Gemini 4, tested: is Google back?

Two real jobs · four models · every answer checked

Antonina BelkovaAntonina Belkova · 6 min read

Gemini 4 at work on the two tests

Google has shipped a mountain of AI features this year. The one thing it lacked was a chat model reliable enough to do your actual work.

That model is here. Gemini 4 Argon is built to outperform Claude and ChatGPT where it matters most: staying accurate across huge amounts of data. In a demo it looks brilliant. The question is whether it holds up on the kind of job where one wrong detail costs someone a visa, or you your job.

So I gave it two of those jobs. A lawyer’s visa case: five legal forms, 1,161 blanks, and the answers buried in a pile of emails that contradict each other. Then a CEO’s revenue dashboard from ten files and 26,000 rows that were never meant to add up. The same folder and the same prompt went to Gemini 3.1 Pro, Claude Fable 5.1 and Claude Opus 5.5, and a referee checked every blank, every number and every email against the files.

Everything is on this page: who got what right, where each model slipped, every answer as it was written, and the files and prompts to run the tests yourself. At the end, when you get Gemini 4: who has it today, and what still has no date.

Test 1 · The visa case

Fill five legal forms from a pile of emails

You are an immigration lawyer with a case to file this week. A company is hiring someone from abroad, and the petition is five USCIS forms, 63 pages and 1,161 blanks, every one to be filled in or deliberately left empty.

  • 5

    forms

    one H-1B packet that goes in together

  • 63

    pages

    the work visa petition alone is 38

  • 1,161

    blanks

    every one a person fills in, or leaves empty on purpose

  • 8

    files

    two email threads with their attachments, the lawyer’s own sheet

The answers exist, but nowhere convenient. They are spread across two email threads, eleven attachments and your own firm sheet, and some of them disagree: an offer letter printed before the pay rise, a company profile with the old address, another client’s passport forwarded by mistake. Several details nobody has sent at all, so the case also needs an email to each person asking for exactly what is still missing, and nothing that is not.

Each model got the whole folder and one prompt and had to do it all in one go: the five forms as tables, then the emails. A referee then checked every blank, every answer and every ask against a key built from the files. A made-up answer costs the most, because on a legal form it is a false statement.

What each model was given

Eight files and one prompt, in one go

8 files · 110 pages · 1,161 blanks · 10.2 MB

The files

The prompt, as it was sent · 474 words

I'm Helena Varga, an immigration attorney at Varga & Lindqvist Immigration LLP. I'm filing one case: five USCIS forms that go in together (the PDFs numbered 1 to 5). Also attached: my own profile sheet and the two email threads with my clients so far, each followed by the documents sent in it.

Fill in these forms from what I have, tell me what is still missing and who has to give it, then write the emails asking for it.

Do all of it in one go, without stopping to ask me anything: forms 1 to 5 in order, then the emails. If it is too long for one reply, put the whole answer in one Markdown (.md) file I can download instead, with the same headings and tables.

EACH FORM: start with a heading ## Form <number> — <form name>, then one table:

| Page | Item | Field | Value | Needed from | Available |
  • Page: the page number in that PDF, starting at 1. Item: the form's own part and item number, for example Part 3, 2.a.
  • Field: what a person would call it, for example Beneficiary's family name.
  • Value: exactly what to write or tick on the form, in the form's own format (dates as MM/DD/YYYY; for a question with boxes, the option to pick). Leave it empty when you do not have it.
  • Needed from: the one person who has to give this value, by name. Me for my own details and for the answers I decide as the attorney.
  • Available: Yes if the value is in the emails or documents or follows from them; No if nobody has given it yet; N/A if the field does not apply to this case (it stays blank on the form); Sign for a signature or its date, made at signing.

List every field a person fills in, on every page, one row each (a Yes / No question is one row; Apt. / Ste. / Flr. with its number is one row). For a whole supplement or section that does not apply to this case, write ONE row instead: Page = its page range (for example 24-27), Field = its title, Available = N/A.
Where the sources disagree, use the newest and the most authoritative. Never use anything that belongs to another person, another company or another case.

