Is Gemini 4 Out? The Leaked 'Gemini 4 Pro' Chart, Fact-Checked — What Google Has Actually Said

Short answer: no. Gemini 4 is not out, Google has published no Gemini 4 benchmarks, and the chart is unverified. What Google has said, on the record, is narrow and recent: on July 21 the Gemini team wrote it had 'already started our most ambitious pre-training run yet, for Gemini 4'; the next day Sundar Pichai told investors 'We are now training Gemini 4'; and on September 23, DeepMind's Koray Kavukcuoglu said the model is in early post-training and that Google's 'intention is to, like — as soon as possible — to release an early post-training output,' hopefully 'much earlier' than the end of 2026 (9to5Google). No date, no model card, no price, no API. Everything else circulating this weekend — the benchmark table, the 2-million-token context, the API prices, the 'Argon' checkpoint on Arena — comes from anonymous accounts and blind-test screenshots, and none of it can be checked.

What the chart claims

The table was posted Sunday night by @JustLingonberry, an account with no stated source for it, and amplified Monday morning by @mark_k with the honest framing 'Huge if true.' As reported by the outlets that reproduced it, the chart gives 'Gemini 4 Pro' figures such as 88.7% on DeepSWE v1.1 (about fourteen points above the Opus 5.5 and Astra columns), 95.3% on Terminal-Bench 2.1, 72.1% on an 'HLE-Verified' benchmark, and 86.8% on OSWorld 2.0, plus a 2M context window and API prices below both rivals. We're not reprinting the table, because there's no evaluation log, no model ID, and no Google page behind any cell of it — and because, as one widely shared response put it, the whole thing could be fabricated in a spreadsheet in ten minutes.

Why it can't be verified — and the tells to check yourself

• No provenance. The poster gives no source — not an internal doc, a partner deck, or a screenshot of a Google property. Real Google launch charts arrive on blog.google and deepmind.google with a model card the same hour.
• Mixed benchmark versions. The chart uses Terminal-Bench 2.1 alongside DeepSWE v1.1 and OSWorld 2.0. Anthropic's and OpenAI's September launch tables both report Terminal-Bench 4.0, where Opus 5.5 scores 66.4% — so a 95.3% on the older 2.1 suite isn't comparable to anything either rival published this month, only to Vals AI's separate 2.1 run (Opus 5.5: 87.6%). A chart that cherry-picks suites across versions is a pattern, not a proof.
• The rival columns are the check. If the Opus 5.5 and Astra numbers in the table don't match those models' published scores — Opus 5.5's are on Anthropic's launch page and Artificial Analysis — the Gemini column is fiction too. Outlets reproducing the chart haven't done that reconciliation; neither could we from the image alone.
• The Arena angle is speculation on top of speculation. Posts claim a Gemini 4 Pro checkpoint, codenamed 'Argon,' has been served on Arena under the label of Gemini 3.8 Flash since mid-September. Arena hasn't commented, Google hasn't confirmed, and there is no Gemini 4 entry on the leaderboard — what exists is testers saying a 'Flash' model answered too well. Anonymous frontier checkpoints on Arena are normal for every lab; naming one is guesswork.

What Google has confirmed, in full

Why the timing makes people want it to be true

OpenAI's DevDay is tomorrow, September 29. Anthropic shipped Opus 5.5 six days ago and took #1 on the independent indexes. Google's last flagship-tier release was the Gemini 3.8 line, and in our chatbot ranking Gemini's case rests on the free tier and the Google integration rather than the top model. A leak that has Google leapfrogging both rivals the night before DevDay is exactly the story half of AI X wants — which is the reason to hold it to a higher bar, not a lower one. Kavukcuoglu's own phrasing is worth rereading: Google wants to ship an early post-training output fast. That's a company signalling it will release something before it's fully polished, not one sitting on a finished model that already beats the field.

How you'll know when Gemini 4 is real

• A post on blog.google or deepmind.google with a model card, and a gemini-4 model ID on ai.google.dev with a price row. Google has never launched a flagship without all three.
• Third-party scores within 24 hours from Artificial Analysis, Vals, and Arena under the model's real name — the same trio that scored Opus 5.5 and GPT-6 Sol last week.
• Availability in the Gemini app for AI Pro and Ultra subscribers, usually the same day as the API.
• Until then, treat every 'Gemini 4 Pro' number as a rumor, and every 'Gemini 4 release date' page — including the dozen that appeared this weekend — as a guess.

Our read

The interesting fact isn't the chart; it's the calendar. Google confirmed in July that Gemini 4 is a bigger model than anything it's shipped, and confirmed last week that it's already in post-training with a stated urge to release early. That's real, and it means the two-lab race we wrote about after Opus 5.5 becomes a three-lab race before the year is out. What the weekend added is noise — a table nobody can trace and an Arena guess nobody can confirm — timed to land under DevDay. We'll update our Gemini review the day Google publishes a number, and not for a screenshot. If the chart turns out to be accurate, we'll say so here; that's a lower bar for Google to clear than for an anonymous account to have cleared already.