Google Cuts Image API Prices and Launches Nano Banana 2.1 With Mask Editing

Google launched Nano Banana 2.1 on October 6, cutting API prices by roughly half at most resolutions, improving editing precision, and — critically for developers — deprecating its predecessor the same day. Teams currently running production workloads on Nano Banana 2 have until October 29 to migrate, a 23-day window that makes this release more than a routine model upgrade.

The new model, accessible via the API identifier gemini-nano-banana-2.1, is built on Google's Gemini 3.6 Flash architecture and is available immediately through Gemini App, Google AI Studio, and the paid API. The release marks the fourth major iteration of Google's Gemini-native image generation line since it launched in August 2025, following an accelerating cadence: Nano Banana (Aug 2025), Nano Banana Pro (Nov 2025), Nano Banana 2 (Feb 2026), and NB2.1 now.

A Price Cut That Comes With a Deadline

The pricing change is the most immediate operational fact for existing users. According to Google's official API pricing, generation at 2K resolution drops from $0.101 to $0.0504 per image — a cut of approximately 50%. The 1K tier also falls substantially, from $0.067 to $0.0336. At 4K, the price moves from $0.151 to $0.113, a reduction of roughly 25%. Batch-mode requests receive an additional 50% discount on top of those figures, bringing 4K generation to approximately $0.057 per image at scale — a significant shift for high-volume applications such as e-commerce product photography pipelines, real estate visualization, and advertising creative production.

The price cut arrives alongside a hard shutdown date. Nano Banana 2, which launched on Gemini 3.1 Flash Image in February 2026, was placed into deprecated status on October 6 and will stop accepting API requests on October 29. Google has not announced an extended grace period. The migration path is technically straightforward — update the model identifier, adjust any prompt engineering that relied on NB2-specific behaviors, and re-run integration tests — but the narrow 23-day window puts time pressure on engineering teams that have not yet begun. By comparison, Google's typical model deprecation windows for Gemini text models have run 90 to 180 days. The compressed timeline for NB2 appears to reflect the faster replacement cycle of the Flash-tier product line.

For consumer users on Gemini App and AI Studio, 50 daily credits remain free, and the upgrade is automatic.

How Gemini 3.6 Flash Changes the Generation Pipeline

Nano Banana 2.1's capabilities extend well beyond the previous version's feature set. The underlying Gemini 3.6 Flash model brings three key architectural additions to the image generation pipeline, alongside several targeted fixes that address known pain points from NB2.

The first addition is mask-based semantic editing. Where Nano Banana 2 required full-image regeneration to change any element, NB2.1 lets users draw a region — or describe it in text — and replace only that area while preserving the surrounding scene. The model segments the selection semantically rather than at the pixel level, then generates replacement content that matches the scene's existing lighting, perspective, and edge transitions. This closes one of the most-cited gaps between NB2 and competitors such as GPT Image 2.5. In practice, it enables workflows like replacing a product background without touching the foreground subject, or swapping a garment color while preserving fabric texture and shadow, operations that previously required manual masking in a dedicated editor.

The second addition is multi-subject consistency. NB2.1 accepts up to 14 reference images in a single context window — enabled by the Gemini family's 1-million-token context — and tracks visual identity for up to four named characters and ten object categories across an editing session. This matters for any workflow involving brand characters, product families, or recurring human subjects: the model can maintain a character's facial structure, clothing, and style across an entire image series without fine-tuning. It eliminates a class of consistency errors that previously required either post-processing compositing or expensive custom model training.

The third addition — and the most distinctive to Google's ecosystem — is Google Image Search grounding. Before rendering, NB2.1 can query Google Web Search and, new in this release, Google Image Search, retrieving current visual references to anchor generated content factually. The capability addresses a known limitation shared by all image generation models trained on historical data: the inability to represent recent products, updated landmark appearances, or current brand assets accurately. A generator that has retrieved a reference image of a product released last month can incorporate its actual design rather than hallucinating an approximate version based on training-set patterns. No competing consumer image model currently offers equivalent live-search integration.

Beyond these three additions, NB2.1 also resolves a persistent tiling defect from Nano Banana 2 that produced visible seam artifacts on wide and panoramic compositions at 2K and 4K resolutions. Google has not published a technical explanation of the fix, but community testing confirms the artifacts are gone on standard wide aspect ratios. The release also includes explicit improvement to Chinese-character text rendering — Chinese typography in AI-generated images has historically suffered from garbled characters, incorrect stroke order, and unstable glyph shapes. NB2.1's improvement, confirmed in independent designer testing, is sufficient for basic Chinese-language commercial poster production.