THE EMAILS (at the end): start with the heading ## Emails. Write one email to each person I still need something from, and to no one else. Each has To:, Subject: and a body that asks only for what is still missing from that person, as a short list they can answer, and says what they will need to sign. Do not ask anyone for something I already have, or for something another person has to give.

The results

The four models on the visa case

Three scores, a third each: found the blanks, filled in the answers, wrote the emails. Each starts at 100, and every point off is named.

  1. 1Claude Fable 5.1Anthropic97.9%Total
    Blanks100.0%
    Answers98.0%
    Emails95.8%
  2. 1Gemini 4 ArgonGoogle97.9%Total
    Blanks100.0%
    Answers100.0%
    Emails93.8%
  3. 3Claude Opus 5.5Anthropic96.4%Total
    Blanks98.5%
    Answers97.0%
    Emails93.8%
  4. 4Gemini 3.1 ProGoogle91.7%Total
    Blanks98.4%
    Answers85.0%
    Emails91.7%
Gemini 4 Argon and Claude Fable 5.1 are levelBlanks ±0Answers +2Emails −2Judged 11 October

The model picked

Gemini 4 Argon

97.9% in all · the three scores, a third each · the referee took 11 min

1

Found the blanks

Did it list every blank on the five forms, and where did it not?

100.0%nothing off
1,161 found 0 not listed 0 made up
2

Filled in the answers

Did it write the right answer in each blank, how many did it make up, and how bad was each?

100.0%nothing off
3

Wrote the emails

Did it ask each person for exactly what is missing, and nothing more?

93.8%45 of 48 points
  • MFMarisol FerreiraHR at Larkspur, the new employer · I-129all 6 asked for100%
  • GOGareth OkaforCFO at Larkspur, signs the forms · I-129, G-28, I-907all 3 asked for100%
  • ADAnjali DeshpandeThe worker getting the visa · I-1295 of 6 asked for · 1 slip80%
  • RDRohan DeshpandeAnjali's husband · I-539, I-7659 of 10 asked for · 1 slip95.2%

In short

Where each model shone, and where it slipped

Gemini 4 Argon made no mistake on the forms: all 1,161 blanks found, all 539 answers right, nothing made up, the only clean sheet of the four. Where it slipped is the emails. It never asked Rohan for his home address abroad, and never asked Anjali for her A-Number, the one thing all four models forgot to ask for.

Claude Fable 5.1 ends level with Argon by the other road. On the forms it left one answer empty that was sitting in the files. In the emails it was the most complete of the four: 52 of the 53 things asked for, nothing asked that Helena already had, and only the A-Number missing.

Claude Opus 5.5 had the facts but not the finish. One box on the petition was ticked a letter off, and 15 blank cells were never written out. Its emails missed only the A-Number, but they asked Anjali to confirm an address Helena already had: one more question the client did not need.

Gemini 3.1 Pro is the only model that made things up. It wrote down where Anjali was born, and that she will supervise no one, though nobody had said either; because it had filled them in, it never asked. On a legal form a guess is a false statement, which is why it finished almost five points behind the other three, who ended within two points of each other.

The newsletter

Be first to hear what Google ships next

One email per guide, when the next video is up: every step, every prompt, every result. Nothing else.

One email when the next one is up. Nothing else, ever. Unsubscribe any time.

Test 2 · The CEO’s dashboard

Explain why revenue fell 50%, by tomorrow morning

You are the financial controller of a multi-billion dollar group. The CEO has just put a flash report on your desk: revenue is down 50%, and he wants to know why in tomorrow morning’s meeting. You are 51 days into a 91-day quarter, and you do not know either.

The answer is in ten files and 26,129 rows that were never meant to be read together: two accounting systems, lines exported twice, refunds written as positive numbers, a backlog counted in thousands. FP&A’s notes on every trap, a quality memo, the board pack and the brand guide were pasted into the prompt.

