Ratings / Google
Gemini 4 Argon
Google · released Oct 2026 · gemini-4-argon · effort medium (default)
Gemini 4 ArgonwithAPIRanked #7 overall
46 runs · ±3.6 · Solid evidenceWhat you can hand it, and what you’ll still have to catch
Dropped 13 Oct 2026, against whatever leads today.
Sharp on people and plans. Shaky on its own sums.
Gemini 4 Argon is the best Gemini we've tested, and on people calls it's up with the leaders: its performance reviews score 93, a point behind GPT-6 Astra, and it took apart an AI SDR plan using only what the customer interviews said. But its Combined score is 69: six points up on Gemini 3.8 Flash, and 13 behind GPT-6 Luna. The reason is almost always the same. It recounts the data you give it, gets the count wrong, and builds the argument on the wrong number. The LLM judge found 15 of its 44 graded written outputs usable with at most a quick edit. Net result: hand it the judgement calls where the numbers are already done, and keep another model for anything that needs a total. At its introductory price it costs about the same per output as GPT-6.1 Sol, which scores 21 points higher, and that price doubles on 1 December.
What you can hand it
- People calls. Its performance reviews score 93, a point behind the leader. Its review of a PM who shipped a lot and moved nothing used only the evidence supplied and was ready to paste into the template.
- Finding the flaw in a plan. Its challenge to an AI SDR for small agencies scored 92. It argued from what the interviews actually said (these are high-trust purchases, and email is a small part of the job), not from generic objections. One fix: it reads the test out on signed deals, when booked meetings would show up well inside the 10 weeks.
- Roadmaps with a hard constraint. On a year of spend management (86), it finished the migration builds by late January to leave four weeks for processor certification, and showed the $6M procurement case needed about 81% adoption before cutting it back to a cheap test. On fourteen requests from one customer (77), it de-duplicated the big account's asks and weighed demand by how many customers asked.
- Prototypes that work. The clinic rebooking (86) and receipt-capture (92) prototypes both worked end to end. Neither said what it tests and what it fakes, so add that line before you share it.
What you’ll still have to catch
- Its own arithmetic. This is behind most of its misses. On the competitor analysis it counted 16 deals lost to the rival where the data shows 17, and on the support-ticket synthesis it counted 72 supplier tickets instead of 75. It got the survey segments wrong on a product-market fit call, mis-summed the traffic CSV it sized an experiment from, and built a pricing table on revenue and cost totals that don't match the data. Re-add every total.
- Context it makes up. It priced US sales reps at $120k–$150k a year in a challenge to a $5B plan, invented 7,000 dormant students in a marketplace strategy, put a 60–90 day hiring backfill in a QBR deck, and told the CEO Product was auditing invoice layouts this sprint. None of it was in the brief.
- Hunches written as findings. "Students are paralyzed by the paradox of choice." "January 1 is the easiest time to switch payroll providers." Both were stated as fact, with nothing behind them.
- No "what would change my mind". In 25 of its 44 written outputs it didn't name the unknowns that could flip the call. Both its North Star memos picked a metric and never said what result would change it.
- Gates that can't fail. Its PRDs and strategy memos set targets with no kill threshold or window. On the AI substitutions launch call it made a clear go/no-go for each region, but gave no signal that would trigger a rollback. And its disputes PRD added steps that made raising a dispute harder than making a payment: its one critical mistake.
