# The Cheapest AI APIs in 2026: Real Prices for Text, Image, Video and Music

> Verified per-call prices for 2026's most-used AI models — from $0.0046 an image and $0.019 a second of video — and the four things that decide your real bill.

- **Published**: 2026-07-25
- **Tag**: Pricing
- **Read time**: 8 min

---

Ask a better question than "what's the cheapest AI API?" — because nobody buys an API. You buy a specific model doing a specific job, and the cheapest text model is irrelevant if your feature generates video.

So this post lists verified per-call prices for the models people actually ship in 2026, grouped by what you're generating, then explains the four things that move your bill more than the sticker price does.

All prices USD, checked 15 August 2026 (1 credit = $0.01). "Direct" is the model owner's own published rate. Percentages are against **standard rates**; buying credits in the $1,250 top-up pack (+10% bonus) adds about 3 more points, which is how the headline **up to 90%** figure is reached.

Savings vary a lot by model — from 5% to 89% — so compare the specific models you call. Every percentage below can be checked against the two price columns beside it.

# Cheapest image generation

| Model | Price per image | Direct | Saving |
|---|---|---|---|
| **Z-Image** | **$0.0046** | $0.005 | −8%, cheapest call here |
| **GPT Image 2 (1K)** | **$0.025** | $0.219 | **−89%** |
| FLUX.2 Pro (1K) | $0.0284 | $0.03 | −5% |
| Nano Banana 2 Lite | $0.03 | $0.0336 | −11% |
| Nano Banana | $0.0302 | $0.039 | −23% |
| Nano Banana 2 (1K) | $0.039 | $0.08 | −51% |
| Seedream 5 Pro (1K) | $0.0412 | $0.045 | −8% |
| GPT Image 2 (2K) | $0.045 | $0.234 | −81% |
| Nano Banana 2 (2K) | $0.059 | $0.12 | −51% |
| Nano Banana Pro (1K / 2K) | $0.059 | $0.15 | −61% |
| Nano Banana Pro (4K) | $0.059 | $0.20 | **−71%** |
| GPT Image 2 (4K) | $0.072 | $0.413 | −83% |
| Nano Banana 2 (4K) | $0.089 | $0.18 | −51% |

Two things worth noticing. **GPT Image 2 is 89% below its direct rate at 1K** — it used to be one of the most expensive image models and is now cheaper than most of the mid-tier. And **Nano Banana Pro is a flat $0.059 at every resolution**, which makes it cheaper than Nano Banana 2 at 4K and identical at 2K, so there is no reason to run Nano Banana 2 above 1K. More detail in [Nano Banana API pricing](/blog/nano-banana-api).

At $0.0046, [Z-Image](/models/z-image) is still around 5× cheaper than GPT Image 2, so draft-cheap / finish-expensive still holds — the gap is just far smaller than it was.

# Cheapest video generation

Per second of output.

| Model | Price | Direct | Saving |
|---|---|---|---|
| **Seedance 2.0 Mini (480p)** | **$0.019** | $0.0721 | **−74%** |
| **Grok Imagine 1.5 i2v (480p)** | **$0.0268** | $0.12 | **−78%** |
| Seedance 2.0 Mini (720p) | $0.0398 | $0.1547 | −74% |
| **Grok Imagine 1.5 i2v (720p)** | **$0.0502** | $0.225 | **−78%** |
| Kling 3.0 (720p) | $0.0795 | $0.084 | −5% |
| Volcengine lip sync (720p) | $0.0814 | — | — |
| Wan 2.7 (720p) | $0.0928 | $0.10 | −7% |
| Kling 3.0 (1080p) / V3 Turbo (720p) | $0.1042 | $0.112 | −7% |
| Kling V3 Turbo (1080p) | $0.1302 | $0.14 | −7% |
| Wan 2.7 (1080p) | $0.1392 | $0.15 | −7% |
| OmniHuman 1.5 (lip sync) | $0.1522 | $0.16 | −5% |
| Kling 3.0 (4K) | $0.3892 | $0.42 | −7% |

Per clip:

| Model | Price | Direct | Saving |
|---|---|---|---|
| Gemini Omni (4s 720p) | $0.31 | — | — |
| InfiniTalk lip sync (480p) | $0.108 | $0.20 | −46% |
| InfiniTalk lip sync (720p) | $0.2321 | $0.40 | −42% |
| **Veo 3.1 Fast (720p)** | **$0.5847** | $1.20 | **−51%** |
| Gemini Omni (4s 4K) | $0.73 | $1.8667 | −61% |
| Veo 3.1 (720p) | $1.7286 | $3.20 | −46% |
| Veo 3.1 (4K) | $2.5584 | $4.736 | −46% |

Video is where the discounts are largest and the absolute prices are smallest: a 6-second 480p clip on Seedance 2.0 Mini costs about **11 cents**. Two rules still apply — render at 720p unless you know you need more, and keep clips short. See [Gemini API pricing](/blog/gemini-api-pricing) for the per-clip maths.

