Gemini 4 Argon: Architecture, Benchmarks, and Real-World Developer Review
Status as of October 1, 2026. Google announced Gemini 4 Argon on September 30, 2026. The model is real, but broad developer access is not yet generally available. Initial access is rolling out to trusted cyber defenders through Google's Fairwind Program, and Google says paid API customers and Google AI Ultra subscribers will be reached first. Independent Rewarx hands-on testing, customer production results, and full model-card details will be added to this article as access stabilizes.
Quick Answer
Gemini 4 Argon is Google's newest flagship model, announced September 30, 2026. It is positioned as Google's most resilient model yet against indirect prompt injection, and Google reports it leads on Gray Swan's IPI benchmark. Argon raises the announced per-request output ceiling to one million tokens, far above prior Gemini generations. As of October 1, 2026 the model is in a controlled rollout, not a public release. Treat every benchmark number in this article as Google-published, source-reported, and subject to harness caveats, not as an independent Rewarx test result.
Confirmed Facts
- Announced September 30, 2026 on the Google blog and the DeepMind model page.
- Initial access is through Google's Fairwind Program for trusted cyber defenders; Google states broader developer, enterprise, and consumer access will follow.
- Introductory API pricing announced at $2 per million input tokens and $10 per million output tokens, stepping up to $4 input and $20 output after the introductory period. Cached input is announced at 95% off the applicable input price. Source.
- Announced per-request output limit is 1,000,000 tokens, up from 64,000 in earlier production Gemini endpoints.
- Reported scores from the evaluation methodology: DeepSWE v1.1 77.9%, FrontierSWE v2 55.0%, Vibe Code Bench 91.9%, Terminal-bench 4.0 57.4%, AutomationBench 51.3%, Vals Index 68.9%, LVBench 91.7%, CWE-bench v1 68.0%, and GraphWalks F1 84.2% on 256K-to-1M-token inputs.
- Argon is described by Google as its most resilient model against indirect prompt injection and as a leader on Gray Swan's IPI benchmark.
Unknowns and Access Limits
- No SWE-bench Verified score has been published for Argon. DeepSWE and FrontierSWE are not substitutes and should not be quoted as SWE-bench results.
- No public batch-mode price for Argon has appeared in the launch post. Refer to Gemini batch documentation for general capability and check current pricing for any post-launch update.
- No architecture disclosure: parameter count, expert count, active-parameter count, tokenizer details, and training-data composition are not in the launch material.
- No public quotas, latency targets, throughput numbers, or 50,000-request scalability claims are stated for Argon. See rate-limit documentation only for non-Argon tiers.
- No fine-tuning support, knowledge-cutoff date, exact public model ID, or production region list for Argon has been published.
- The 0.7% indirect-prompt-injection attack-success figure sometimes associated with Argon is not stated in the official launch material. Treat it as unverified until Google or Gray Swan publishes it.
Architecture and Training Approach
Google has not disclosed whether Argon is a dense transformer, a Mixture-of-Experts model, or a hybrid. Active-parameter counts, expert routing, tokenizer changes, pre-training corpus composition, and post-training mixture weights remain unknown. Treat any third-party "Argon architecture" claim that is not on the DeepMind Gemini page as speculation. For developer purposes, plan against what you can observe: token throughput, latency distributions, function-calling reliability via the function-calling guide, and structured-output schema adherence. None of these is a substitute for a public model card, but they are what an engineer can actually measure on day one.
Reported Benchmark Numbers and Harness Caveats
Every figure below is Google-published or source-reported. Each entry notes the original methodology and its limitations.
| Benchmark | Reported score | What it measures | Key caveat |
|---|---|---|---|
| DeepSWE v1.1 | 77.9% | Software engineering tasks | Custom harness; not SWE-bench |
| FrontierSWE v2 | 55.0% | Frontier software tasks | Vendor-curated; not comparable to public SWE-bench Verified |
| Vibe Code Bench | 91.9% | Natural-language-to-code | Newer benchmark, limited public replication |
| Terminal-bench 4.0 | 57.4% | Shell and terminal reasoning | Tool access and sandbox depend on harness |
| AutomationBench | 51.3% | Multi-step automation | Near chance; treat as directional |
| Vals Index | 68.9% | Long-form reasoning | Source-defined rubric |
| LVBench | 91.7% | Long-video understanding | Video ingest pipeline differs from text |
| CWE-bench v1 | 68.0% | Vulnerability remediation | Synthetic issues, not live CVE triage |
| GraphWalks F1 | 84.2% | Multi-hop graph reasoning | Evaluated on 256K-1M input contexts |
Harness parity is the real story. Two vendors reporting 77% and 75% on "coding" may be running different scaffolds, different thinking budgets, and different test-set slices. Until Google's methodology and independent reproductions agree, prefer pass@1 with a fixed tool environment, fixed prompt template, and a frozen evaluation harness.
Long-Context Output Limit vs Input Context
Argon's headline number is a 1,000,000-token output limit per response, up from 64,000. That is an output ceiling, not an input context window. The official evaluation does report GraphWalks tests over 256K-to-1M-token input contexts, which is evidence of long-input behavior, but a public API model card with the complete input-limit specification is not yet published. When you read the long-context guide and context-caching docs, treat the documented limits as the prior Gemini generation's defaults until Argon's model card replaces them.