- Best setupGemini 4 Argon · API68.7#7 overall
- Against the leader−20.8GPT-6 Astra · ChatGPT
Task by task
Its Combined score on each task type, across 46 outputs.
| Model | Overall | Product Craft | Growth | Metrics & Experimentation | Strategy | Leadership | ||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Discovery | Research guide | PRD | Roadmap | Prototype | Launch call | Activation | Growth loop | GTM plan | Retention | Test design | Test readout | North Star | OKRs | Strategy | Challenge | 1000x | Rivals | PMF call | Pricing | Exec update | Hiring | Reviews | ||
| GPT-6 AstraOpenAI | 89#1 | 96Best | 81 | 84Best | 94Best | 70 | 93 | 91 | 85 | 98Best | 100Best | 76 | 96Best | 89 | 99Best | 88Best | 82 | 91 | 92 | 98Best | 84 | 95Best | 84 | 90Best |
| Gemini 4 ArgonGoogle | 69#7 | 67 | 70 | 54 | 77 | 89 | 71 | 68 | 67 | 69 | 76 | 66 | 64 | 70 | 71 | 58 | 77 | 56 | 57 | 72 | 52 | 65 | 75 | 89 |
Combined scoreEach cell is the model’s best setup on that task type.
Every tested harness
Same model, different setup. The harness often matters more than the launch post says.
| Rank | Harness | Combined score | Decision model | LLM judge | Evidence |
|---|---|---|---|---|---|
| 7 | Gemini 4 ArgonwithAPIOpenRouter 2026-09 | 70.8 | 68.6 |
Published runs
46 outputs, each graded blind on its task’s current checklist
| Task type · task | Harness | Combined score |
|---|---|---|
| 1000x an ideaBudgets for two | API | |
| 1000x an ideaSharing a shopping list, 1000x | API | |
| Activation & onboarding reviewThe checklist isn't the moment | API | |
| Activation & onboarding reviewAnalytics tool losing users at setup | API | |
| Analyse experiment resultsUp 6%, and significant | API | |
| Analyse experiment resultsThe underpowered onboarding test | API | |
| Build a roadmapA year of spend management, with a hard deadline | API | |
| Build a roadmapFourteen requests from one customer | API | |
| Call product-market fitDo we have fit, or a spike? | API | |
| Call product-market fit38%, and the retention that isn't | API | |
| Challenge an ideaFive billion dollars, on paper | API | |
| Challenge an ideaAn AI SDR for small agencies | API | |
| Competitor analysisFree forever, funded | API | |
| Competitor analysisThe competitor isn't who you think | API | |
| Customer research call guideToo expensive, by default | API | |
| Customer research call guideWhy are warehouses leaving? Two execs, two theories | API | |
| Develop product strategySupply or demand for a stalled marketplace | API | |
| Develop product strategyEnterprise, says the CEO | API | |
| Diagnose a retention problemFlat, while sign-ups grow | API | |
| Diagnose a retention problemThe collapse that wasn't | API | |
| Draft OKRsA great quarter, on paper | API | |
| Draft OKRsEleven key results and a bonus | API | |
| Experiment specificationShowing the delivery fee up front | API | |
| Experiment specificationShow the fee sooner | API | |
| Extract discovery insightsHidden validation case | API | |
| Extract discovery insightsThe loudest ticket | API | |
| Find the growth loopGrowing on the surface, decaying underneath | API | |
| Find the growth loopThe badge on every form | API | |
| Hire a PMWhat our interviews predict | API | |
| Hire a PMA loop for the first growth PM | API | |
| Make the launch callGo/no-go for AI substitutions, from the rollout data | API | |
| Make the launch callOffline, on older phones | API | |
| One-shot prototypeClinic appointment rebooking | API | |
| One-shot prototypeExpense receipt capture | API | |
| Pick a North Star metricGrowing, on paper | API | |
| Pick a North Star metricRevenue, actives, or something else? | API | |
| Price and package a productFifteen dollars for everyone | API | |
| Price and package a productUnlimited pages, falling margins | API | |
| Write a GTM planPayroll, for the people who fear it | API | |
| Write a GTM planEnterprise, without killing the engine | API | |
| Write a performance reviewFourteen ratings, three managers | API | |
| Write a performance reviewShipped a lot, moved nothing | API | |
| Write a PRDDisputes that aren't disputes | API | |
| Write a PRDMeeting summaries with action items | API | |
| Write a stakeholder updateOn track, mostly | API | |
| Write a stakeholder updateA quarterly business review deck, in brand | API |