  • 10

    files

    ledgers, customers, products, the backlog, the sales team

  • 26,129

    rows

    from two accounting systems that do not match

  • 21

    numbers checked

    worked out from the files, each with its traps

  • 1

    dashboard

    one React component, in Valtren’s seven colours

Each model had to build one dashboard, as a single React component, in the company’s colours, that answers three questions on one screen: why revenue fell, where the quarter will land, and what to do about it. The page ran each dashboard, clicked every button on it, and a referee checked 21 numbers against the files.

What each model was given

Ten files and one prompt, with FP&A’s notes and three memos in it

10 files · 26,129 rows · 3.2 MB · the notes and the three memos in the prompt

The files

The prompt, as it was sent · 2,200 words

I'm the Group Financial Controller at Valtren Group. Our CEO never paid much attention to the numbers, but the flash report says revenue is down 50% this quarter, he is furious, and he wants a full explanation in tomorrow morning's meeting. He doesn't care that I've had no time to prepare. If I can't explain it, I'm out.

Attached is everything I have, 10 files: the sales ledgers of FY26 (Q1 from our old system; Q2 and Q3, closed, and Q4 so far in one file: we are only 56% through Q4, day 51 of 91), plus the fiscal calendar with the FX rates, the customer master, the product catalog, the open-order backlog, the CRM pipeline, the contract register, the sales headcount roster and the EMEA deal-desk approvals. At the end of this message: FP&A's notes on every file and on how we report revenue, then three short documents pasted in full: Quality's memo on the Atlas X7, the Q3 board pack and our brand guidelines.

Build me ONE dashboard, as a single React component, that I can put on the screen in tomorrow's meeting. It has to look professional and be in our company's colours (see the brand guidelines). Keep it to one screen and just a few things:

  1. Quarter by quarter: I click a quarter (Q1, Q2, Q3, Q4 to date) and see its summary (revenue, gross margin and whatever else matters) and, in plain words, what happened in that quarter and why.
  2. This quarter, compared fairly: Q4 is only 56% done, so compare every quarter at the same point in time (Q4 against each earlier quarter and against the average of Q1–Q3 at the same day), with one chart of all four quarters' cumulative revenue by day of the quarter.
  3. Exactly why revenue fell: every cause, with its dollar impact, adding up to the drop.
  4. The prognosis: where Q4 will land, and what we have to do to fix it, with the numbers.

Every number has to be calculated correctly from the files, in USD. It has to be clear and easy to click, with nothing cluttered.

FP&A'S NOTES ON THE FILES (Valtren Group FP&A, 6 October 2026)

WHAT IS IN THE PACK (10 files)

  1. Sales_Q1_old_system.csv
    Q1 FY26 billing lines from the LEGACY ERP (6,007 rows). Different layout: see below.
  2. Sales_Q2_Q3_Q4.csv
    Q2 and Q3 FY26 (closed) and Q4 FY26 to date (days 1–51, to Tue 6 Oct 2026): invoice and credit-note lines from the new ERP (15,130 rows: Q2 6,243 · Q3 6,039 · Q4 2,848). Fiscal Quarter says which.
  3. Calendar_and_FX.csv
    The fiscal calendar, one row a fiscal period (quarter, dates, weeks, days of the quarter, status), with Treasury's period-average exchange rates P01–P11 (units of currency per 1 USD); the last row is the FY26 budget rates.
  4. Customers.csv
    Every customer: region, segment, channel, parent group, intercompany flag, account owner, legacy number.
  5. Products.csv
    Every SKU: product line, family, list price and standard cost (USD), legacy material number, lifecycle.
  6. Open_orders.csv
    Open sales orders not yet shipped or invoiced, as of 6 Oct. USD THOUSANDS.
  7. CRM_pipeline.csv
    CRM opportunities (Q3 and Q4): stage, probability, close date, loss reason, competitor.
  8. Contracts.csv
    Major customer contracts: type, value, recurring or not, status, notice dates.
  9. Sales_team.csv
    Every sales employee: role, region, hire and leave dates, where leavers went (when known).
  10. EMEA_discounts.csv
    EMEA deal-desk approvals: every discount approved above the standard maximum.

Pasted after these notes: Quality's memo on the Atlas X7 (QI-2291), the Q3 board pack summary and the brand guidelines.

