Google pushed a new image-generation model into production on October 6, 2026, and the rollout arrived with almost none of the keynote fanfare that usually accompanies a Gemini release. Nano Banana 2.1, listed in Google’s developer documentation under the model ID gemini-nano-banana-2.1, is now live across the Gemini app, Google AI Studio, AI Mode in Search, Flow, Stitch, Google Ads, and the Gemini Enterprise Platform, according to Google DeepMind’s own model card. The headline number: standard image output pricing lands at $30 per million tokens, roughly half the generation cost Google charged for the outgoing Nano Banana 2 tier, per figures confirmed on Google’s official Gemini API pricing page.
That single pricing move reshapes the economics of the AI image-generation market just as rivals OpenAI, Adobe, Midjourney, and xAI compete for the same enterprise and creator budgets. It also lands in the middle of a broader pattern this fall: Google has now shipped or updated three major AI products — Gemini 4 Argon, Nano Banana 2, and now Nano Banana 2.1 — inside roughly six weeks, each time trimming free access or prices somewhere else in the stack.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Nano Banana 2.1 actually changes
Nano Banana 2.1 succeeds both Nano Banana 2 and Nano Banana Pro on what Google frames as its faster, lower-cost “Flash”-style tier, according to listings on Google’s Gemini API model documentation. The model is part of the Gemini 3 family and handles both text-to-image generation and conversational, iterative editing — meaning a user can ask for a follow-up change to an image without restating the entire original prompt.
Reported improvements, drawn from outlet coverage of the October 6 rollout, include stronger consistency when a reference image (a face, a product, a logo) needs to survive multiple edits, sharper text rendering inside generated images, and support for 1K, 2K, and 4K output resolutions. Coverage of the launch specifically called out improved visual quality and consistency as the standout upgrade over Nano Banana 2. None of these claims come with a full official changelog from Google, so they should be read as reported improvements rather than a verified line-by-line spec sheet.
One detail did not improve: input pricing. Nano Banana 2.1 charges $1.50 per million tokens for text, image, or video input, roughly three times the $0.50 per million Nano Banana 2 charged, according to pricing comparisons reported by TokenPost. Google appears to have redistributed the cost curve rather than cut it everywhere — output got dramatically cheaper, input got notably more expensive. For developers running high-input-volume workflows (long conversational editing sessions, bulk reference-image uploads), that tradeoff matters more than the headline “half price” framing suggests.
The pricing breakdown, token by token
Google’s published Gemini API pricing converts into the following per-image costs, confirmed directly from the official pricing documentation:
| Output resolution | Token consumption | Standard price per image | Batch price per image |
|---|---|---|---|
| 1K (1024×1024) | 1,120 output tokens | $0.0336 | $0.0168 |
| 2K (2048×2048) | 1,680 output tokens | $0.0504 | $0.0252 |
| 4K (4096×4096) | 3,780 output tokens | $0.113 | $0.0567 |
Those figures sit underneath a simpler underlying rate: $30 per million output tokens on standard processing, or $15 per million on batch jobs that can tolerate delayed delivery. Input tokens — covering text prompts, reference images, or video frames fed into the model — run $1.50 per million on standard and $0.75 per million on batch. Text and “thinking” output, separate from the image tokens themselves, is billed at $7.50 per million standard and $3.75 per million batch, per the same Google documentation.
Developers already running Gemini-based pipelines through Nano Banana 2 or Imagen 4 integrations will need to re-test prompts against the new input/output cost ratio before assuming the upgrade is a straightforward savings. For teams that were primarily output-bound — generating lots of final images from short prompts — the switch is close to a 50% cut. For teams that lean on long conversational edit chains with heavy reference-image input, the math is less favorable.
A sample API call under the new pricing
For developers testing the switch, the model ID change in a standard Gemini API request looks like this:
curl -X POST \
"https://generativelanguage.googleapis.com/v1beta/models/gemini-nano-banana-2.1:generateContent" \
-H "x-goog-api-key: $GOOGLE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"contents": [{
"parts": [{"text": "Edit the uploaded product photo: swap the background to a studio gradient, keep the product identical"}]
}],
"generationConfig": {
"responseModalities": ["IMAGE"],
"imageConfig": {"resolution": "2K"}
}
}'Swapping the model string from a prior Nano Banana release to gemini-nano-banana-2.1 is, per Google’s own documentation, the only change required to route existing integrations onto the new pricing and capability tier — though teams should budget for the input-token increase described above before shipping to production.
