Claude-modellen in Gemini Agent: wat het lek onthult en wat Google al had opgeschreven

A leak posted on 2026-10-08 claims that G​emini Argon 4, G​emini Flash 3.8, Claude Opus 5 and Claude Sonnet 5.5 will all be available inside the G​emini agent that G​oogle announced the same day for G​emini Business. Three of those four names are wrong, and the correction is more interesting than the leak. The model the post calls "G​emini Argon 4" is G​emini 4 Argon — G​oogle's own frontier model, announced on 2026-09-30 by Koray Kavukcuoglu and still without a callable model ID. The one it calls "G​emini Flash 3.8" is Gemini 3.8 Flash, which has been in the G​emini API since 2026-09-02. And the two A​nthropic models are not the current generation: the leak names Claude Opus 5 (shipped 2026-07-24) and Claude Sonnet 5.5 (shipped 2026-09-28) where G​oogle's own agent documentation already routes to Claude Opus 5.5 and Claude Sonnet 5.5. What survives the check is not the model list. It is the sentence G​oogle published itself.

Because here is the thing a leak can obscure: G​oogle's announcement for the G​emini agent already says A​nthropic is in it. It says so without a version number, without a model ID, and without a single word about G​emini Business. Everything below separates what G​oogle put on the record from what an X account inferred from it, because the two are not the same document and they do not say the same thing.

Wat het lek eigenlijk zegt, en waar het vandaan komt

The signal is one post on X from the account @testingcatalog, dated 2026-10-08, whose full text reads: "G​emini Argon 4, G​emini Flash 3.8, Claude Opus 5 and Claude Sonnet 5.5 will be available in the recently announced G​emini Agent for G​emini Business. It is the first time when Claude models would appear on G​emini, along with G​oogle's models." It carries no screenshot, no source inside G​oogle, no document, and no date for the claimed availability.

That is a single-source claim about a product surface that, as of today, has not shipped. It is worth tracking precisely because the last Claude-versus-G​oogle rollout looked like this too and turned out to be real — but the truthful description of the state of play is that a third-party observer has read G​oogle's own announcement and concluded a product line from it. The verification of a leak is not whether the leaker is credible. It is whether the underlying document says the thing.

What G​oogle already confirmed on 2026-10-08, in G​oogle's own words

G​oogle Cloud published "Welcome to G​emini at Work 2026: Introducing the G​emini agent" on 2026-10-08, adapted from Thomas Kurian's keynote at the G​emini at Work 2026 event, and it settles the A​nthropic question outright. The post's section on model orchestration reads, verbatim:

"It runs each job on the model that fits best — orchestrating across our G​emini family of models and Claude models from A​nthropic today, and other leading private and open models in the future — to deliver optimal quality and lower your costs."

So the claim that Claude is coming to G​oogle's agent is not a leak. It is a published first-party statement, and the leak's framing of it as news is a day late. What G​oogle does not do is name a version. There is no Claude Opus number, no Claude Sonnet number, no model ID, no rate card, no quota and no availability date attached to that sentence anywhere in the post. G​oogle names A​nthropic and names its own lines — Argon for frontier reasoning, Flash for speed and volume, Omni for generative media, G​emma for lightweight open-weights edge workloads — and stops there.

The reasoning G​oogle gives for the architecture is the part worth keeping, because it is a routing argument, and it is the same argument this blog makes about not hardcoding a single model:

• "Het beste model voor de taak is niet altijd het grootste."

• "Het toonaangevende model verandert om de paar maanden, dus die keuze openhouden betekent dat je context, je vaardigheden en je data op hun plek blijven wanneer dat gebeurt."

G​oogle's stated mechanism for that is a set of features it describes as multi-model orchestration, Smart Routing, and real-time spend caps enforceable in the Cloud Billing console — agents that pause when a cap is hit and resume with one click, tracked per project for chargeback. If you have ever priced a month of agent traffic after the fact, that is the part of the announcement that changes what you can decide.

The G​emini Business line is the weakest part of the leak

Of everything in the post, "for G​emini Business" is the piece with the least behind it, and it deserves to be stated plainly rather than passed along.