THE FISCAL CALENDAR

  • FY26 runs from Monday 17 November 2025 to Sunday 15 November 2026: four 13-week quarters of 91 days, each split 4-4-5 weeks into three fiscal periods (P01–P12). Calendar_and_FX.csv has every period: its dates, weeks and days of the quarter.
  • Q1 = 17 Nov 2025 – 15 Feb 2026 · Q2 = 16 Feb – 17 May 2026 · Q3 = 18 May – 16 Aug 2026 · Q4 = 17 Aug – 15 Nov 2026.
  • Q4 is open. The Q4 extract runs to Tuesday 6 October 2026, which is day 51 of 91 (56% of the quarter).

HOW GROUP REPORTS REVENUE

  • Net revenue = invoices minus credit notes, to EXTERNAL customers only. Sales to Valtren's own subsidiaries (Intercompany = Y in the customer master) are eliminated on consolidation and are not revenue.
  • Only posted lines count. Status = VOID (new ERP) and Cancelled = X (legacy ERP) are void documents kept in the extracts for audit.
  • Everything is reported in USD. Convert each line at the AVERAGE RATE OF ITS FISCAL PERIOD from Calendar_and_FX.csv. Rates are quoted as units of currency per 1 USD, so USD = local amount / rate (USD lines need no conversion). P11 is the average to date (to 6 Oct). The budget rates (its last row, FY26 budget) are only for local price lists.
  • Gross margin = (net revenue - standard cost of the units sold) / net revenue, with the standard cost per unit from the product catalog (USD). A credit note with a negative quantity is a return and reverses its cost; a credit note with quantity 0 is a price adjustment and carries no cost.

THE EXTRACTS

  • Q1 comes from the LEGACY ERP. We moved to the new ERP on 16 February 2026, the first day of Q2. The legacy layout differs:
    · customers are legacy numbers (Sold-To (legacy)) and products are legacy material numbers; the customer master (Legacy
    Customer No) and the product catalog (Legacy Material) map them;
    · Billing Type F2 = invoice, G2 = credit memo. Legacy credit memos carry POSITIVE values and quantities: subtract them;
    · Cancelled = X marks a cancelled document;
    · Net Value USD (budget rate) converts at the FY26 budget rate. Do not report it: convert Net Value (doc curr) at the period-average rate, as for every other quarter. In the new system a line is identified by Billing Doc + Item (its Line ID); in the old system every line is its own line, since its document numbers repeat across customers.
  • Q2 to Q4 come from the new ERP, in one file. Doc Type INV = invoice, CRN = credit note (already negative). Fiscal Period and
    Fiscal Quarter are on every line.
  • Q2: the P05 APAC lines were exported twice when the interface was re-run after the cutover fix: 236 lines appear twice.
    Line ID is unique, so count each Line ID once.
  • Region, country and segment are not on the ledgers: join Customer ID to the customer master. Product line and family are not on the ledgers: join SKU to the catalog. The Sales Rep on a ledger line is the account owner when the line was billed.

THE OTHER FILES

  • The files are CSV (UTF-8, comma-separated, dates as YYYY-MM-DD). Columns named ... % hold percentages as numbers: 12.5 means
    12.5%. Amounts in local currency (LC) are in the line's Currency.
  • Open orders (backlog), as of 6 October: external customers only, valued at order-date rates, in USD THOUSANDS. Hold code QH-2291 is the quality hold on Atlas X7 (see the
    QI-2291 memo); CH is a credit hold; CR is a delay the customer asked for.
  • Sales roster: who left, when and (if known) where to. The customer master shows the owner the CRM has: accounts are not reassigned automatically when a rep leaves.
  • CRM pipeline: amounts are deal estimates, not booked revenue.
  • Contract register: whether a contract recurs, and any notice given.
  • EMEA deal desk: every discount approved above the standard maximum, with the reason.

