Gemini 4 Argon: $2/$10 Through Nov 30, Then It Doubles

Google finally shipped its next flagship. On October 13, 2026, the company made Gemini 4 Argon generally available — the first Gemini model above the Flash tier in more than seven months, according to lineage tracking by Artificial Analysis (via Tech Times). It arrives with a deliberate price hook: $2 per million input tokens and $10 per million output tokens through November 30, 2026. After that, the price doubles.

That deadline is the real story of this launch. Reuters and VentureBeat report the introductory window ends November 30; after that, Argon costs exactly what Claude Opus 5.5 costs today. This article lays out what the model is, where you can actually use it, what it really costs, and where independent testing backs up — or undercuts — Google's claims.

Quick answer

• What it is: Google's first model in the Gemini 4 family, built for long-horizon "deep work" — coding, finance, legal, cybersecurity and science.
• When: Announced September 30, 2026; generally available since October 13, 2026.
• Price: $2 per million input tokens / $10 per million output tokens through November 30; then $4/$20 (reported by Reuters and VentureBeat).
• Who gets it: Paid Gemini API developers, Google AI Pro and Ultra subscribers in the Gemini app and AI Mode in Search, and Ultra subscribers in Antigravity. No free-tier access.
• Independent verdict: 53 on Artificial Analysis's Intelligence Index — level with GPT-6 Astra, five points behind Claude Opus 5.5, but first place on business-workflow benchmarks.

What Gemini 4 Argon actually is

Google's positioning word for Argon is "deep work": sustained, multi-step professional tasks rather than quick chat replies. The official launch post describes a model that maintains focus across long, complex workflows — code migrations, large codebase refactors, legal and financial analysis, and cybersecurity defense. It is the first model in what Google says will be the Gemini 4 family.

The technical headline is context. Argon offers a 1 million token unified input-output limit through the Gemini API — a single pool shared across input, internal reasoning and output — which Google describes as industry-leading. It also supports long decode (up to 1 million output tokens) and tunable thinking levels (low, medium, high), so developers can trade reasoning depth against latency and cost. On modalities, Argon accepts text, images, video and audio as input but outputs text only.

The launch itself was deliberately staged. On September 30, Argon went only to trusted cyber defenders through Google's Fairwind Program, and Google said it was participating in the U.S. government's voluntary pre-release access process. Two weeks later came the general release. Google says thousands of its own engineers were already using it — though Bloomberg reported some internal testers doubted the benchmark strength would transfer to real work, a claim Google disputed.

Where you can actually use it (and where you can't)

Two things stand out. First, Argon skipped the free tier entirely — Google is treating it as a premium-only flagship from day one. Second, the audience is enormous: Google said in August that the Gemini app had passed 1 billion monthly users, which the company described as its fastest-growing product ever (PYMNTS).

The price — and the December catch

Here is the rate card as listed on Google's pricing page (via Tech AI Wire's read of the Gemini API pricing page):

The fine print matters more than the headline:

• Output prices include thinking tokens — Argon is a heavy thinker. Artificial Analysis measured an average of 62,000 output tokens per index task versus 27,000 for GPT-6 Astra. So the per-token rate advantage is partly offset by how many tokens the model consumes per job — the advantage comes from cheaper rates, not from efficiency (N884 AI Brief).
• What tokens buy. A million input tokens is roughly 750,000 English words — several thousand pages of text. But Argon's headline use case is long reasoning runs, where output tokens (billed at 5x the input rate) are where the bill lands.
• Cached input at 95% off is the pressure valve. Repeat context — retrieved documents, system prompts, long histories — bills at $0.10 per million tokens during the intro window. Workflows built around caching will feel the December doubling much less.
• After December 1, Argon costs exactly what Claude Opus 5.5 costs: $4/$20 (DataNorth). The intro pricing is a land-grab — Google is inviting teams to benchmark and integrate now, betting switching costs will keep them once rates double.
• The cheaper sibling stays cheaper. Gemini 3.8 Flash is listed at $0.75 input / $3.75 output (Tech AI Wire) — and free-tier users stay on Flash-class models regardless.

The alternative is running your own weights — open-weight models like Meta's 816B Muse Spark Open, which we covered when it launched earlier this month, trade per-token billing for infrastructure you host yourself.

Benchmarks: what Google says vs what independent tests measure

Google's own table (vendor-run, Google-chosen tests) shows Argon leading 12 of 19 rows, per VentureBeat's count: 77.9% on DeepSWE v1.1 (claimed state of the art for long-horizon software engineering), 68.9% on the Vals Index (ranked #1), 68.0% on CWE-bench, and a 0.7% prompt-injection success rate versus 8.5% for GPT-6 Astra. The same table has gaps: Astra wins FrontierSWE v2 by over 10 points, and Opus 5.5 wins Terminal-Bench 4.0 — which model leads on coding depends on which test you pick.

