The AI Stack Cost Report · October 2026 · The Image Stack. StackSpend's monthly read on what the modern AI engineering stack actually costs. There was no September issue, so this edition covers everything that changed between early August and October 6, 2026. Also in this issue: the frontier briefing, the pricing map, and the coding stack.
Use this when you run an image feature in production (product shots, ad creative, user-generated edits) and need a current answer to "which model, for which job, at what monthly cost".
The fast answer: OpenAI's GPT-Image-2.5 Sunburst and Flare (September 8) now hold the top two places on both Arena image leaderboards. Google shipped Nano Banana 2.1 today (October 6) at roughly half of Nano Banana 2's per-image price, and its predecessor is scheduled to retire on October 29. FLUX 3 Image finally reached the API on October 1, with a steep 4K price. And Imagen 4 is gone from the Gemini API as of August 17. If you have not migrated off Imagen 4 or
gemini-3.1-flash-image, that is this month's work, and it is also a chance to cut the unit price.
This guide covers the current leaders, what each costs per image or per token, which model fits which job, and a worked monthly bill showing how much the choice matters. Prices are list rates checked on October 6, 2026; treat them as a point-in-time snapshot.
Quick answer
- Best overall quality and editing: GPT-Image-2.5 Sunburst, No. 1 on Arena text-to-image (1425) and image edit (1524). Token-priced at $5 text in, $8 image in, $30 image out per 1M tokens.
- Best quality for production volume: GPT-Image-2.5 Flare, No. 2 on both boards, same token rates, and OpenAI says up to 50% lower latency than GPT-Image-2.
- Best value among the leaders: Nano Banana 2.1: $0.0336 at 1K, $0.0504 at 2K, $0.0756 at 4K, half that on Batch. No. 5 on text-to-image after one day of votes.
- Strong token-priced challenger: MAI-Image-2.6 (Microsoft), No. 4 on text-to-image, reported at $38 per 1M image-output tokens (about $0.039 per image by Arena's estimate).
- Flat per-image pricing: Grok Imagine Image 2.0 ($0.04), Seedream 5.0 Pro ($0.045 up to 2.36 MP), Qwen-Image-3.0 Pro ($0.04), FLUX 3 Image ($0.048 at 1K).
- Specialists: Recraft V4.1 for design and vector output ($0.035 raster, $0.08 vector; V4.1 Flash at $0.007), Ideogram 4.0 for typography, Midjourney V8.x for art direction (subscription, no public API price).
- Open weights: Qwen-Image-2.1 (7B, native transparency) is non-commercial; FLUX.2 [klein] from $0.014 on BFL's API is the cheapest current-generation option we could verify.
What changed since August
GPT-Image-2.5 Flare and Sunburst (OpenAI, September 8). ChatGPT Images 2.5 shipped with two API models split by workload: Flare as the faster default for high-volume pipelines, Sunburst as the slower, detail-first model for precise edits. Both list at $5 / $8 / $30 per 1M text-input / image-input / image-output tokens, which is the same rate card as gpt-image-2. OpenAI's pricing page sends you to a calculator for per-image estimates rather than publishing them for 2.5.
MAI-Image-2.6 (Microsoft, August 10; Foundry public preview September 4). It launched at No. 2 on Arena, 79 Elo points above MAI-Image-2.5, and then reached public preview in Microsoft Foundry with a cheaper MAI-Image-2.6-Flash. Microsoft says Flash is 2.8x faster than GPT-Image-2-Medium. Reported rates: $5 / $8 / $38 per 1M tokens for 2.6, and $1.75 / $2.50 / $19 for Flash. Microsoft's blog posts do not state the rates themselves, so treat them as reported, not first-party.
Grok Imagine Image 2.0 (xAI, around August 7–8). xAI's second-generation image model is listed at $0.04 per image in its model docs and sits at No. 6 (canvas) and No. 8 on Arena text-to-image.
Seedream 5.0 Pro (ByteDance, early August). It is reported at $0.045 per image up to 2.36 MP and $0.09 above, with each extra reference image costing $0.003. It holds No. 12 on Arena text-to-image with over 93,000 votes.
Imagen 4 shut down on the Gemini API (August 17). All three endpoints (imagen-4.0-generate-001, -ultra-, -fast-) are past their shutdown date. Google's deprecations page now names gemini-nano-banana-2.1 as the replacement. gemini-2.5-flash-image, the original Nano Banana, was also scheduled to shut down on October 2.
