Which Gemini Model to Use in October 2026, Now That Gemini 4 Argon Is Out

Which Gemini model should a team pick in October 2026? For most text work the answer is gemini-3.8-flash, released September 2, 2026. For high-volume sorting and extraction it is gemini-3.5-flash-lite. For the hardest work there is now gemini-4-argon, which Google made generally available on October 13. Gemini 3.1 Pro is still a preview, and Gemini 3.5 Pro has not shipped.

This piece is a plain picker for agencies, CMS developers and small publishers who call the Gemini API for clients. It shows which Gemini model to try first for each job: text, Pro, voice, images and video. After that come the prices from Google’s pricing page and the model IDs that are closing or now answer with a different model. Every price comes from Google’s pricing page as read on October 14, 2026, and every date from Google’s own changelog, model and deprecation pages.

The short version on which Gemini model to use: start text work on 3.8 Flash and measure what it costs per finished task. Move bulk classification to 3.5 Flash-Lite. For hard problems, test Argon before the 3.1 Pro preview, and price the test at Argon’s December 1 rate. Do not wait for 3.5 Pro. If your code still names gemini-3.7-flash or gemini-3.5-flash, a different model answers it now. For voice, use 3.8 Live. For images, start with Nano Banana 2.1, which costs the same as 2 Lite at 1K.

Which Gemini model fits which job

The table sorts common jobs by the model to try first. It is a starting point for your own tests, not a benchmark result.

Gemini 3.8 Flash is the default for text

Google has shipped four Flash generations since May, according to the Gemini API changelog. Gemini 3.5 Flash arrived May 19, 3.6 Flash on July 21, 3.7 Flash on August 13 and 3.8 Flash on September 2. On October 8, Google deprecated 3.5 and 3.7 Flash. Their IDs still work, but requests now go to newer models.

Google’s changelog calls 3.8 Flash "our most intelligent Flash model, engineered for long-horizon software engineering, autonomous agents, and complex enterprise workflows." It takes text, images, video, audio and PDF, and returns text. The context window is 1,048,576 tokens in and 65,536 out. Thinking levels are low, medium and high; the model page says minimal is not supported.

That makes 3.8 Flash the first test for drafting, summarizing and chat. It is not always the cheapest to run. Our look at what Gemini 3.8 Flash costs per task found it produces far more output tokens than 3.7 Flash on the same benchmark, at the same rate card. That cheaper option is gone. Since October 8, requests to gemini-3.7-flash go to gemini-3.8-flash, per the changelog. A team on 3.7 Flash now gets 3.8 Flash output lengths without changing a line of code. Requests to gemini-3.5-flash now go to gemini-3.6-flash. If either ID is in your code, check output length and your bill from October 8 on.

Flash-Lite for volume: pick 3.5, not 3.1

Two Flash-Lite models are stable. gemini-3.5-flash-lite went GA on July 21. Its model page describes it as a low-latency, low-cost model for high-throughput work such as subagent tasks and document parsing. It has the same 1,048,576-token input window as 3.8 Flash, supports thinking, and costs $0.30 in and $2.50 out per million tokens.

gemini-3.1-flash-lite is cheaper, at $0.25 in and $1.50 out. It is also closing. The deprecations page gives it a shutdown date of May 7, 2027, with 3.5 Flash-Lite as the replacement. Building a new pipeline on it now means migrating it within seven months.

Flash-Lite suits jobs with short prompts and short answers: tagging an archive, routing support email, pulling fields from PDFs. These jobs rarely need deep reasoning, and volume makes the rate matter. Run a sample through both Flash-Lite and 3.8 Flash. If the cheaper model gets the same labels, keep it.

Pro: 3.1 Pro is still a preview, and Argon now sits above it

The newest Pro model on the Gemini API is gemini-3.1-pro-preview. Google released it on February 19, 2026, per the changelog. No Pro model has shipped since. The models page now lists it only in its table of all Gemini 3 models. Gemini 4 Argon has taken the top spot on that page.

Google’s model page says 3.1 Pro Preview offers better thinking, improved token efficiency and more grounded, factually consistent answers than the Gemini 3 Pro series it refines. Its context window matches Flash: 1,048,576 tokens in and 65,536 out. It does not support the Live API or image generation.

Input costs $2.00 per million tokens up to 200,000 tokens and $4.00 above that. Output costs $12.00 and $18.00. There is no free tier. Until November 30, stable Argon costs less than this preview on output.

