SPIRITT×
Gemini 4 ArgonComplex codebases. Deep research. Multi-step business work. Google’s new frontier model brings up to 1M output tokens to longer problems. Explore the early results, introductory pricing and phased rollout.
Google’s announcement, independent evaluations, visual benchmark examples and early reactions. Access is still limited; these are launch and evaluation posts, not a gallery of public hands-on demos.
Introducing Gemini 4 Argon – our new frontier model. It’s built for complex workflows across coding, enterprise knowledge work, and cybersecurity defense – rolling out today to a set of trusted testers through our Fairwind Program.
— Google DeepMind (@GoogleDeepMind) 2026-09-30T20:03:30Z
Lots of discussion out there about our next model(!), so I wanted to give an early look as soon as possible. Introducing Gemini 4 Argon! It shows frontier performance in complex workflows, cyber defense and software engineering. Teams are using it extensively at Google, from coding to quantum computing, great feedback. Here’s a look at the benchmarks:
— Sundar Pichai (@sundarpichai) 2026-09-30T20:02:58Z
Gemini is on top of the Vals Index for the first time. Gemini 4 Argon takes the #1 spot on the Vals Index at 68.9%.
— Vals AI (@ValsAI) 2026-09-30T20:04:56.000Z
With high reasoning (the highest available), it scores 53 on the Artificial Analysis Intelligence Index, matching GPT-6 Astra (max, 53) and 1 point ahead of GPT-6.1 Sol (max, 52), with gains driven by lower hallucinations and stronger agentic capabilities. […] At its current 50% pricing discount and with cache discounts increased to 95%, Gemini 4 Argon costs $1.99 per Intelligence Index task, 60% of GPT-6 Astra (max), but 2.7x GPT-6.1 Sol (max). After the discount ends, this will rise to $3.98 (~1.2x GPT-6 Astra (max)). Gemini 4 Argon is currently being rolled out to selected users and is not publicly available. The 50% discount is an initial promotion. Google has not yet confirmed the promotion end date […] We tested Gemini 4 Argon with Long Decode Continuation, a new Gemini API feature that pauses long responses and resumes them across follow-up calls.
— Artificial Analysis (@ArtificialAnlys) 2026-09-30T20:21:32Z
Big news: Gemini 4 Argon (High) by @GoogleDeepMind just landed #1 in Text Arena with 1525 pts, and #8 in Code Arena: WebDev with 1679 pts!
— Arena.ai (@arena) 2026-09-30T20:30:24Z
Say hello to Gemini 4 Argon 👋 @GoogleDeepMind’s new model just became the first to clear 50% on AutomationBench Use it when your workflow crosses a bunch of apps: ops routing & approvals, marketing campaigns, sales pipeline follow-through
— Zapier (@zapier) 2026-09-30T20:11:18Z
It ranks #1 on Blueprint-Bench 2, a benchmark where AI agents draw floorplans from photographs of apartment interiors.
— Andon Labs (@andonlabs) 2026-09-30T20:06:33Z
Gemini 4 Argon is in the top 5 on 20 of the 22 benchmarks we ran. It is particularly strong on finance, legal, coding, and security. It is also a major step up from its predecessor, Gemini 3.8 Flash, which released earlier this month. It improved on every benchmark we ran.
— Vals AI (@ValsAI) 2026-09-30T20:04:57.000Z
Introducing Gemini 4 Argon, our new frontier model, rolling out to cyber defenders starting today, and more widely as soon as possible. I am really excited by the progress we have made here. Argon is priced at $2 in and $10 out during introductory pricing!
— Logan Kilpatrick (@OfficialLoganK) 2026-09-30T20:03:23Z
It keeps happening. AIs start to lie and cheat once they get good at making money. Gemini 4 Argon is #3 on Vending Bench 2, a huge leap for Google. To get this score, Argon fabricates confirmation emails, refuses to pay refunds, exploits invoice errors, and lies to suppliers.
— Andon Labs (@andonlabs) 2026-09-30T20:16:36Z
Today we’re introducing Gemini 4 Argon. It delivers frontier performance in complex workflows across real-world software engineering, knowledge work, and cybersecurity defense with an industry-leading 1M token output limit.
— Google (@Google) 2026-09-30T20:03:44Z
Google says Argon agents are helping migrate C/C++ to Rust, including an 800K+ line kernel. The same announcement says rigorous manual review before production. I'd like to see the reviewer-hours chart next to the coding benchmarks.