Independent measurement tells a less clean story. Artificial Analysis — which ran Argon at its High reasoning setting against most rivals at their maximum — puts the model in the top tier but not at the top:

Four honesty notes before you draw conclusions:

• Those benchmark scores reflect High mode, not necessarily the subscriber default in the Gemini app (Tech Times). What you get in the consumer app may differ from what the leaderboards show.
• The 15% hallucination rate is the genuine standout — Argon says "I don't know" far more honestly than GPT-6 Astra. But the 50% accuracy figure on the same test shows it is more honest about ignorance, not more knowledgeable (Emergent.sh, The Decoder via NeuralTrust).
• Settings aren't identical across vendors: AA tested Argon at High against most rivals at max, and each vendor ran its own harness. Treat cross-vendor tables as directional, not precise.
• Hands-on tests are mixed. One DevOps-focused test (Bird and Bear) had Argon write 12 Kubernetes manifests that passed linting but crash-looped on a real k3s cluster, while rivals got it right first time — Argon fixed it once shown the error, and was cheaper and faster doing so. Benchmarks measure the ceiling; production measures the floor.

Pros and cons

Pros:

• Introductory pricing ($2/$10) plus 95%-off cached input is genuinely cheap for frontier-tier work — through November 30.
• Strong independent business scores: 1st on the Vals Index and AutomationBench-AA, with standout finance, legal and cyber sub-scores.
• The lowest hallucination rate among top-tier models — valuable for internal tools that summarize policies or answer customer questions.
• 1M unified context window, tunable thinking, and broad availability: API, app, Antigravity, Copilot, Vertex.

Cons:

• The price doubles on December 1 to match Claude Opus 5.5 — budget at $4/$20, not $2/$10.
• Heavy output-token usage means real bills will run higher than the per-token rates suggest.
• No free tier and no open weights — this is a closed, premium-only model.
• Behind Sonnet 5.5 and Opus 5.5 on agentic coding benchmarks, and High-mode benchmark scores may not match the subscriber default.

The practical take: test now, budget for December

The right way to read this launch is as a five-week evaluation window. Run real workloads through Argon before November 30, measure output-token consumption per task (not just per-token rates), and build caching into your retrieval patterns — that's where the 95% discount makes the December doubling survivable.

FAQ

Can I try Gemini 4 Argon for free?

No. Argon is available only through the paid Gemini API and AI Studio, and to Google AI Pro/Ultra subscribers in the Gemini app and AI Mode. There is no free-tier access on any surface.

What exactly happens on December 1?

According to Reuters and VentureBeat, the introductory $2/$10 rate ends November 30 and standard $4/$20 pricing applies. Google's own posts confirm the $2/$10 price but don't name the end date — plan for December 1 and you're covered either way.

Is Gemini 4 Argon better than Claude Opus 5.5?

It is strong on business benchmarks (Vals Index #1, AutomationBench-AA #1) and hallucinates far less — but Opus 5.5 leads the AA index 58 to 53. Independent testing paints a strong, uneven model, not a clear winner.

What's the model ID and context length?

The API model ID is gemini-4-argon, with a 1 million token unified input-output limit and up to 1 million output tokens.

Who is already using Argon?

Google names Deloitte, LexisNexis, Ambience Healthcare and Isomorphic Labs as early enterprise testers (via reporting). Their individual results aren't public — the named-partner list is Google's, not an independent verification.

Key takeaways

• Gemini 4 Argon went generally available October 13 after a September 30 Fairwind-only announcement — Google's first flagship-tier Gemini in over seven months.
• It costs $2/$10 through November 30, then doubles to $4/$20 — the same rate as Claude Opus 5.5. Budget at the higher number.
• Independent testing scores it 53 on Artificial Analysis's index: level with GPT-6 Astra, behind Claude Opus 5.5 and Sonnet 5.5.
• It leads business-workflow benchmarks (Vals Index #1, AutomationBench-AA #1) and has the lowest hallucination rate among top-tier models.
• No free tier, no open weights, and heavy output-token usage means real-world bills will exceed what the per-token rates imply.

As of October 23, 2026, the cheap window has about five weeks left. Whether Argon keeps its users after December 1 will answer the question Bloomberg's sources raised at launch: does the benchmark leader actually do the work?

Sources: Google blog, Introducing Gemini 4 Argon (Sep 30, 2026) and Gemini 4 Argon is now available for everyone (Oct 13, 2026); Google Cloud blog: Gemini 4 Argon available on Google Cloud; pricing via Tech AI Wire's read of the Gemini API pricing page (dev.to); launch-price end date reported by Reuters and VentureBeat via DataNorth; independent benchmarks via Emergent.sh, Tech Times, N884 AI Brief, NeuralTrust, Vanda Team, Bird and Bear; internal skepticism reported by Bloomberg via The Latent; access and app rollout via Lindy, Bushletter, PYMNTS.