Gemini 4 Argon wins 8 of 15 headline benchmarks — and loses the hardest coding ones
Asked (summary):
Data-driven research on Google's Gemini 4 Argon, with graphs and the benchmark numbers.
Gemini 4 Argon is Google DeepMind's frontier model, announced 30 September 2026 for long-horizon reasoning, coding, enterprise knowledge work, multimodal understanding and defensive cybersecurity. This report covers 15 canonical benchmark comparisons against named rivals (two near-duplicate rows were dropped from a 17-row set) plus launch facts from Google's announcement and independent trackers, as of 20 October 2026. Access is still phased — cyber defenders and trusted testers first — so most rows cannot yet be independently reproduced.
Benchmark margins: Argon vs its named comparator
Each row is one benchmark; the circle is Argon, the open circle its comparator, sorted by Argon's effective margin. The Gray Swan row measures attack success rate, where lower is better — its bar colour already reflects that reversal. Scores are not averaged: the benchmarks measure different things on different scales.
Argon ahead
Argon behind
Tie
Independent evaluation
Google / vendor-reported
Scorecard: 8 wins, 2 ties, 5 losses
Against the named comparator on each of the 15 benchmarks, after reversing the lower-is-better Gray Swan row.
Where it wins and where it loses, by workload
Effective margin in points versus the comparator (Gray Swan reversed so right always means Argon is better).
Launch facts
Announced30 September 2026, by Google DeepMind —
blog.google
AccessPhased: trusted cyber defenders (Fairwind Program) and testers first; paid API customers and Google AI Ultra subscribers planned next. Not broadly available —
blog.google
Output limit1,000,000 tokens, up from 64K —
blog.google
Input contextIndependent listings record 1M tokens; Google's announcement does not state the public API input limit —
aievals.app
Introductory pricing$2 per million input tokens, $10 per million output; standard $4 / $20 after the intro period —
blog.buildfastwithai.com
Cached input95% discount off the input rate —
blog.google
Artificial Analysis Intelligence Index53, tied with GPT-6 Astra —
dev.to
Cost per Index taskAbout $1.99 at introductory prices; Argon is verbose, using ~110M output tokens against a median of 81M —
thejayant.in
Read the numbers with care
The benchmark table mixes independent third-party tests, Google-run tests, and figures Google quotes from other labs' reports or leaderboards. Harnesses, reasoning settings, video frame budgets, safety filters and test subsets differ: DeepSWE used a minimal agent harness, OSWorld 2.0 is a partial best-of-three offline subset, video comparisons gave Gemini 1 frame per second while rivals got fixed 300–800 frame budgets, safety-filter blocks counted as empty answers only for Argon, and restricted access prevents broad reproduction. Treat the rows as workload signals, not a single definitive ranking —
thejayant.in,
aievals.app.
Largest win: Harvey Legal Agent, 19.6% vs GPT-6 Astra's 5.4% (+14.2 pts) — professional knowledge work is Argon's clearest edge.
blog.buildfastwithai.comOn indirect prompt injection, Argon's attack success rate is 0.7% vs Astra's 8.5% — a 12× lower hit rate, the one row where a small number is the good one.
iotdigitaltwinplm.comLosses cluster in agentic coding and computer use: −10.5 pts on FrontierSWE v2 and Terminal-Bench Science, −9.0 on Terminal-Bench 4.0 vs Claude Opus 5.5.
blog.buildfastwithai.comIndependent evaluators partly agree: Vals AI puts Argon first on its Index (68.9 vs Opus 5.5's 67.0), and Wiz scores it 70.9 vs 58.2 on penetration testing.
tomsher.com,
torontodadlife.com
All 15 benchmark rows
| Benchmark | Workload | Comparator | Argon | Rival | Margin | Evaluator | Source |
|---|
Method: 15 canonical benchmark comparisons retained from a 17-row shortlist (a legacy Artificial Analysis Index v4.3.2 row and a duplicate generic OSWorld-2.0 row were dropped); scores are percentages or index points as each benchmark defines them, compared against the single named comparator per row; margin is Argon minus comparator in points, reversed in the win/loss tally for the lower-is-better Gray Swan row. Rows span six source hosts; the most-cited single URL backs 7 of 15 rows. Product facts come from Google's 30 September 2026 announcement, AIEvals and two independent reviews, read up to 20 October 2026. Benchmarks mix independent and Google/vendor-reported figures and are not averaged.