— Vraj Talati (@VrajTalati1) 2026-09-30T21:14:58Z
Google just dropped Gemini 4 Argon. New frontier model with a 1M token output limit. Limited rollout for now — first to trusted cybersecurity teams. […] Not first everywhere: FrontierSWE and Terminal-Bench Science go to Astra, Terminal-bench 4.0 and PostTrainBench to Opus 5.5. The model looks built for long agentic workflows in knowledge work, coding, and defense — not just chat. Curious to see it when access opens up.
— Yurash (@Yurash1104) 2026-09-30T20:55:34Z
Argon leads the Vals Index for professional work and pushes long-horizon coding forward. The broader benchmark picture shows both strengths and trade-offs.
Google announced Gemini 4 Argon on September 30, 2026, targeting software engineering, legal and financial work, multimodal research, and cybersecurity defense. Access starts with trusted testers and cyber defenders through the Fairwind Program. Wider access is planned, not generally available at announcement.
Google is expanding the output limit from the previous 64K to up to one million tokens. Artificial Analysis tested Argon at high reasoning with Long Decode Continuation enabled. A larger output budget creates room for extended reasoning and generation; it is not a speed or low-cost guarantee.
Argon scores 68.9% on Vals Index v2.1, leading its professional-work evaluation. On Artificial Analysis’s broader Intelligence Index v4.3.2, it scores 53, matching the rounded score of GPT-6 Astra at max effort while trailing the leading Claude configurations.
Introductory API pricing will be $2 per million input tokens and $10 per million output tokens. Google says $4/$20 rates apply after the introductory period, without announcing its end date. Cached input is discounted by 95%. These are Google API prices, not SPIRITT subscription prices.
Inside Google, Argon has helped with large code migrations, memory optimization and specialized research. Google reports memory optimizations that would free over 300 TiB once rolled out, and a 2.7× speedup versus an existing Rust video-decoder port. Critical code rewrites are still subject to extensive auditing and testing; these are specific internal examples, not universal productivity promises.
Artificial Analysis, September 30, 2026, Index v4.3.2. Selected configurations; rounded index points, not percentages. Argon uses high reasoning with Long Decode Continuation, up to 1M output tokens. Claude fallback and effort settings are shown. Matching rounded scores do not establish equal capability on every task.
Professional work, coding, science, long context and multimodal understanding, with independent evaluations kept separate from Google’s compiled results. Compare the setup as well as the score.
Vals AI’s September 30 launch snapshot. GDP-weighted finance, coding, legal and tax tasks. Argon: high reasoning, temperature 1, 262K output budget. Comparator effort settings are not specified in the captured chart. This is a different evaluation from the AA Intelligence Index.
Google’s comparison, citing Zapier’s private-set leaderboard. These scores are separate from AutomationBench-AA’s objective-completion metric. Model settings and safeguards follow the cited runs, not a uniform rerun.
Google’s comparison, citing Vals AI. Tests financial research that requires multiple steps and reading source documents. The displayed model settings follow Google’s evaluation report.
Google’s comparison, citing Vals AI. Argon leads the listed systems, but a 19.6% score still leaves substantial room to improve. Relative leadership is not reliable completion of every legal task.
Google computes Argon with a mini-swe-agent harness; Astra comes from the public leaderboard, and Claude values from system cards. Highest-scoring reported thinking levels are selected. These are compiled results, not one shared rerun.
Google’s comparison, citing Proximal’s leaderboard. Argon trails all three displayed peers here. Strength on DeepSWE does not establish a lead on every coding evaluation.
Google’s comparison, citing Vals AI’s leaderboard. All four listed systems score highly; the small point-estimate gaps do not establish a decisive advantage for every app-building task.
Argon is Google’s run; other scores come from the benchmark leaderboard at each model’s highest-scoring reported thinking setting. Argon trails the listed peers. Keep this separate from AA’s mini-swe-agent results.
Google runs all models with OpenCode, ten hours and one NVIDIA H100 per attempt. Weighted aggregate across four base models and seven benchmarks. Argon trails Opus 5.5 in this setup.
Google’s Argon run uses a 6× verifier timeout to address verification timeouts; other scores come from the public leaderboard. Argon trails Astra and Opus. The timeout difference matters when comparing results.
Google evaluates all four models with a Linux terminal, bioinformatics tools, Python, R and internet access. These are tool-assisted results, not closed-book biological knowledge scores.
Google’s comparison, citing Surge’s public leaderboard. Argon has the highest displayed score in this snapshot; general reasoning ability is broader than one mathematics evaluation.
Google-run breadth-first-search F1 on 650 items with contexts up to 128K tokens. The 99.7% and 98.7% scores are close. F1 is not a whole-task completion rate.