Benchmark claims: a 1,050 Elo score, with caveats
The most specific performance numbers in circulation come from TokenPost’s reporting on an internal preference-Elo evaluation with “thinking” enabled. Nano Banana 2.1 reportedly scored 1,050, ahead of Nano Banana 2’s 990 and Nano Banana Pro’s 935. On a separate infographic-factuality metric — essentially, how often the model gets labels, numbers, and charts right inside a generated image — Nano Banana 2.1 scored 0.521 against 0.179 for Nano Banana 2 and 0.265 for Nano Banana Pro.
Those are meaningful jumps if they hold up under independent testing, but it’s worth being precise about what has and hasn’t been verified. The Elo figure is not confirmed to be the same benchmark used on the public LMArena leaderboard that tracks model-versus-model voting, and no sample size, methodology writeup, or confidence interval accompanies the reported scores. Readers comparing Nano Banana 2.1 against competing image models should treat the 1,050 number as a vendor-adjacent data point, not an independently audited leaderboard rank, until LMArena or a comparable third party publishes its own placement.
How Nano Banana 2.1 stacks up against the competition
The most direct, apples-to-apples comparison available right now is against OpenAI’s GPT Image 2.5 family, released as the Sunburst and Flare variants roughly a month earlier. Both OpenAI models charge the identical $30-per-million-token rate for standard image output, confirmed on OpenAI’s own GPT Image 2.5 Sunburst documentation. That means Google’s “half price” move doesn’t undercut OpenAI on output cost — it matches it. Where the two diverge is input: GPT Image 2.5 charges $8 per million tokens for image input and $5 per million for text input, both notably above Nano Banana 2.1’s $1.50 blended input rate.
| Model | Standard image output (per 1M tokens) | Input pricing (per 1M tokens) | Max output resolution |
|---|---|---|---|
| Nano Banana 2.1 (Google) | $30.00 | $1.50 (text/image/video, blended) | 4K |
| Nano Banana 2 (Google, prior gen) | ~$60 equivalent* | $0.50 | 2K |
| GPT Image 2.5 Sunburst/Flare (OpenAI) | $30.00 | $8.00 (image), $5.00 (text) | Not uniformly published |
| GPT Image 1 (OpenAI, legacy) | $40.00 | $10.00 (image), $5.00 (text) | 1536×1536 |
*Derived from the reported roughly-50%-cheaper framing applied to Nano Banana 2.1’s confirmed $30 rate; Google has not published a standalone Nano Banana 2 output-token price for direct citation, so this figure is an approximation, not an official Google number.
Midjourney and Adobe Firefly remain outside this kind of direct comparison because both sell access through subscription tiers and consumption credits rather than a published per-token or per-image API rate, and neither company has put out a 2026 pricing page specific enough to convert into the same unit economics. xAI’s Grok Imagine is in a similar position: no verifiable public API price for image generation has surfaced as of this writing. That opacity is itself a competitive data point — Google and OpenAI are now the two companies willing to publish transparent, comparable, developer-facing image-token pricing, which makes them easier to benchmark against each other and harder to benchmark against the subscription-based players.
The scale behind the release
Nano Banana’s cumulative usage gives some sense of why Google keeps iterating on this specific model line so aggressively. At Google I/O in May 2026, CEO Sundar Pichai said the Nano Banana family had generated more than 50 billion images since its August 2025 launch — up from roughly 5 billion images reported just months earlier, in October 2025. Those figures come from Pichai’s own keynote remarks as reported by multiple outlets tracking the event; Google has not published a separate, independently auditable counter for verification, so the number should be read as a company-disclosed figure rather than a third-party-measured one.
Whatever the precise count, a model generating billions of images a month justifies the kind of granular pricing engineering Google applied here: shaving the output price while raising the input price is the kind of change that only moves the needle at genuine scale. A hobbyist generating a handful of images a day won’t notice; a company running automated product-photo pipelines or ad-creative generation at thousands of images per hour will feel both sides of that tradeoff immediately.
Market size and where the money is going
Estimates for the overall AI image-generation market in 2026 vary by a wide margin depending on scope. Narrower estimates that count only dedicated image-generation tools put the figure closer to $4.8 billion, while broader estimates that fold in image features bundled into advertising platforms, productivity suites, and creative software — the kind of bundling Google is doing by pushing Nano Banana 2.1 into Search, Ads, and Workspace-adjacent tools simultaneously — push the number toward $12.4 billion, a figure reported via MarketsandMarkets research summaries. Projected compound annual growth across these estimates runs between roughly 17% and 40%, reflecting just how unsettled the market’s boundaries still are.
