Google DeepMind

Gemini 4 Argon

Google's Gemini API docs say gemini-4-argon is now available (checked 2026-10-19) and is not on the Free tier, ending the announced-only state this page carried after the September 30 post. Introductory API pricing is $2/$10 per million tokens through 2026-11-30, then $4/$20. Three independent scores remain: AA HLE 57.09, AA-AnalystAgent 48.75, and vals.ai Terminal-Bench 4.0 at 57.58, all at high effort where tagged.

Gemini 4 Argon benchmarks and pricing, every number sourced: 5 tracked Gemini 4 Argon benchmark scores (3 independently run, 2 still resting on a vendor’s own claim), priced at $4.00 per million input tokens and $20.00 per million output.

Released
2026-09-30
License
proprietary
Context window
1M tokensvendor-stated (ai.google.dev/gemini-api/docs/models/gemini-4-argon: unified context window 1,048,576; checked 2026-10-19)
Knowledge cutoff
Not disclosed
Verified
sources checked 2026-10-02–2026-10-25
Parameters
Not disclosed
Architecture
Not disclosed

Gemini 4 Argon’s verified record

Gemini 4 Argon’s most-compared rival is Gemini 3.8 Flash: 2 leads, 1 tie, and 1 not callable across their 4 shared comparisons. Gemini 4 Argon is priced at $4.00/$20.00 per 1M tokens (in/out) vs Gemini 3.8 Flash’s $0.75/$3.75.

Against the 118 head-to-head comparisons Gemini 4 Argon shares with other tracked models:

30 real gaps5 inside the band83 not called

A gap counts for Gemini 4 Argon only where both sides were run independently and the benchmark still separates models. Losses are listed alongside wins on purpose — Gemini 4 Argon trails on 1 of them.

HLE · no tools Reasoning

±2 is noise
Ahead
Kimi K3.1 Pro +4.6 · MiMo-V2.6-Pro +7.7 · MiniMax M3.1 Flash +9.1 · Gemini 3.8 Flash +9.3 · Gemini 3.1 Pro Preview +10.1 · Kimi K3 +10.2 · Step 5 Preview +10.6 · Qwen3.8-Max +14.0 · DeepSeek V4 Pro (0813) +16.1 · GLM-5.3-Flash +17.2 · MiniMax M3 +18.1 · Qwen3.8-Flash-Next +19.1 · Mistral Large 4 +22.1 · Beam 1.0 +22.4 · Nemotron 3.5 Super VL +26.4 · 4 previous: Claude Opus 4.8 +8.4 · Gemini 3.7 Flash +9.2 · Grok 4.6 +14.2 · GLM-5.2 +16.0
Unverified
1 model — vendor-reported on one side
Setup-dependent
21 models — scored on a different harness or effort tier

Terminal-Bench 4.0 · high Terminal ops

±12.4 is noise
Ahead

AnalystAgent Spreadsheet & document analysis

±11.2 is noise
Behind
1 previous: Gemini 3.7 Flash −11.3
Ahead
MiMo-V2.6-Pro +11.3 · Step 5 Preview +13.8 · MiniMax M3 +38.8
Tie
5 models within ±11.2
Setup-dependent
18 models — scored on a different harness or effort tier

No verdict for Gemini 4 Argon anywhere on DeepSWE, OSWorld 2.0 · v2026 08 08 offline ( nothing independently confirmed on both sides).

  • None of Gemini 4 Argon’s coding comparisons are independently confirmed on both sides yet.

Gemini 4 Argon API pricing

$4.00 in / $20.00 out per 1M tokens — official pricing source

What Gemini 4 Argon costs per job

Gemini 4 Argon cost for three reference workloads, computed from its list rates
WorkloadTokens in / outCost
One long chat turn100K / 10K$0.600
A codebase review1,000K / 100K$6.00
A day of agent work10,000K / 1,000K$60.00

Computed from Gemini 4 Argon’s list rates above — cache discounts, its introductory rate through 2026-11-30, and batch tiers are not applied.

Gemini 4 Argon is one of 7 Google DeepMind models tracked on this site, at these official list prices.

Google DeepMind model pricing, official list rates
ModelIn / 1MOut / 1M
Gemini 4 Hydrogen$1.50$7.50
Gemini Decision 1$0.10$0.00
Gemini 4 Carbon$4.00$20.00
Gemini 4 Argon$4.00$20.00
Gemini 3.8 Flash$0.75$3.75
Gemini 3.7 Flash(previous version)$0.75$3.75
Gemini 3.1 Pro Preview$2.00$12.00

Gemini 4 Argon benchmark scores

Gemini 4 Argon benchmark scores, provenance, and source links
BenchmarkScore
HLE(no tools)[1]
Reasoning · ±2 is noise
Terminal-Bench 4.0(high)[2]
Terminal ops · ±12.4 is noise
OSWorld 2.0(v2026 08_08_offline)[3]
Long-horizon computer use · ±9.7 is noise
AnalystAgent[4]
Spreadsheet & document analysis · ±11.2 is noise
DeepSWE[5]
Long-horizon coding · ±9.5 is noise

Who ran these numbers: 3 of 5 independent — artificialanalysis.ai (2), vals.ai (1); vendor self-reported (2).