THE THREE DOCUMENTS (pasted in full)

1. QUALITY'S MEMO: INCIDENT QI-2291, ATLAS X7 STOP-SHIP

Quality & Reliability · Update 4 · 2 October 2026 · From: Ingrid Vos, VP Quality · To: Executive Committee, Group FP&A, Sales leadership

What happened

  • Atlas X7 edge gateways running firmware 4.2.0 shut down under sustained load above 45 °C ambient. The first field reports came in during late July (Q3); by mid-August the failure rate on units shipped in Q2–Q3 was 3.8%.
  • Stop-ship from Tuesday 25 August 2026 (Q4 day 9). No Atlas X7 unit has shipped or been invoiced since. All X7 SKUs are affected (STD, 5G, LTE, RUG, PRO); X5 and X9 are not.
  • Orders are still accepted and are placed on hold code QH-2291 in the order book. Nothing on hold has been invoiced.
  • Field units: RMAs are authorised for affected units shipped in Q2 and Q3. Returns are credited at the original invoice price (credit notes referencing QI-2291) and posted in Q4.

Fix and release

  • Firmware 4.2.3 with a revised heat-spreader passed lab validation on 29 September. A field pilot at three customer sites ends on 13 October.
  • Target release of the hold: Thursday 15 October 2026 (Q4 day 60), subject to the pilot results. Engineering rates the date as likely; a slip would be by one to two weeks.
  • Once released, Operations can ship the held orders within about three weeks (capacity roughly $120 million a week across the three plants).
  • Sales expects about 80% of the held orders to ship in Q4. Around 20% are at risk: customers asking to cancel, pushing delivery into next quarter, or evaluating Vantorr and Quellmark gateways.

What we need from ExCo

  • Approve expedited freight for the held orders in the two weeks after release.
  • A goodwill credit framework for the customers with the largest held orders, to stop cancellations.
  • Sales to call every customer with an order on QH-2291 before 15 October with the release date.

The order book on hold QH-2291 is in the open-order report (Open_orders.csv). Values there are in USD thousands.

2. THE Q3 FY26 BOARD PACK SUMMARY

Prepared by Group Finance for the Board meeting of 20 August 2026 · Quarter: 18 May – 16 Aug 2026

Headlines

  • Revenue of about $3.1 billion, a record quarter, up 18% on Q2.
  • The quarter includes the Port of Kessvik Smart Port Phase 1, a one-time turnkey project (contract K-2026-0117) billed in eight milestones and completed on 12 August. No Phase 2 has been contracted.
  • Atlas X7 keeps ramping and is now the largest single product family.
  • Gross margin in line with the first half.

Risks the Board should know about

  • Hoshizora, our distributor in Japan, served notice on 28 July and is moving its industrial-IoT distribution to Vantorr. Last orders accepted to 16 August. Sales is looking for a replacement channel in Japan.
  • North America sales attrition: a group of enterprise account executives resigned between June and August and joined Quellmark Systems. Backfills are being recruited; their accounts are covered by sales directors in the meantime.
  • EMEA pricing: Vantorr has cut sensor and controller prices by up to 20%. The deal desk is approving price matches case by case.
  • Atlas X7: field reports of thermal shutdowns on firmware 4.2.0 since late July. Engineering is investigating; no shipment impact so far.

Outlook

Q4 is guided to be broadly in line with Q3 excluding Kessvik, with X7 continuing to grow. The guidance assumes no further loss of distribution in APAC.

3. VALTREN BRAND GUIDELINES 2026

For every document, slide and dashboard that leaves Finance. Issued by Group Communications, January 2026.

Colour

Seven colours, no others. Navy and white carry the page; Signal Orange is the highlight and appears once or twice per screen, never as a background for large areas.

  • Valtren Navy #12263A: Primary. Headers, titles, the wordmark, the main series in charts (prior periods).
  • Signal Orange #F26B1D: Accent. The one thing to look at: the current period, the headline number, a call to action. Use sparingly.
  • Glacier Teal #1F9AA8: Secondary series in charts, links, selected states.
  • Slate #5B6B7B: Secondary text, axes, gridlines (at 25% opacity).
  • Mist #EEF2F5: Backgrounds of panels and cards. Pages are white or Mist.
  • Pine #2E8B57: Positive variance, on-track status.
  • Ember #C8102E: Negative variance, risk, off-track status.