Why the preview label matters for client work

Preview models can close quickly. The deprecations page shows that gemini-3-pro-preview was released November 18, 2025 and shut down March 9, 2026. That is under four months. Its replacement was 3.1 Pro Preview, which has no shutdown date announced.

On which Gemini model to use for hard problems, our view is that most agency work does not need either. If testing shows Flash cannot do the job, test Argon first. It is stable, and 3.1 Pro is not. Keep the model ID in one config setting, so a forced switch is a one-line change.

Gemini 3.5 Pro: what Google has and has not said

As of October 14, Google’s changelog and models page do not list a Gemini 3.5 Pro. Everything else about it comes from press reports.

TechCrunch reported on July 21 that the model, expected in June, had not shipped, citing Bloomberg. On August 13, Forbes reported that the delay was continuing. Google has not published a release date.

For planning, treat 3.5 Pro as unannounced, and leave it out of which Gemini model you pick for a client. If it ships, test it against your current model on your own work before you switch. Launch-day benchmark tables are a weak basis for that decision, for the reasons in our piece on vendor benchmarks and independent leaderboards.

Gemini 4 Argon: generally available since October 13

Google announced Gemini 4 Argon on September 30 for its Fairwind security program. On October 13 the changelog listed gemini-4-argon as generally available, and the models page lists it as Stable. Any paid Gemini API project can call it. Google AI Pro and Ultra subscribers also get it in the Gemini app and AI Mode.

Argon has a 1,048,576-token window shared by input, thinking and output. Normal calls stop at 65,536 output tokens; background execution goes up to about 1M. It costs $2 per million input tokens and $10 output through November 30, then $4 and $20 from December 1. There is no free tier.

Two changes affect old code. Argon ignores temperature, top_p and top_k, and it does not support explicit caching. On which Gemini model to use for a long report or a code migration, Argon is now the first test. For chat and drafting, 3.8 Flash is still far cheaper. Our Gemini 4 Argon GA post has the full details and a test plan. Our earlier explainer on what Gemini 4 Argon is covers the Fairwind release.

Voice: Gemini 3.8 Live or 3.8 Live Extended Thinking

Gemini 3.8 Live and 3.8 Live Extended Thinking went GA on September 15. The launch post says both detect 97 languages and can call tools in the background while the conversation continues.

For voice, which Gemini model to use is a short list. Google’s model page calls gemini-3.8-live the default for most low-latency voice agents. gemini-3.8-live-extended-thinking is for complex, multi-step problems during a live call. The standard model takes text, images, audio and video. Both have a 131,072-token input window.

The pricing page lists one set of rates for both. Audio input costs $3.00 per million tokens, or about $0.005 a minute. Audio output costs $12.00 per million, or about $0.018 a minute. Start a member-services line on 3.8 Live. Move to Extended Thinking only for calls that fail because the agent cannot reason through several steps.

Two older IDs have named replacements. The deprecations page points gemini-3.1-flash-live-preview to gemini-3.8-live, and gemini-3.1-flash-tts-preview to the new 3.8 Flash TTS or 3.8 Flash-Lite TTS models, which went GA on September 22. Our guide to building voice agents on Gemini 3.8 Live covers the full stack.

Images: Nano Banana 2.1, 2 Lite or Pro

For images, which Gemini model you need depends on resolution and text. Nano Banana 2.1 (gemini-nano-banana-2.1) went GA on October 6, per the changelog read October 7. It replaces Nano Banana 2 (gemini-3.1-flash-image), which the deprecations page now lists as deprecated, with no shutdown date. Nano Banana Pro (gemini-3-pro-image) went GA on May 28. Nano Banana 2 Lite (gemini-3.1-flash-lite-image) followed on June 30.

The image generation guide sets them apart:

• Nano Banana 2 Lite: Google’s fastest and cheapest image model. It outputs 1K only and does not take multiple reference images.
• Nano Banana 2.1: the general-purpose model, an update to Nano Banana 2. It outputs 1K, 2K or 4K, supports Google Search and Image Search grounding, handles text in images, and takes up to 10 object references plus 4 character images.
• Nano Banana Pro: the premium choice for complex visual work, with 4K output and separate slots for character and style references.