That spread matters for how this release should be read. Google isn’t just competing in a standalone “image generator” category against Midjourney; by wiring Nano Banana 2.1 into Google Ads, AI Mode in Search, and the Gemini Enterprise Platform at once, it’s positioning the model as infrastructure inside products that already have massive existing user bases, not as a separate app fighting for new downloads.
Part of a six-week pattern of quiet Google releases
Gemini 4 Argon set the tone in late September
Nano Banana 2.1 is the third major Google AI update inside roughly six weeks. Gemini 4 Argon launched September 30 as Google’s newest frontier text-and-reasoning model, and within days the company had already tightened free-tier access to its older Gemini models. The pattern repeating here — ship the new flagship, then adjust pricing and access around it — is now showing up on the image-generation side of the product line too.
A rollout with no keynote
What’s notable about Nano Banana 2.1 specifically is how it shipped. Several outlets described the launch as quiet — no onstage reveal, no dedicated press event, just an updated model card and a rollout across products that most users encounter without realizing the underlying model changed. Some coverage even noted the model briefly surfaced inside Google’s Flow video tool before Google formally confirmed the release elsewhere, suggesting the rollout itself was still being finalized across products on launch day.
That’s a deliberate shift in how Google treats incremental model updates. Flagship releases like Gemini 4 Argon get blog posts and benchmark charts; point-release updates like Nano Banana 2.1 get a model card and a pricing page update. It’s a sign the model-release cadence across the industry has outpaced the audience’s ability to track every version bump — competitors are shipping updates fast enough that quiet releases are becoming the norm rather than the exception.
Historical context: from Nano Banana to Nano Banana 2.1 in 14 months
The original Nano Banana launched in August 2025, less than fourteen months before this update. In that window, Google has already cycled through Nano Banana 2, Nano Banana Pro, and now Nano Banana 2.1 — a release cadence that mirrors what happened on the text-model side of Gemini, where the jump from Gemini 3.8 to Gemini 4 Argon took a comparable few months. Image-generation models, once treated as a slower-moving category than text LLMs, are now iterating on nearly the same timeline.
That acceleration lines up with how OpenAI has handled its own image line. GPT Image 1 launched in 2025 and reportedly generated 700 million images in its first week alone. OpenAI followed with GPT Image 2.5’s Sunburst and Flare variants roughly a year later, cutting the per-token output rate from $40 to $30 in the process — almost the same percentage reduction Google just applied to Nano Banana 2.1’s output pricing. Both companies appear to be converging on $30-per-million-output-tokens as something close to a market rate for frontier image generation, at least on the standard (non-batch) tier.
What this means for developers and enterprise buyers
For teams evaluating whether to build on Nano Banana 2.1, GPT Image 2.5, or stick with an existing integration, three practical considerations stand out:
- Workloads dominated by bulk final-image generation (marketing creative, product catalogs) benefit most from Nano Banana 2.1’s output-price cut.
- Workloads built around long conversational editing chains or heavy reference-image uploads should model out the new $1.50-per-million input rate before migrating, since the net cost could rise rather than fall.
- Enterprises already standardized on the Gemini Enterprise Platform get Nano Banana 2.1 bundled in automatically, which changes the calculus compared with teams that would need to stand up a brand-new integration from scratch.
Google has not published a complete account of Nano Banana 2.1’s safety or watermarking behavior specific to this release. Google has used its SynthID invisible-watermarking system across prior Nano Banana generations, but no dedicated Nano Banana 2.1 safety report or confirmation of SynthID’s exact implementation on this version — including whether watermarks survive conversational edits — has surfaced in the documentation reviewed. Enterprise buyers with provenance or compliance requirements should treat that as an open question to raise directly with Google’s enterprise sales team rather than an assumed carryover from earlier models.
Competitor reactions so far: mostly silence
As of this writing, no on-record statements from OpenAI, Midjourney, Adobe, or xAI directly addressing the Nano Banana 2.1 launch have surfaced in available reporting. That’s not unusual this early in a release cycle — pricing responses in this market have historically taken weeks rather than days — but it does mean any claim of a price war remains speculative rather than confirmed until a competitor actually moves its own rate card. What is confirmed is the structural overlap: Google and OpenAI are now charging identical headline output rates, which puts pressure on both companies to differentiate on either input economics, model quality, or platform bundling rather than sticker price alone.