The Kurian post never uses the name G​emini Business. The tiers it names are G​emini Enterprise — deployed by "nearly 90% of the Fortune 100" in G​oogle's own framing — plus G​emini for Education, and the industry builds G​emini for Financial Services and G​emini for Legal, with Government, Healthcare and Retail marked coming soon. G​emini in G​oogle Workspace is described separately, inline in Gmail, Drive, Docs, Slides, Sheets, Chat and Calendar. The word "Business" does not appear as a product tier anywhere in the announcement.

There is a real surface at business.gemini.google, and it is the kind of thing that produces exactly this confusion: the URL redirects to a G​oogle Workforce Identity Federation sign-in endpoint, i.e. an SSO gate for a G​oogle Workspace-hosted G​emini deployment rather than a named SKU with a page and a price. That is a hosting arrangement, not an announcement. If the leak is right that G​emini Business is where Claude lands first, the evidence is not in G​oogle's post — and if it is wrong, the error is the ordinary one of reading a login domain as a product name.

Where A​nthropic's models already sit inside G​oogle products

The leak's "first time Claude models would appear on G​emini" also needs a correction, because Claude is already selectable inside a G​oogle product, and has been for a week.

G​oogle's Antigravity documentation lists third-party reasoning models next to G​oogle's own, and as of this week the rows read:

• Claude Sonnet 5.5 (thinking) — G​oogle AI Pro (non-trial subscriptions only) and G​oogle AI Ultra

• Claude Opus 5.5 (thinking) — G​oogle AI Pro (non-trial subscriptions only) and G​oogle AI Ultra

• Claude Sonnet 4.6 (thinking) and Claude Opus 4.6 (thinking) — Free, G​oogle AI Plus and G​oogle AI Pro, with removal listed for 2026-11-02

• GPT-OSS-120b — Free, G​oogle AI Plus, G​oogle AI Pro and Ultra, with removal listed for 2026-11-02

• Gemini 3.8 Flash, G​emini 3.7 Flash, G​emini 3.6 Flash and G​emini 3.1 Pro — all tiers, unchanged

Note what is absent from that list: Claude Opus 5 and Claude Sonnet 5, the two models the leak names. G​oogle's existing Claude surface runs the 5.5 generation, on paid tiers, with the older 4.6 rows on a documented retirement date. A leak that names the previous Opus generation for a new G​oogle surface is either describing a cost-optimisation tier or is simply stale — and there is no document either way. Treat the model list in the X post as the least reliable line in it.

Kun je vandaag een van deze vier modellen aanroepen?

This is the question the leak cannot answer and G​oogle's documentation can, at least for the G​emini side.

• G​emini 4 Argon — no. G​oogle announced it on 2026-09-30 with introductory pricing of $2 per million input tokens and $10 per million output, a staged rollout with only the first stage dated (cyber defenders through the Fairwind Program), and no model ID. As of 2026-10-09 the G​emini API's model list contains no Argon entry and no gemini-4 identifier; the newest entries are the 3.8 family. Artificial Analysis has measured Argon at an Intelligence Index of 52.56 on its v4.3.2 revision, on a model that a reader still cannot put into a request.

• Gemini 3.8 Flash — yes, and it has been callable since 2026-09-02, so it is already available in the G​emini API, AI Studio, and on our routes as google/gemini-3.8-flash. G​oogle's own pricing page lists it at $0.75 per million input and $3.75 per million output through 2026-12-31, rising to $1.50 / $7.50 from 2027-01-01. On the same index revision it scores 40.9 — roughly twelve points below Argon, and the gap is the whole reason G​oogle wants both lines in one agent.

• Claude Opus 5 — yes. A​nthropic shipped it on 2026-07-24 at $5 / $25 per million tokens with a 1M-token context window, and Artificial Analysis scores it at 51 on Intelligence Index v4.3, seventh of roughly 200 models measured.

• Claude Sonnet 5.5 — yes. A​nthropic shipped it on 2026-09-28 at $2 / $10 per million tokens, and it scores 56 on the same revision, two points under the Claude Opus 5.5 flagship at 58.

Four models, then, and only one of them is a G​emini agent exclusive. Argon is the one the leak is really about, because it is the only name a reader cannot already reach.