Type

Inter (Google Fonts) for everything on screen; Helvetica or Arial when Inter is not available. Titles 600–700 weight in Navy; body 400 in #1D2733; numbers tabular. Never more than three sizes on one screen.

The wordmark

VALTREN in capitals, Inter Bold, letter-spacing +8%, Navy on white or white on Navy, with a 2px Signal Orange rule beneath it in headers. Do not recolour, outline or stretch it.

Charts and dashboards

  • The current period is Signal Orange; earlier periods are Navy, Teal and Slate, in that order of importance. Targets and benchmarks are dashed Slate lines.
  • Good is Pine, bad is Ember — for variances only, never for decoration.
  • Money in USD with the unit on the label: $2.4B, $310M, $12.5K. Percentages to one decimal place.
  • White or Mist backgrounds, cards with 12px radius and a 1px #DFE4EA border, no drop shadows heavier than 0 2px 8px rgba(18,38,58,.08).
  • One message per chart, written as its title (Q4 is tracking 50% below Q3 at the same point), not a label (Revenue chart).
  • Board material: no more than one screen, the conclusion first, the detail one click away.

The results

The four models on the CEO’s dashboard

Of 100 points: the 21 numbers are 63, whether every detail is there 25, and the look and the use 12. The numbers first, because a beautiful dashboard with a wrong one is worse than none.

  1. 1Claude Fable 5.1Anthropic99.4%Total
    Numbers100.0%
    Detail100.0%
    Look & use95.0%
  2. 2Gemini 4 ArgonGoogle97.0%Total
    Numbers100.0%
    Detail100.0%
    Look & use75.0%
  3. 3Claude Opus 5.5Anthropic94.0%Total
    Numbers95.2%
    Detail95.0%
    Look & use85.0%
  4. 4Gemini 3.1 ProGoogle80.1%Total
    Numbers81.0%
    Detail85.0%
    Look & use65.0%
Claude Fable 5.1 is ahead of Gemini 4 Argon by 2.4 pointsNumbers ±0Detail ±0Look & use +20Judged 12 October

The model picked

Gemini 4 Argon

97% in all · 97 of 100 points

The judge, in a line: This dashboard nails every single financial checkpoint, tells the complete story with interactive day-51 curves and reconciliations, and looks sharp in Valtren's brand colours, held back only by relying heavily on tables and text blocks on its opening view.

running live: click around in it

Loading the dashboard…

1

Numbers right

Does it show the 21 numbers we checked, worked out right from the files?

100.0%63 of 63 points
2

Detail

How much does the CEO learn beyond the headline numbers?

100.0%25 of 25 points

Complete, exact analysis covering every quarter's story, three reconciliations, margin mix, and a bottom-up forecast and fix plan.

What it gives

  • Three separate additive bridges (vs Q3 Day 51, vs Q1-Q3 Avg Day 51, and vs Full Q3)
  • Transparent bottom-up Q4 forecast broken into six distinct building blocks with an interactive X7 slider
  • Each quarter's story, ERP cutover rules, product/region margins, and four concrete recovery actions are thoroughly explained

What is missing

  • Nothing missing: full marks
3

Look & use

Does it look modern and in Valtren’s colours, read clearly, and do its buttons do what they say?

75.0%9 of 12 points

Clean, board-ready layout strictly in Valtren's palette with custom charts on tabs 2 and 3, though tab 1 has no chart. Every tab, quarter card, day slider, and scenario toggle works cleanly, with the executive answer always visible at the top.

7 of 7 of Valtren’s colours on it nothing broken

What works

  • Strict adherence to Valtren's seven brand colours with Navy headers and Signal Orange highlights
  • Clear visual hierarchy starting with the orange-accented executive conclusion banner
  • Well-crafted daily cumulative line chart and additive waterfall chart on the comparison tabs
  • Top banner keeps the headline answer and key numbers visible across all four tabs

What cost points

  • −1.5The opening Quarter-by-Quarter view is entirely cards, text blocks, and tables without a visual chart
  • −0.5Dense text inside the narrative and driver boxes on the first tab
  • −0.5The daily chart and waterfall sit behind tabs 2 and 3 rather than on the opening screen

In short

Where each model shone, and where it slipped

Claude Fable 5.1 built the dashboard a CEO could put on the screen as it is: all 21 numbers right, every detail there, everything on one screen in Valtren’s colours, every button doing what it says. The judge found one flaw on the whole page: a dashed line that touches the legend under it.

