Google made Gemini 4 Argon generally available on October 13, 2026 at an introductory $2 / $10 per million tokens, with cached input at $0.10. Every line of that deck doubles on December 1: $4 / $20 standard, $0.20 cached. The same release changes the request contract — sampling parameters are ignored, thinking_budget returns a 400, explicit context caching is gone — and from November 13 a long completion that times out is still billed.
per 1M tokens · through Nov 30, 2026
same model, same meters · 2× on every line
input + thinking + output in one limit
USD per million tokens, from Google’s Gemini API pricing page observed October 19, 2026. Output includes thinking tokens. The December column is price / cached input / output.
| Tier | Input (to Nov 30) | Output (to Nov 30) | Cached input | From Dec 1 |
|---|---|---|---|---|
| Standard | $2.00 | $10.00 | $0.10 | $4.00 / $20.00 / $0.20 |
| Batch | $1.00 | $5.00 | $0.05 | $2.00 / $10.00 / $0.10 |
| Flex | $1.00 | $5.00 | $0.05 | $2.00 / $10.00 / $0.10 |
| Priority | $3.60 | $18.00 | $0.18 | $7.20 / $36.00 / $0.36 |
Sources: ai.google.dev · Gemini API pricing (tier ladder and December rates) · ai.google.dev · What's new in Gemini 4 Argon (cached-input line item) · cloud.google.com · Argon available on Google Cloud (introductory rate through Nov 30).
The Gemini API rate is not the whole rate card. Vertex AI prices the multi-region endpoints 10% higher, and runs a credit that the model page never mentions.
| Surface | Region | Input | Output | Cached | Note |
|---|---|---|---|---|---|
| Gemini API | global | $2.00 | $10.00 | $0.10 | No free tier |
| Vertex AI | global | $2.00 | $10.00 | $0.10 | 50% provisioned-throughput credit through Nov 30 |
| Vertex AI | us / eu | $2.20 | $11.00 | $0.11 | 10% regional uplift; $4.40 / $22 from Dec 1 |
The Vertex credit is a billing adjustment, not a discount rate: Google Cloud injects a monthly credit equal to 50% of net eligible Provisioned Throughput spending from October 13 through November 30, 2026, issued on the 7th of the following month and expiring 30 days later, and the credits apply to on-demand Argon usage as well as provisioned-throughput SKUs. It does not apply to Gemini 4 Argon Cyber Permissive, and it has no equivalent on the Gemini API. Because the credit is sized by provisioned-throughput spend, a pay-as-you-go-only deployment earns nothing from it — so a quote built on the published rate card alone is wrong in both directions, depending on how the capacity is bought.
Argon is not available on the Gemini API free tier.
These are silent behaviour changes and hard errors, not deprecation warnings. A request migrated from Gemini 3.8 Flash keeps working — and stops behaving the way it was tuned.
Argon has a unified 1,048,576-token limit that includes input, thinking, function calls and the final answer. A 900K-token document leaves roughly 148K tokens for reasoning and output on the same request, so a workload sized on “1M context + 65K output” does not fit. Google recommends raising max_output_tokens to at least 32,768, because Argon can spend more of the limit on thinking than earlier models.
The Gemini API page puts the implicit-cache minimum at 4,096 tokens and the Vertex model page at 6,144 tokens, with durable caching on by default. Both agree that a cache hit discounts cached input by 95%, and that explicit caches cannot be created. The practical consequence is the same on either number: only a stable prefix that clears the threshold earns the discount, and a rotating context does not.
Migration order from Google: raise max_output_tokens, set background: true above 64K tokens, move client timeouts to at least 15 minutes, and carry state with previous_interaction_id.
A new billing rule turns a failed long-decode request into a paid one.
Argon can generate completions up to one million tokens. From November 13, 2026, a completion longer than 32K tokens is charged for retrievable token usage even if the request ultimately fails because of a client or server timeout. A retry loop around a long decode therefore pays twice: once for the abandoned attempt and once for the retry. Set timeouts and output caps before enabling long decode, not after the first invoice.
What a deployment record needs to carry for this model.
GA on October 13, 2026 in the global, us and eu locations, with data residency, CMEK and VPC-SC on online prediction, batch inference and context caching. Reachable through GenerateContent, the Interactions API and the OpenAI-compatible Chat Completions surface. The published retirement date is October 13, 2027 or later.
Client-side continuation runs only on the v1beta1 GenerateContent API and needs a continuation_token loop; the Google Gen AI SDK and Agent Studio handle it automatically. Background execution runs the loop server-side, which is the simpler option unless you need control over retries, backoff or whether the interaction is stored.
The current ALLM snapshot is cat_2026_09_22, observed September 22, 2026, so Argon is absent by construction — a sync lag, not a missing endpoint. The same snapshot still carries google/gemini-3.7-flash, which Google has deprecated and now auto-routes to gemini-3.8-flash on request.
Two changes in the October 13–19 window that move a route rather than a benchmark.
As of October 15, fast mode on claude-opus-5-5 no longer needs the fast-mode-2026-02-01 beta header or an access request, and requests that still send the header keep working. It delivers up to 2.5× higher output tokens per second at $8 / $40 per million tokens, twice the standard Opus 5.5 rate, on a rate limit separate from standard Opus that returns a 429 with retry-after when exhausted. It is now also available on Claude in Microsoft Foundry for deployments hosted on Azure — not for Anthropic-hosted Foundry deployments. Fast mode on Opus 5 and Opus 4.8 remains a research preview.
On October 15, OpenAI added native Decisions integration in GPT-Live: a live voice session can delegate to /v1/decisions directly, or trigger a decision as input or output transcript deltas arrive, so the application can act while the user is still speaking. Voice sessions remain billed per minute, and the decision is a separate request on /v1/decisions — so one interaction can now span two endpoints and two billing surfaces, neither of which a token-only gateway route covers.
All official provider pages observed October 19, 2026 (Europe/Paris). Prices are USD per million tokens as the provider states them; the Vertex table separates the global endpoint from the us and eu multi-regions.
Related live views: the October cheap-tier reset, the GPT-5.6 Sol price-deck report and models with a published end-of-life date render the current catalog records.
Deployment intelligence, explained
A launch price is a temporary rate attached to a permanent integration. Argon's $2 / $10 applies through November 30, 2026; the deck that runs from December 1 is $4 / $20, so a product margin set on the launch rate halves on a date the provider has already published. Teams that adopt a model inside its introductory window inherit both the request contract and the cliff.
The rate card is also surface-specific. On Vertex AI the global endpoint matches the Gemini API rate, the us and eu multi-regions add 10%, and a 50% provisioned-throughput credit runs through November 30 — a discount that never appears on the model's own price page. Read the surface, the region and the credit window before comparing two quotes for the same model.
We use technologies necessary for the website to operate. With your consent, we use PostHog and Google Analytics to measure pseudonymised audience data. Refusal does not affect access to the website.