Ideogram API price cut (August 25, reported). Ideogram 4.0 API pricing reportedly halved to $0.03 / $0.06 / $0.10 per image for Turbo / Default / Quality. Ideogram's own pricing page loads its rates dynamically, so we could not confirm them first-hand.
Recraft V4 Styles (late August). The new Styles family is priced at $0.035 raster, $0.05 vector, and $0.10 / $0.12 for Pro, plus $0.005 per style-reference request.
Qwen-Image-2.1 (Alibaba, September 20). This is a 7B open-weight model that generates, edits and outputs native RGBA transparency from one checkpoint. Unlike its Apache-2.0 predecessor, it ships under a non-commercial research license. Commercial use needs a separate agreement.
FLUX 3 Image (Black Forest Labs, October 1). The image tier of FLUX 3, which was missing from our August issue, is now on the API: one endpoint for generation and pixel-exact local edits, up to ten reference images, layout control via bounding boxes in the prompt, and output up to 4K. Pricing is $0.041 (768 square), $0.048 (1K), $0.100 (2K), $0.607 (4K) per image. It does not yet appear on Arena's text-to-image board.
Nano Banana 2.1 (Google, October 6). Built on Gemini 3.6 Flash, it is rolling out across the Gemini app, AI Mode, AI Studio, Flow and the API. At $0.0336 / $0.0504 / $0.0756 for 1K / 2K / 4K, it costs about half of Nano Banana 2 ($0.067 / $0.101 / $0.151). gemini-3.1-flash-image (Nano Banana 2) has an October 29, 2026 shutdown date.
Midjourney: no new version. There is no V9. The changes are alpha changelog updates (September 16, September 23, October 1) and editor improvements, so V8.x remains current.
The October 2026 leaderboard
| Rank | Model | Text-to-image score | Image-edit score | Note |
|---|---|---|---|---|
| 1 | GPT-Image-2.5 Sunburst (OpenAI) | 1425 | 1524 (No. 1) | Detail-first, slower |
| 2 | GPT-Image-2.5 Flare (OpenAI) | 1398 | 1481 (No. 2) | Speed-first default |
| 3 | GPT-Image-2 medium (OpenAI) | 1383 | 1462 (No. 3) | Previous generation, same token rates |
| 4 | MAI-Image-2.6 (Microsoft) | 1333 | 1428 (No. 5) | Token-priced, Foundry preview |
| 5 | Nano Banana 2.1 (Google) | 1328 ±9 | 1428 (No. 6) | ~5,300 votes on day one: provisional |
| 6 / 8 | Grok Imagine Image 2.0, canvas / standard (xAI) | 1321 / 1297 | 1439 / 1425 | Flat $0.04 per image |
| 9 | Muse Image (Meta) | 1274 | 1403 | Meta's first in-house image model |
| 12 | Seedream 5.0 Pro (ByteDance) | 1256 | 1394 (No. 10) | Strong quality per dollar |
| 16 | Nano Banana Pro, 2K (Google) | 1248 | — | Now outranked by the cheaper 2.1 |
Arena leaderboards, checked October 6, 2026. OpenAI holds the top three places on both boards, which is a clearer lead than in August. The more useful finding for cost is Nano Banana Pro at No. 16, below Nano Banana 2.1 at about a quarter of its 1K price. Google's own model card shows 2.1 beating the older Pro model on some internal editing evaluations, though independent reviewers note Pro still produces better images in some hands-on tests. FLUX 3 Image, Midjourney and Ideogram 4.0 at its top setting are either absent or ranked lower (Ideogram 4.0 Quality is No. 20), so arena rank says little about their specialist strengths. Benchmark on your own prompts before switching.