For featured images, social cards and product shots at web size, start with Nano Banana 2.1. At 1K it costs the same as Lite and does more. Use Lite when speed matters most. Keep Pro for layouts that the cheaper models get wrong. If a site calls gemini-3.1-flash-image today, plan the switch; Google’s guide recommends 2.1 for all new projects. Our explainer on what Nano Banana is covers the family’s background.

Video: Gemini Omni 1.1 Flash

For video, there is one answer to which Gemini model to use. gemini-omni-1.1-flash went GA on August 27. The Omni guide says a single generation runs up to 10 seconds, and extensions can bring a clip to 40 seconds. Output comes in 360p, 720p (the default), 1080p and upscaled 4K, in 16:9 or 9:16. Every output carries an invisible SynthID watermark. Editing or extending uploaded video is not available in the EEA, Switzerland or the UK.

Omni has no free tier. The pricing page bills video output at 5,792 tokens per second of 720p video, which it puts at about $0.10 a second. A 40-second 720p clip therefore costs roughly $4 in output before any retries. Draft at 360p to settle the prompt, then render the final version at the resolution you need. Our Gemini Omni Flash explainer covers the earlier preview.

Gemini API text and voice prices, as of October 14

These are standard paid-tier rates per million tokens from the Gemini API pricing page. Batch and Flex cost half. Priority costs 1.8 times the standard rate.

The pricing page lists the 3.6 and 3.8 Flash rates through December 31, 2026. On January 1, 2027 they double, and cached input doubles with them. We covered the same end date when Gemini 3.7 Flash launched. Any annual quote built on today’s Flash rate needs the January figure in it. Argon’s rate doubles a month earlier, on December 1. The 3.5 and 3.7 Flash rows are gone, since those IDs now route to newer models.

Image and video prices

Google prices image and video output per token, then gives per-item estimates. These are the estimates on the pricing page, read October 14.

At 1K, Nano Banana 2.1 costs the same as Lite and about half what Nano Banana 2 does. Pro costs about four times as much. At hundreds of images a month, the tier choice matters.

The free tier has a data trade-off

The pricing page answers one question for every model: whether Google uses your data to improve its products. On the free tier the answer is yes. On the paid tier it is no.

That matters for nonprofits, associations, schools and agencies handling member or donor records. The free tier is fine for testing with made-up data. Once real client content goes through the API, use a paid project. The free tier is also missing for Argon, the 3.1 Pro preview, the Nano Banana models and Omni, so those require billing from the start.

Model IDs that are closing or now answer differently

Two October 8 changes come first. deep-research-pro-preview-12-2025 is set to shut down on October 23, the day this post runs. Google names deep-research-preview-04-2026 and deep-research-max-preview-04-2026 as replacements; change the agent value in your requests. And gemini-3.7-flash and gemini-3.5-flash now route to gemini-3.8-flash and gemini-3.6-flash. Those endpoints are not shutting down, so nothing fails, but the answers and the bill can change. In effect, Google has changed which Gemini model answers those calls.

As read October 14, the deprecations page lists gemini-omni-flash-preview for shutdown on October 22, the day before this post runs. It now gives gemini-2.5-flash-image, the original Nano Banana, a shutdown date of March 15, 2027, with Nano Banana 2 Lite as the replacement. Our roundup of Gemini shutdown dates covers those moves.

Gemini 2.5 is the other one to watch. Since September 18, the Gemini API has limited 2.5 Pro, Flash and Flash-Lite to past users. The deprecations page lists no shutdown date. As of October 14, the Firebase AI Logic models page said "To be determined" for the Gemini Developer API and October 20, 2026 for the Agent Platform Gemini API. If your 2.5 calls run through Google Cloud, check your console now.

Further out, gemini-3.1-flash-lite closes on May 7, 2027. gemini-3-flash-preview points to gemini-3.6-flash with no date. The Live and TTS previews now close November 17, 2026, and point to the 3.8 models above.

How to confirm which Gemini model is right for your work

A table narrows the choice. Only your own workload settles which Gemini model to use. Take 30 to 50 real tasks from the job in question. Run them through the first-choice model and one step down. Compare the output, then compare the bill.

Count cost per finished task, not cost per million tokens. Two models with the same rate card can produce very different bills, because one writes more. Check thinking levels too. A lower level is often good enough for chat and tagging, and it cuts both latency and output tokens.

Keep the model ID in a single setting in each site or plugin. Google has shipped four Flash models since May. The next change should be a config edit, not a code search.

Frequently Asked Questions