What to watch next
Several developments are worth tracking over the coming weeks. First, whether LMArena or another independent benchmark operator publishes its own Nano Banana 2.1 ranking, which would settle the question of how the reported 1,050 Elo score holds up outside Google’s own testing. Second, whether OpenAI, Adobe, or Midjourney respond with pricing or feature changes of their own, confirming or disproving the price-war framing some early coverage has applied to this release. Third, whether Google extends the same input/output price restructuring to Nano Banana Pro or to Imagen-branded products, which would suggest this is a company-wide pricing strategy rather than a one-off adjustment. Fourth, whether the older-API shutdown referenced in some early coverage of the rollout materializes on a confirmed timeline, which would force remaining Nano Banana 2 and Imagen 4 integrations to migrate on a deadline rather than at their own pace. Fifth, whether Google publishes an updated cumulative image-generation count beyond the 50-billion figure disclosed at I/O in May, which would be the clearest signal yet of how much of this market Google has already captured relative to OpenAI, Midjourney, and Adobe combined.
The bigger picture for AI model pricing
Nano Banana 2.1 is a small release in terms of announcement volume but a meaningful one in terms of what it signals about where frontier AI pricing is heading. Output costs, the part of the bill tied directly to what a customer actually receives, are compressing toward a shared market rate near $30 per million tokens across the two biggest players. Input costs, the part tied to how much context, reference material, or conversation history a workflow demands, are where companies are now experimenting with increases to offset those output cuts. That’s a more complicated pricing picture than the simple half-price headline suggests, and it’s one enterprise buyers evaluating any frontier model family, text or image, should expect to see repeated across the rest of 2026 as the major labs continue shipping updates on compressed timelines.
Frequently Asked Questions
What is Nano Banana 2.1?
Nano Banana 2.1 is Google’s latest image-generation and conversational-editing model, released October 6, 2026, under the API model ID gemini-nano-banana-2.1. It succeeds Nano Banana 2 and Nano Banana Pro on Google’s Flash-tier lineup.
How much does Nano Banana 2.1 cost to use?
Standard image output is priced at $30 per million tokens ($0.0336 for a 1K image, $0.0504 for 2K, $0.0756 for 4K), with batch processing available at half those rates. Input tokens cost $1.50 per million on standard processing, per Google’s official Gemini API pricing page.
Is Nano Banana 2.1 cheaper than Nano Banana 2?
Output pricing is roughly 50% cheaper, according to reporting on the launch. However, input pricing rose from $0.50 to $1.50 per million tokens, so total cost depends heavily on how input-heavy a given workflow is.
How does Nano Banana 2.1 compare to GPT Image 2.5?
Both models charge an identical $30 per million tokens for standard image output. GPT Image 2.5 charges higher input rates ($8 per million for image input, $5 per million for text input) compared with Nano Banana 2.1’s $1.50 blended input rate.
Where is Nano Banana 2.1 available?
It is rolling out across the Gemini app, Google AI Studio, AI Mode in Google Search, Flow, Stitch, Google Ads, and the Gemini Enterprise Platform, according to Google’s own rollout announcement.
Is the Nano Banana 2.1 benchmark score independently verified?
No. The reported 1,050 Elo preference score and 0.521 infographic-factuality score come from third-party reporting (TokenPost) rather than a confirmed public leaderboard such as LMArena. Independent verification has not yet surfaced.
What happened to older Nano Banana API versions?
Some early coverage referenced an eventual shutdown of an older image API tied to this release, but no confirmed shutdown date has been independently verified as of this writing. Developers on older integrations should monitor Google’s API deprecation notices directly.
Does Nano Banana 2.1 watermark its images?
Google has used SynthID invisible watermarking on prior Nano Banana models, but no Nano Banana 2.1-specific confirmation of SynthID’s exact implementation, including behavior after conversational edits, has been published.
Related Coverage
- Nano Banana 2 API: Migrate From Imagen 4 in 13 Steps [2026]
- Z-Image Turbo Setup: 6B Params, 14 Steps [2026]
- Zetaris Launches Cloud AI Data Harness, Teams With SpaceXAI for Global Hackathon
- The Best AI Image Generators Compared: Features, Use Cases and Pricing
- 5 Best Free AI Image Generators in 2026: Tested and Compared