NB2.1 also introduces configurable thinking levels — Minimal, Medium (the default), and High — that let callers trade inference time and cost against output quality. This adapts a pattern from reasoning models, where inference-time computation is treated as a tunable parameter, and applies it to image synthesis. The exact latency and cost differences between levels are not published in Google's official documentation, but community testing suggests Minimal roughly halves generation time compared to Medium while producing visibly lower detail at 4K.

Where NB2.1 Sits Against GPT Image 2.5

The independent Arena crowd-preference leaderboard placed Nano Banana 2.1 at rank 5 in text-to-image generation with an Elo score of 1,328, and at rank 6 in image editing with 1,428, at the time of its launch on October 6. Both represent meaningful gains over Nano Banana 2. The rankings above NB2.1 in text-to-image include GPT Image 2.5 variants in the top positions, followed by GPT Image 2, Grok Imagine 2.0, and MAI Image 2.6 from Microsoft. Arena scores are generated by large-scale blind preference voting rather than technical accuracy tests, which means they measure aesthetic preference and prompt adherence, not factual accuracy or downstream task utility.

Community side-by-side comparisons between NB2.1 and GPT Image 2.5 consistently report that NB2.1 matches or exceeds GPT Image 2.5 on generation speed but falls short on the finished quality of photorealistic human subjects — specifically, micro-detail consistency in eyes, skin texture, hair, and fabric at 4K magnification. For abstract illustration, graphic design, typographic layouts, and product photography with non-human subjects, the gap is substantially smaller and may not be operationally significant.

NB2.1's pricing is competitive with GPT Image 2.5 at the 2K tier and is favorable at scale, where Google's per-image rates compare well for high-resolution batch generation. GPT Image 2.5 uses token-based billing rather than per-image flat rates, which makes direct comparison dependent on specific generation parameters, but high-resolution outputs on both platforms are priced within a similar range for most workloads.

The competitive factor that has no direct equivalent in GPT Image 2.5 is Google's ecosystem integration. NB2.1 operates natively within Google Flow, where generated images can be handed directly to the Veo video generation model for animation, creating a continuous production path from static concept to animated video inside a single platform without file export and re-import steps. For studios, agencies, and content operations already embedded in Google Workspace, that integration is a practical workflow advantage that Arena Elo scores do not capture.

Two Limitations Developers Should Test Before Scaling

Alongside the improvements, two quality issues are active concerns for production use.

The first is the watermark accumulation bug. NB2.1 embeds Google's SynthID provenance watermark invisibly into every generated image, combined with C2PA Content Credentials metadata in the file header, enabling downstream detection of AI-generated content in compliance with emerging disclosure requirements. In single-generation workflows, the invisible watermark causes no visible degradation. In iterative editing sessions — where the same image undergoes multiple successive edits — the pixel-level watermark accumulates with each round, progressively introducing noise, color drift, and texture distortion. Community tests document images becoming visibly degraded after three to five edit rounds. Google has not disclosed a timeline for a fix. Teams designing multi-step editing pipelines should test their specific round-trip counts before deploying NB2.1 at scale and may need to implement intermediate export-reimport steps as a workaround.

The second limitation is the quality ceiling on photorealistic human subjects, which remains below GPT Image 2.5 in community evaluations. This is most visible at 4K in portrait and close-crop contexts. The gap is narrower than it was with Nano Banana 2, but it has not closed. Organizations whose core use case is high-fidelity human portraiture should run comparative tests on their specific prompt types rather than treating the Arena benchmark improvement as a proxy for their workflow.

The Migration Deadline and What Comes Next

The practical priority for the next 20 days is migration. Developers running Nano Banana 2 via API need to update the model identifier, re-run quality and regression tests against the new architecture's output characteristics, and revise cost projections using the new pricing tier. The output style of NB2.1 differs from NB2 — particularly in color grading and sharpness at 4K — so prompts optimized for NB2 may require adjustment to produce equivalent results. Google's deprecation notice does not offer extensions; October 29 is the confirmed shutdown date.

Looking beyond the transition, the Nano Banana product line's history suggests a Pro-tier update is the next expected release. Nano Banana Pro, launched November 2025, delivered the highest quality in the family at a premium price point. Whether a Nano Banana 2 Pro built on a higher-capability Gemini 3.x Pro model is planned, and whether it would close the remaining gap with GPT Image 2.5 on photorealistic output, is the next competitive question worth watching. Until then, NB2.1's case is strongest where the evidence is clearest: for developers who need fast, cost-effective image generation with native Google ecosystem integration and live search grounding, the price drop and editing improvements represent a substantial practical upgrade — provided migration is complete before October 29.