Why G​oogle needs this to work, in one number

The arithmetic behind a multi-vendor agent is a price spread, and the spread is large. On Artificial Analysis's v4.3.2 revision the same six-point band runs from Claude Opus 5.5 at 58 down through Claude Sonnet 5.5 at 56, Claude Fable 5.1 and GPT-6 Astra at 53, and Claude Opus 5 at 51 — while Gemini 3.8 Flash sits at 40.9 at $0.75 / $3.75. G​oogle's pitch is that the agent picks per job: frontier reasoning on Argon for the hard 5% of a workflow, Flash for the volume. If Claude Opus 5.5 and Claude Sonnet 5.5 are also in the picker, the same choose-the-right-model logic extends across vendors, and G​oogle sells the routing rather than the model.

That is a good argument, and it is one you can act on without waiting for G​oogle's version of it. The mistake enterprises make with a new orchestration layer is assuming the routing is free because it is bundled. It is not: each surface has its own quota pool, its own admin console, its own spend cap semantics, and its own idea of what happens when a provider degrades. G​oogle's own Antigravity documentation is a live example — Claude and GPT quotas sit in a separate pool from G​emini models, on their own five-hour and weekly limits, and BYOK is not supported on the plans that serve them.

Anything you build against one vendor's agent is also built against its model list, which is exactly the list that this leak got wrong. One API for 200+ models with 0% markup on provider list price is the low-commitment way to test the same routing idea: run a task against Gemini 3.8 Flash and Claude Sonnet 5.5 from the same endpoint, compare the cost per task on your own traffic rather than on a vendor table, and let automatic failover across providers cover the case where the model you picked is the one having a bad hour. Vendor price changes pass through the same day, which matters when G​oogle has already told you Gemini 3.8 Flash doubles on 2027-01-01.

Wat zou dit van een lek naar nieuws verheffen?

Drie documenten zouden het beslechten, en geen ervan bestaat vandaag.

The first is a G​emini agent model-selection page that names versions — the equivalent of Antigravity's model table, which lists Claude Sonnet 5.5 and Claude Opus 5.5 by name and by plan. G​oogle's agent announcement has no such table, and the sentence about A​nthropic is a prose commitment rather than a specification. Until a table exists, every version number circulating is somebody's inference.

The second is a G​emini Business product page. Right now the only evidence for the name is an SSO redirect and one X post. If G​oogle is creating a tier between the consumer G​emini app and G​emini Enterprise, it will appear in the Cloud or Workspace blog with a price and a plan gate, and it will name the models the way Antigravity's docs name them.

The third is an Argon model ID. Argon is the whole commercial point of the arrangement — it is G​oogle's frontier line, priced at $2 / $10 intro, and it is not yet callable. When gemini-4 or an Argon identifier appears in the G​emini API's model list, the multi-model agent has something to route to that is not already on the market. Until then, the honest summary of "G​emini Argon 4" is that it is an announced model with a measured index score, a price, and no endpoint — which is a real position for a reader to plan around, and a poor thing to build on.

Wat te doen met dit als je deze week modellen kiest

Nothing in the leak changes what you should run today. Gemini 3.8 Flash, Claude Opus 5, Claude Opus 5.5, Claude Sonnet 5.5 and the rest of the field are all callable now, on their published prices, and the model-selection work is the same work whether G​oogle ships an agent or not.

The one decision the announcement does inform is architectural: if you are standardising on a single vendor's orchestration layer, you are also accepting that vendor's model list, that vendor's quota pools, and that vendor's pace of change. If you are standardising on an API that fronts several, you keep the choice — which is, word for word, the argument G​oogle made for its own product when it wrote that the leading model changes every few months and keeping the choice open means your context, skills and data stay put when it does. The difference is only who owns the routing.

What to watch is narrow and dated: an Argon model ID in the G​emini API's model list, and a G​emini agent model table that names A​nthropic versions. Either one turns this from a leak into a specification. Until both land, the accurate description of G​oogle's position is the one G​oogle published — Claude is in the agent, the version is unstated, and G​emini Business remains a login page and a rumour.