Google-run breadth-first-search F1 on 200 problems with contexts from 256K to 1M tokens. This evaluates graph reasoning at those lengths; it is not a complete specification of the future public API context window.
Google’s Argon run: default ALE-Claw, five-hour window, safeguards enabled, binary task pass rate. Astra and Opus come from the public leaderboard. Fable has no reported comparable value and is omitted.
Google reports the best of three Argon runs, not an ordinary average: 1080p screenshots, 500-step limit, compaction, parallel tools and safeguards. Astra is from its launch report. Partial credit on the offline subset is not full-task success; incompatible Claude combined-subset scores are omitted.
Google’s comparison, citing Surge’s leaderboard. No tools. Argon and Astra are close on this chart-understanding evaluation; the gap alone is not evidence of a decisive lead.
Google evaluates all four models without tools. Gemini uses 1 frame/second; Astra, Fable and Opus are capped at 800, 300 and 600 frames respectively because of API limits. Unequal sampling limits direct comparability.
Selected leading models from the leaderboard reproduced by Google. Argon, Grok 4.7 and Astra tie at 68% pass@1; pass@4 breaks leaderboard ties. Different agent harnesses are shown, so this is not a unique first-place claim.
Google’s private dataset measures recall of recent confirmed vulnerabilities across popular open-source projects in 20 languages. Models receive source code through an internal Antigravity harness. These published scores are not an independently reproduced public benchmark.
Google reports Wiz’s private benchmark of real web vulnerabilities without source-code access. Published internal-evaluation scores, not a general vulnerability-detection guarantee or an independently reproduced public benchmark.
Google reports Gray Swan’s attack-success rates at 15 attempts; lower is better. Selected peers shown. Per-model effort and harness details are not specified here. Low measured success does not mean immunity to prompt injection.
Artificial Analysis: dataset 1.0.6, 657 private tasks, 50-turn cap. Average share of guardrail-adjusted objectives completed, with errors or guardrail violations scoring zero. Different metric and setup from Zapier’s 51.3% result; not a whole-task pass rate.
Artificial Analysis’s rounded launch scores: mini-swe-agent, 66 tasks, three repeats, 500-step cap, no compaction. Claude configurations use the evaluator’s default fallbacks. These scores are separate from Google’s compiled results and Vals’ 57.6% run.
Artificial Analysis’s task-cost estimate, not cost per successful completion. Argon high uses about 62K output tokens per task versus Astra max at 27K. The introductory advantage comes from pricing; after the promotion, Argon costs more in this comparison. These are not Vals task costs.
Artificial Analysis’s benchmark-specific hallucination metric, lower is better. Argon also answers fewer questions correctly: see the adjacent accuracy comparison. More abstention can reduce wrong guesses without adding knowledge; this is not a real-world error rate.
Artificial Analysis’s rounded share of questions answered correctly. Argon’s lower hallucination metric does not translate to higher accuracy here. Its overall Omniscience score is 42 versus Astra’s 43.
Google’s announced USD API rates. The promotion end date is not specified, and general access is still rolling out. Cached input is 95% off the applicable input rate. Ancillary charges and product-specific terms are not fully specified; these are not SPIRITT plan prices.
Google’s announced USD API rates. The promotion end date is not specified, and general access is still rolling out. Cached input is 95% off the applicable input rate. Ancillary charges and product-specific terms are not fully specified; these are not SPIRITT plan prices.
Sources: Google’s September 30, 2026 announcement and evaluation methodology; Artificial Analysis Intelligence Index v4.3.2 and launch results; Vals Index v2.1. Google’s comparison table combines its own runs with evaluator and provider-reported results. Versions, reasoning settings, tools, safeguards and scoring differ. Missing scores are omitted, not plotted as zero. Close point estimates do not establish a meaningful advantage. API token prices, evaluation task costs and SPIRITT subscription prices are different things.
Bring the project into SPIRITT, start with your available models, and assess new releases against work that matters to you.
A codebase to improve, research to turn into a decision, or a workflow to automate. Give SPIRITT the files, context and success criteria so the work starts with a clear destination.
Google is starting with trusted testers before expanding access. Check the official rollout and the models or connections enabled in your workspace. This release brief does not mean Argon is already available in SPIRITT.
Start with the models available to you today. When evaluating a new model, compare the quality of the finished work, total cost, time and human review needed, not just a leaderboard score.
Give SPIRITT your goal, files and context. Build with the models available in your workspace today, and evaluate new releases against real work as access becomes available.