What image generation costs (October 2026)
| Model | Pricing basis | ~1K image | 2K / 4K | Batch / notes |
|---|---|---|---|---|
| GPT-Image-2.5 Flare / Sunburst | Tokens: $5 text in, $8 image in, $30 image out per 1M | Not published; use OpenAI's calculator | Up to 3840 px edge | Batch 50% off |
| GPT-Image-2 | Same token rates | $0.006 low / $0.053 medium / $0.211 high (1024×1024) | — | Official per-image estimates |
| MAI-Image-2.6 / 2.6-Flash | Tokens: $5/$8/$38 and $1.75/$2.50/$19 per 1M (reported) | ~$0.039 (Arena estimate) | Up to 1.5K | Foundry public preview |
| Nano Banana 2.1 | Per image | $0.0336 | $0.0504 / $0.0756 | Batch: $0.0168 / $0.0252 / $0.0378 |
| Nano Banana 2 (retiring Oct 29) | Per image | $0.067 | $0.101 / $0.151 | Batch 50% off |
| Nano Banana Pro | Per image | $0.134 | $0.134 / $0.24 | Premium editing |
| FLUX 3 Image | Per image by tier | $0.048 | $0.100 / $0.607 | $0.041 at 768 square |
| FLUX.2 [pro] / [klein] | Per megapixel | From $0.03 / from $0.014 | Scales with MP | [flex] from $0.05, [max] from $0.07 |
| Grok Imagine Image 2.0 | Per image | $0.04 | — | xAI model docs |
| Seedream 5.0 Pro | Per image (reported) | $0.045 (≤2.36 MP) | $0.09 above 2.36 MP | +$0.003 per extra reference |
| Qwen-Image-3.0 Pro / Standard | Per image (reported) | $0.04 / $0.03 | $0.075 at 2K (reported) | Closed weights |
| Recraft V4.1 / V4.1 Flash | Per image | $0.035 / $0.007 | Pro 2K $0.21 | Vector $0.08 |
| Ideogram 4.0 Turbo / Default / Quality | Per image (reported) | $0.03 / $0.06 / $0.10 | — | After Aug 25 cut |
The table spans roughly 30x at 1K, from $0.007 (Recraft V4.1 Flash) to $0.211 (GPT-Image-2 high). The resolution columns matter more than the headline. Nano Banana 2.1 charges 2.25x more for 4K than for 1K, which works out to about $0.0045 per megapixel at 4096×4096 against $0.032 at 1024×1024. FLUX 3 Image charges 12.6x more for the same step ($0.048 to $0.607). Moving a pipeline from 1K to 4K is a modest change on Google and a large one on FLUX 3. OpenAI and Microsoft bill by token, so their per-image cost depends on resolution, quality and the number of input images in each call, and the published rate cards cannot tell you your bill.
Which model for which job
| Job | First choice | Cheaper alternative | Why |
|---|---|---|---|
| Hero and campaign creative, complex briefs | GPT-Image-2.5 Sunburst | GPT-Image-2.5 Flare | Top of both Arena boards; Sunburst preserves unedited regions across successive edits |
| Product catalog at volume | Nano Banana 2.1 (Batch) | FLUX.2 [klein], Recraft V4.1 Flash | Top-five quality at $0.0168 per 1K batch image |
| Iterative editing in an app | GPT-Image-2.5 Flare | MAI-Image-2.6-Flash, Grok Imagine Image 2.0 | Edit-arena leaders at interactive latency |
| Precise layout and multi-reference composition | FLUX 3 Image | Seedream 5.0 Pro | Bounding boxes in prompt, up to ten references; keep it at 1K–2K |
| Typography, posters, ads with text | Ideogram 4.0 | MAI-Image-2.6 | Ideogram's specialism; MAI reports +91 Elo on text rendering |
| Logos, icons, vector assets | Recraft V4.1 vector | Recraft V4 Styles vector ($0.05) | Native SVG output |
| Art direction and mood | Midjourney V8.x | — | Aesthetic lead, subscription-billed |
| Transparent assets, self-hosted research | Qwen-Image-2.1 | GPT-Image-2.5 (transparent backgrounds) | Native RGBA, but non-commercial license |
The trade-off that matters most this month is in product imagery. GPT-Image-2.5 is clearly better on the leaderboards, while Nano Banana 2.1 is about two arena tiers behind and much cheaper. For a catalog where the acceptance bar is "clean, on-brand, correct product", the cheaper model usually clears it. For a hero image that goes on a billboard, the extra spend is easy to justify. Run both on 50 of your own SKUs, then score how many images pass first time, because the cost of rejected images belongs in the comparison.