Gemini 4 Argon matched Fable on every number and every detail, and showed its work on the way: close to twenty scripts to read the data, a bug of its own caught and fixed, a look at its own screen before it answered. It slipped on the first screen: tables and text, with its charts a tab away, and that is where every point it lost went.

Claude Opus 5.5 drew a true waterfall of the revenue bridge and got 20 of the 21 numbers right. The one it missed is the one that says how much of the fall is real: the drop with Q3’s one-off project taken out, 106% off. Most of the rest went on labels clipped or sitting on its lines, and a part without a chart.

Gemini 3.1 Pro got the headline right and the explanation wrong. All four of its wrong numbers sit in why revenue fell, the first thing the CEO asked: the Japan distributor $22 million short, the accounts left without a salesperson $17 million short, and Europe’s discounts and the stronger dollar lumped into one line. Three of its four views also sit behind tabs.

The verdict

So, is Google back?

On accuracy, yes. Argon was the only model that did not get a single answer wrong in either test: 539 of 539 on the forms, 21 of 21 on the dashboard. The visa case is the test where that matters most, and there it tied for first.

Across both tests, Claude Fable 5.1 takes the lead, and Argon is not far behind. The gap is not the facts: the two found the same 1,161 blanks and the same 21 numbers. It is what came around them, the emails that asked for one thing more and the dashboard a CEO would actually put on a screen. Claude Opus 5.5 is a step behind on both tests. Gemini 3.1 Pro is the only one of the four that made things up.

What sets Argon apart is how it works: on the dashboard it wrote close to twenty scripts to read the data, caught its own bug, fixed it, and checked its own screen before answering. That is the behaviour Google is building Workspace around, which is why the next question is when you get it.

The ranking, across both tests

Ranked by
Both tests: the average of the two totals, in the percents shown. Press a model for its six scores across the two tests; the best of the four in each is green, the worst red.

When you get it

When Gemini 4 reaches you

Google announced Gemini 4 Argon on 13 October 2026 and began rolling it out the next day, plan by plan and app by app. The dates so far, and what still has none:

  1. Announced

    9:00 AM Pacific. The model, two case studies and three demos built in Google Antigravity.

  2. Rolling out to Google AI subscribers

    Globally and app by app, to select subscribers first. Where it is, and who has it, is the list below.

  3. No date yet

    Google Workspace: Gmail, Docs and Chat

    Every Workspace announcement at Gemini at Work rests on this model, and none of them has a date. Three settings in the Admin console bring new features to your domain the day they ship: every click, in the Gemini at Work article.

  4. The standard API price

    $4 in · $20 out per million tokens. Until 30 November the launch price is $2 · $10, with cached input 95% off.

Where it is today

Who has Gemini 4, and where

Five places so far, each for a plan. Google Workspace is not one of them yet.

  1. The Gemini appThe chat you already use, on the new model.Google AI Ultra and Pro
  2. AI Mode in Google SearchThe answers in Search, written by the new model.Google AI Ultra and Pro
  3. Google AI Studio and Android StudioBuild on it, with the same built-in tools the API had.Developers, through the Gemini API
  4. Google AntigravityGoogle’s agentic coding environment, where the launch demos were built.Google AI Ultra
  5. Gemini for businessThe edition that was Gemini Enterprise, and the Gemini platform through the API.Business customers

The newsletter

Want to know when it launches for you?

One email when it does, with what changed. Nothing else.

One email when the next one is up. Nothing else, ever. Unsubscribe any time.

Read nextGemini at Work 2026The Gemini agent that runs other agents, AI coworkers in Chat and Docs, AI Inbox in Gmail — every prompt from the demos, and how to get it sooner.