sells, and the ordinary institutional preference for the thing the company controls. Quota-limited access to a rival for “specialized use cases” is entirely consistent with Gemini being the right default for most Google code and wrong for some of it.

What the reversal does establish is narrower and more durable than a benchmark claim: single-vendor model policy failed a cost-benefit test at the company best positioned to make it work. Google has the cheapest possible access to its own weights, total control of the toolchain, and a strategic interest in the answer coming out the other way. It still concluded the productivity loss was not worth it.

That is the finding that generalises. Your organisation has none of Google’s advantages and all of the same pressure.

Model choice and harness choice are different purchases

The structural lesson here is one this desk keeps arriving at from different directions.

A harness — Antigravity, Claude Code, Cursor, Copilot, Codex — is not a neutral pipe between you and a model. It owns prompt scaffolding, context assembly, tool definitions, sandbox behaviour and retry logic. The same weights behave measurably differently across two harnesses, which is why “Opus 5 in Antigravity” is a genuinely different product from “Opus 5 in Claude Code”, and why Google’s engineers did not get what they asked for even though they got the model they asked for.

It follows that these are two decisions, and they fail independently:

• The model can be taken away by a supplier. OpenAI’s termination of Cursor’s model access in August was the clean demonstration: multi-model support was never an architecture, it was a contract term, and contract terms have counterparties. The Antigravity tier table is a quieter version of the same fact.
• The harness can be taken away by its owner. Google sunset Gemini CLI and the Code Assist extensions on 18 June with no grace period, making Antigravity 2.0 the only path forward for consumer-tier users. Anthropic cut Claude Code’s weekly limits by 17% on 14 September. GitHub let promotional Copilot credits expire on 1 September, revealing the real cost baseline.

Tools that let you change one without changing the other are worth more than their feature lists suggest. That is the axis the harness roundup and the Claude Code vs Antigravity comparison are organised around, and it is the one Google just ratified with an internal policy change.

What to do with this

1. Ask whether you can change models without changing tools. If your setup allows it, you are insulated from a supply decision you do not control. If it does not, write it down as a concentration risk and revisit at renewal. Do not migrate today over this.
2. Confirm tier-level model availability before you pilot. Specifically for Antigravity, ask your account team which third-party models are available on the tier you intend to buy, and get it in the contract if it matters. The published table currently says Enterprise gets Gemini only.
3. Treat internal-adoption stories as a signal about engineers, not about products. They reliably tell you which model developers push for. They tell you almost nothing about which product is best, because the version that survives the reversal is always shaped by the employer’s control requirements.
4. If you run a single-vendor model policy, price the exception. Google’s answer — allow the model, keep the harness, meter the access — is a reasonable middle. It preserves audit logs, sandbox policy and code-retention posture while removing the complaint. It is easier to defend than a flat ban, and cheaper than approving a second full agent stack.
5. Do not read a coding-model ranking out of this. Google published a procurement decision about specialised use cases, not a benchmark. The coding tools roundup and developer guidance still turn on your own evals against your own repository.

The honest read

Google blocked outside coding tools all year, its engineers kept asking, and this week it gave them the smallest thing that would answer the question: a rival’s weights, inside Google’s own tooling, rationed. Nobody should be surprised, and nobody should read it as capitulation — Gemini’s coding story has had a difficult year, but a default is not a verdict.

The genuinely useful artifact is the mismatch. A company described its internal setup as consistent with what it sells, and what it sells — per its own published table — does not include the model on the tier an enterprise buys. That is probably a stale docs page. It is worth ten minutes of your account manager’s time to find out, because the alternative is discovering it during a rollout.

Questions or corrections? Email Pick Right. Want the full list? See all news.