Worked example: 50,000 product images a month
Assume an e-commerce team generates 50,000 product images a month at ~1K (1024×1024), one output per request, with no reference images beyond the product photo.
| Model and setting | Unit price | Monthly bill | With 1.5 attempts per accepted image |
|---|---|---|---|
| Recraft V4.1 Flash | $0.007 | $350 | $525 |
| Nano Banana 2.1, Batch | $0.0168 | $840 | $1,260 |
| Nano Banana 2.1, standard | $0.0336 | $1,680 | $2,520 |
| FLUX.2 [pro] (from) | $0.03 | $1,500 | $2,250 |
| MAI-Image-2.6 (Arena estimate) | ~$0.039 | ~$1,950 | ~$2,925 |
| Grok Imagine Image 2.0 | $0.04 | $2,000 | $3,000 |
| Seedream 5.0 Pro | $0.045 | $2,250 | $3,375 |
| FLUX 3 Image, 1K | $0.048 | $2,400 | $3,600 |
| GPT-Image-2, medium | $0.053 | $2,650 | $3,975 |
| Nano Banana 2 (if not migrated) | $0.067 | $3,350 | $5,025 |
| Nano Banana Pro | $0.134 | $6,700 | $10,050 |
| GPT-Image-2, high | $0.211 | $10,550 | $15,825 |
The same workload ranges from $840 to $10,550 a month across models that would all be considered reasonable choices, a 12.5x spread before retries. Three results matter in practice:
- The free win is migrating from Nano Banana 2 to 2.1. A team still on
gemini-3.1-flash-imagepays $3,350. On 2.1 the same volume costs $1,680, or $840 on Batch if the images are not needed in real time, and the model ranks higher. The October 29 shutdown forces the move anyway. - Resolution changes the ranking. At 4K the same 50,000 images cost $3,780 on Nano Banana 2.1 ($0.0756) and $30,350 on FLUX 3 Image ($0.607). Check the 4K price before committing.
- Retries cost as much as the choice of model. At 1.5 attempts per accepted image, the bill rises by 50% with no change to the rate card. A cheaper model that needs two attempts can cost more than a dearer one that needs one.
GPT-Image-2.5 is left out of the table on purpose. OpenAI has not published per-image figures for it, and its token rates match GPT-Image-2's. If 2.5 used a similar number of output tokens per image, its cost would fall in the $0.053–$0.211 band shown for GPT-Image-2, but that is an inference. Measure it in your own account before budgeting. You can use the LLM cost calculator for the text side of a mixed pipeline.
The cost trap: token meters and silent multipliers
Image spend gets away from teams in the same few ways each month:
- Token-priced images have no unit price. On GPT-Image-2.5 and MAI-Image-2.6, a 4K edit with three reference images and a 1K generation are different line items on the same rate card. Streaming partial previews adds 100 image-output tokens per partial on OpenAI.
- Deprecations become emergency migrations. Imagen 4 (August 17), Gemini 2.5 Flash Image (October 2) and Nano Banana 2 (October 29) each forced a model change within one quarter, and each change moves the unit price.
- Variants multiply. A UI that shows four options per request multiplies the bill by four.
- Spend is spread across providers. Many teams now run a token-metered OpenAI or Microsoft model, a per-image Google or BFL model, and a Midjourney subscription, with three units of billing and no combined view.
Image API spend is usage-based and spiky. A batch job that loops on retries, or a product change that defaults to 4K, shows up on the invoice weeks later. StackSpend tracks AI provider spend daily with same-day anomaly alerts, so a jump in image spend is visible the day it happens. Start with AI cost monitoring, or OpenAI cost monitoring if GPT-Image-2.5 is your main model.
FAQ
What is the best AI image generator in October 2026?
On Arena's leaderboards as of October 6, 2026, OpenAI's GPT-Image-2.5 Sunburst is No. 1 for both text-to-image (1425) and image editing (1524), and GPT-Image-2.5 Flare is No. 2 on both. "Best" still depends on the job: Nano Banana 2.1 is the strongest quality-per-dollar option, Ideogram leads typography, Recraft leads vector output, and Midjourney V8.x leads art direction.
How much does GPT Image 2.5 cost?
GPT-Image-2.5 Flare and Sunburst are both priced at $5 per 1M text-input tokens, $8 per 1M image-input tokens and $30 per 1M image-output tokens, with a 50% Batch discount. OpenAI does not publish per-image figures for 2.5. For reference, GPT-Image-2, which has the same token rates, is estimated at $0.006 (low), $0.053 (medium) and $0.211 (high) per 1024×1024 image.
How much does Nano Banana 2.1 cost, and do I need to migrate?