  1. HLE: At AA's high effort tier, the only tier AA publishes for it. Read 2026-10-02 from the model page: Terminal-Bench 2.1, GPQA and LiveCodeBench were null; Terminal-Bench 4.0 read 57.07, which this site does not track. AA-AnalystAgent was null on that 2026-10-02 read; a later row on the same page is 48.75 (recorded separately, checked 2026-10-15, still 0.4875 on 2026-10-19).
  2. Terminal-Bench 4.0: vals.ai board row — mini-swe-agent harness (single bash tool), pass@1 averaged over 3 full passes, raw value 57.576 at effort high. Artificial Analysis same-conditions run (High): 57.07. Google's model card self-computes 57.4 on 'Terminal-bench 4.0' at the card's highest thinking setting — not stored as a row; this cell holds the independent run.
  3. OSWorld 2.0: Google DeepMind's benchmark table for Gemini 4 Argon (model announced 2026-09-30; table row "OSWorld-2.0 — Offline subset, Partial score"; table checked 2026-10-08). Google's evals methodology page discloses the conditions: self-computed, maxed over 3 runs with a single attempt per run — a best-of-3 selection on the test set; official evaluator with the maintainers' 08.08 patch; 1080p, 500 steps; Gemini CUA harness with parallel batch tool calling plus a compaction procedure copied from the Opus 5.5 system card. The same table quotes Astra at 72.6 from OpenAI's blog. No effort tier applies (best-of-runs). Safety-flagged responses return empty strings but the episode continues. Verified via reader-rendered table and methodology page.
  4. AnalystAgent: AA's own run, board row 'Gemini 4 Argon (High)': analystAgent 0.4875 = 48.75 pass^5. pass@1 is 57.75. Read from the Gemini 4 Argon (High) page on 2026-10-15.
  5. DeepSWE: Google's launch blog states verbatim: 'It sets a new state of the art on DeepSWE v1.1 (77.9%)'; the model card's methodology PDF says the number is self-computed with a mini-swe-agent harness at the card's highest thinking setting (Argon's tier scale is low/medium/high). Cited to the blog because the deepmind.google card page now renders client-side.

Notes on the record

Google's Gemini API model page for gemini-4-argon states the model is now available (checked 2026-10-19). The API identifier is gemini-4-argon. It is not on the Free tier. Context window is vendor-stated at 1,048,576 tokens (unified input and output limit); default max output is 65,536, with long decode up to 1,048,576. That replaces the earlier note that only Artificial Analysis carried a 1M context figure.

Paid pricing on ai.google.dev/gemini-api/docs/pricing (checked 2026-10-19): input $2.00 and output $10.00 per million tokens through 2026-11-30, then $4.00 / $20.00 from 2026-12-01. Context caching is $0.10 then $0.20 on the same schedule; implicit caching only. This page's headline price stays the post-intro $4/$20 rate so the column matches other models' undiscounted basis; the introductory $2/$10 is the price_alt.

Independent scores on the sheet are unchanged: Artificial Analysis HLE without tools 57.09 and AA-AnalystAgent 48.75 pass^5 at high effort, and vals.ai Terminal-Bench 4.0 at 57.58 (high). Artificial Analysis also embeds Terminal-Bench 4.0 at 57.07 on the same high-effort object; this site keeps the vals.ai row as the recorded independent TB 4.0 figure and does not mix harnesses in one cell. The OSWorld self-report stays as previously sourced.

The model card also reports Agents' Last Exam at 39.5 binary pass rate via the ALE-Claw harness in a 5-hour window, split unstated — a different protocol from the tracked Overall pass-rate column, so it is not recorded as a row.

Compare with

FAQ

Can you buy Gemini 4 Argon yet?

Yes. Google's Gemini API docs list the endpoint gemini-4-argon as available as of the 2026-10-19 check. It is excluded from the Free tier. OpenRouter also lists google/gemini-4-argon.

What does Gemini 4 Argon cost?

Through 2026-11-30 the paid tier is $2.00 input and $10.00 output per million tokens, including thinking tokens. From 2026-12-01 those rates become $4.00 / $20.00. Cached input is $0.10 then $0.20 on the same dates. Figures are from Google's pricing page, checked 2026-10-19.

What is Gemini 4 Argon's context window?

1,048,576 tokens, stated on Google's own gemini-4-argon model page as a unified input and output limit (checked 2026-10-19). Default max output is 65,536 tokens; long decode can use up to the full 1,048,576.

Which Gemini 4 Argon scores are independent?

Three: Artificial Analysis HLE without tools at 57.09, AA-AnalystAgent at 48.75 pass^5, and vals.ai Terminal-Bench 4.0 at 57.58. OSWorld remains Google's own figure. No coding-board independent row is on this page yet beyond that Terminal-Bench run.

Is Gemini 4 Argon open source?

No. Google publishes no weights. Access is through the Gemini API and listed partners under Google's commercial terms.

How does Gemini 4 Argon compare to Muse Wonder on price?

After the introductory window ends, both list $4/$20 per million tokens. Argon's intro rate of $2/$10 runs through 2026-11-30; Muse Wonder's Meta card, checked earlier, showed a single $4/$20 rate with no dated intro step.

Further reading

Benchmark guides