Pricing Math and Cost Modeling
Use the announced numbers exactly:
- Introductory: $2 per million input, $10 per million output. Cached input at 95% off the applicable input price.
- Post-introductory: $4 per million input, $20 per million output. Same caching formula.
Worked example for a 100,000-token generation reading a 250K-token cached prefix at post-intro rates: input 250K × ($4 × 0.05) = $0.05, output 100K × $20 = $2.00, total ≈ $10.05 per million-token workload at the cache rate, or $10.05 per 350K total tokens. The same workload without caching would be 250K × $4 = $1.00 plus 100K × $20 = $2.00, totaling $3.00 per 350K tokens. Cache or no-cache, plan a 2x–3x safety margin because batch pricing, regional surcharges, and Argon-specific tiers are not yet published. Check current pricing before any commitment.
Indirect Prompt Injection and Defensive Posture
Google states Argon is its most resilient model yet against indirect prompt injection and leads on Gray Swan's IPI benchmark. That is a real and significant claim, but two caveats apply: the exact attack-success percentage is not stated in the launch material, and any production deployment still needs belt-and-suspenders controls. Treat untrusted content as data, segregate tool permissions per the function-calling guide, enforce structured-output validation via the structured-output docs, and require qualified human review for security, legal, financial, medical, compliance, trading, and infrastructure actions. This article does not provide offensive intrusion or evasion instructions.
Production Evaluation Plan
- Confirm access. Verify your account appears in the Fairwind or paid API tier before any benchmark run.
- Pin the model ID and version. Capture the exact string returned by the model directory at evaluation time.
- Freeze the harness. Lock the prompt template, system prompt, tool schema, and temperature to one configuration per claim.
- Run a parity matrix. Reuse the same harness against Claude 4.5 Opus, GPT-class baselines, and any open-weights candidate you can self-host.
- Measure latency, throughput, and tail. Capture p50, p95, p99 TTFT on a representative workload. Do not trust vendor latency claims; measure your own.
- Test long-context behavior separately. Use the long-context guide to construct needle-in-haystack and multi-hop graph probes at 256K and 1M input sizes.
- Test caching economics. Compare cache-on vs cache-off cost for your top-3 prompt templates using the caching guide.
- Red-team prompt injection on real data. Use the Google cyber defense page as a starting point, but bring your own corpus.
- Deploy behind a guardrail. Run output through validators, and require qualified human review for any high-risk action class.
- Document unknowns. Anything you cannot measure, label "unknown" in your rollout notes rather than guessing.
FAQ
Was Gemini 4 Argon actually released?
Argon was announced on September 30, 2026. As of October 1, 2026 it is in a controlled rollout through trusted cyber defenders via the Fairwind Program. Google says broader developer, enterprise, and consumer access will follow for paid API customers and AI Ultra subscribers first.
Does Argon have a 1M token context window?
No, that wording is misleading. Argon's announced per-request output limit is one million tokens, up from 64K. The official evaluation reports GraphWalks at 256K-to-1M input contexts, but a public API model card with the full input-limit specification is not yet available.
What does Argon cost?
Google's announced introductory pricing is $2 per million input and $10 per million output, stepping to $4 and $20 after the introductory period. Cached input is announced at 95% off the applicable input price. No Argon batch price has been published.
What is Argon's SWE-bench Verified score?
Google has not published a SWE-bench Verified score for Argon. DeepSWE v1.1 at 77.9% and FrontierSWE v2 at 55.0% are separate benchmarks with separate harnesses and should not be quoted as SWE-bench results.
Is Argon safer against prompt injection?
Google describes Argon as its most resilient model yet against indirect prompt injection and as a leader on Gray Swan's IPI benchmark. Treat any specific attack-success percentage as unverified until Google or Gray Swan publishes it, and keep qualified human review in the loop for any high-risk action.
Final Verdict
Gemini 4 Argon is a meaningful release on paper: a 1M-token output ceiling, a broad coding-and-agent benchmark suite, and a clear security posture around indirect prompt injection. None of it is independently verified yet. On October 1, 2026, the right move is to freeze your harness, queue your evaluation matrix, and wait for general access rather than declaring a winner from incomparable vendor reports. Rewarx will publish hands-on results, latency distributions, and parity tests once the API tier is available to a normal developer account. Until then, the headline numbers belong to Google, and your conclusions should be earned, not borrowed.
Primary Sources and Verification Links
- Google: Introducing Gemini 4 Argon
- Google DeepMind Gemini 4 Argon
- Gemini 4 Argon evaluation methodology
- Gemini 4 Argon cybersecurity defense
- Gemini API model directory
- Gemini API pricing
- Gemini API rate limits
- Gemini API context caching
- Gemini API long-context guide
- Gemini API function calling
- Gemini API structured output
- Gemini API batch mode
- Google Gen AI SDK
- Vertex AI generative AI documentation
- Google Cloud Run documentation
- Google Cloud compliance resource center
Build Better AI Product Visuals
Rewarx Studio AI helps ecommerce teams create and review production-ready product imagery with brand and SKU fidelity in the workflow.
Explore Rewarx Studio AI