Nano Banana 2.1 (released October 6, 2026) costs $0.0336 per 1K image, $0.0504 at 2K and $0.0756 at 4K, half that on the Batch API. That is about half of Nano Banana 2's price. Google's deprecations page lists an October 29, 2026 shutdown for gemini-3.1-flash-image (Nano Banana 2), with gemini-nano-banana-2.1 as the replacement, so yes, migrate now.
Is Imagen 4 still available?
Not on the Gemini API. imagen-4.0-generate-001, imagen-4.0-ultra-generate-001 and imagen-4.0-fast-generate-001 were shut down on August 17, 2026, and Google's deprecations page names Nano Banana 2.1 as the replacement. Plan for a per-image price change when you migrate: Imagen 4 Fast was $0.02, and Nano Banana 2.1 is $0.0336 at 1K, or $0.0168 on Batch.
How much does FLUX 3 Image cost?
FLUX 3 Image launched on Black Forest Labs' API on October 1, 2026 at $0.041 (768 square), $0.048 (1K), $0.100 (2K) and $0.607 (4K) per image. The 4K tier costs 12.6x the 1K tier, so it suits 1K–2K production work, or 4K only where the extra resolution earns its cost.
What is the cheapest way to generate product images at scale?
Among models near the top of the leaderboards, Nano Banana 2.1 on the Batch API ($0.0168 per 1K image) is the cheapest we could verify. Below that, Recraft V4.1 Flash ($0.007) and FLUX.2 [klein] (from $0.014) are cheaper but lower-ranked. Whichever you choose, track attempts per accepted image as closely as the unit price.
About The AI Stack Cost Report
The AI Stack Cost Report is StackSpend's monthly briefing on what the modern AI engineering stack costs, compiled from primary vendor pricing and release pages with source-check dates in the frontmatter. This issue follows the August 2026 issue; there was no September issue. The October 2026 issue covers the frontier briefing, the pricing map, the coding stack, and the image stack.
Between issues, the LLM API Pricing Index re-syncs published per-1M-token rates daily, and the model changelog logs releases and deprecations by month.
A monthly snapshot can't tell you what your own stack spent yesterday. StackSpend shows per-provider AI spend with same-day anomaly alerts; setup is read-only.
Related reading
- The Latest in AI Image Generation, August 2026, the previous issue
- The Latest in LLMs, October 2026
- LLM Model Pricing in October 2026
- The Best Coding Models in October 2026
- The Hidden Cost of AI APIs
References
- OpenAI — API pricing (GPT Image token rates, Batch)
- OpenAI — Image generation guide (GPT-Image-2 per-image estimates, partial-image tokens)
- OpenAI — Introducing ChatGPT Images 2.5
- OpenAI Developers on X — GPT-Image-2.5 Flare and Sunburst
- OrcaRouter — GPT-Image-2.5 Flare vs Sunburst (resolution limits, positioning)
- Google — Gemini API pricing (Nano Banana 2.1, Nano Banana 2, Nano Banana Pro)
- Google — Gemini API deprecations (Imagen 4, gemini-3.1-flash-image, gemini-2.5-flash-image)
- Google DeepMind — Nano Banana 2.1 model card
- The Decoder — Nano Banana 2.1 generates better images for less money
- Arena — Text-to-image leaderboard
- Arena — Image edit leaderboard
- Microsoft AI — MAI-Image-2.6 launches at No. 2 on Arena
- Microsoft AI — Pushing the quality-cost frontier with MAI-Image-2.6
- RuntimeWire — MAI-Image-2.6 in Foundry at about 4 cents an image (token rates, Flash rates)
- Black Forest Labs — Release notes (FLUX 3 Image, FLUX 3 Video)
- Black Forest Labs — API pricing (FLUX.2, FLUX 3 Image, Kontext)
- xAI — Models and pricing (Grok Imagine Image 2.0)
- gradually.ai — Grok Imagine Image 2.0 (release date)
- Noqta — Meta launches Muse Image
- Atlas Cloud — Seedream 5.0 Pro price
- OrcaRouter — Qwen-Image-3.0 general availability
- AlternativeTo — Qwen-Image-2.1: 7B, native transparency, license change
- Recraft — API pricing
- UsagePricing — Ideogram price change, August 25, 2026
- Midjourney — Updates and alpha changelogs
- ThursdAI — September 2026 AI releases