# Cheapest text and LLM

Per million tokens, input / output.

| Model | Our price | Direct | Saving |
|---|---|---|---|
| **Gemini 2.5 Flash** | **$0.1574 / $1.3117** | $0.30 / $2.50 | −48% |
| Gemini 3 Flash | $0.2624 / $1.5741 | $0.50 / $3 | −48% |
| Claude Haiku 4.5 | $0.506 / $2.5673 | $1 / $5 | −49% |
| GPT-5.6 Luna | $0.5097 / $3.0583 | $1 / $6 | −49% |
| **GPT-5.6 Terra** | **$0.65 / $3.90** | $2.50 / $15 | **−74%** |
| Gemini 2.5 Pro | $0.6596 / $5.247 | $1.25 / $10 | −47% |
| Gemini 3.6 Flash | $0.7871 / $3.9353 | $1.50 / $7.50 | −48% |
| Gemini 3.1 Pro / 3 Pro | $0.9745 / $6.2215 | $2 / $12 | −51% / −48% |
| GPT-5.4 | $1.2743 / $8.6944 | $2.50 / $15 | −49% / −42% |
| **GPT-5.6 Sol** | **$1.35 / $8.10** | $5 / $30 | **−73%** |
| Claude Sonnet 5 | $1.5367 / $7.702 | $3 / $15 | −49% |
| **Claude Opus 5 / Opus 4.8** | **$2 / $10** | $5 / $25 | **−60%** |
| GPT-5.5 | $2.5486 / $15.2916 | $5 / $30 | −49% |
| Claude Opus 4.7 / 4.6 / 4.5 | $2.5673 / $12.8554 | $5 / $25 | −49% |
| **Claude Fable 5** | **$4 / $20** | $10 / $50 | **−60%** |

Two things changed recently and are worth acting on. **GPT-5.6 Terra at $0.65 in / $3.90 out is now cheaper than Gemini 3.1 Pro** while sitting a tier above it on capability — it's the best value in the frontier band. And **Claude Opus 5 and Opus 4.8 are now cheaper than Opus 4.7** ($2 vs $2.5673 input), reversing what was true a few weeks ago, so there is no longer a reason to stay on the older revision. Full breakdown in [Claude API pricing](/blog/claude-api-pricing).

# Cheapest music and audio

| Model | Price |
|---|---|
| Gemini 3.1 Flash TTS | $0.0257 per call |
| Gemini Omni Audio | $0.0642 per generation |
| **Suno — full song** | **$0.1321** per generation |

[Suno](/models/suno) has no official public API, which makes per-generation price the only real comparison point. See [the Suno API guide](/blog/suno-api-guide).

# The four things that actually decide your bill

Sticker price is where most people stop. These four move the number more.

1. **Do you pay for failures?** On most providers a failed, errored or empty generation is still billed. Here it's refunded automatically — which matters most on video, where one failed clip can cost more than a hundred image calls. ([Why we built it that way.](/blog/never-pay-for-failed-jobs))
2. **Output length and resolution.** LLM output is around 5× input. On image and video, resolution is the dial — though note Nano Banana Pro now costs the same at 1K, 2K and 4K, so check per model before assuming.
3. **Tier discipline.** Cheap model for drafts, expensive model for the artifact the user keeps.
4. **Expiring credits and minimums.** Prepaid balances that expire, seat fees and monthly minimums never show up in a per-token table. Our credits don't expire and there are no seats: $5 gets you 500 credits to test with, and the larger packs add bonus credits (+5% at $500, +10% at $1,250).

# The cheap-stack pattern

- **Route by difficulty.** Flash and Luna-class models for the 80% that's routing, extraction and classification; premium models for the rest.
- **Draft cheap, finish expensive.** Z-Image for iteration, GPT Image 2 or Nano Banana Pro for the final render.
- **Measure per feature.** Every create response returns `creditsCharged` — sum it per feature and you have unit economics without building a metering pipeline.
- **Keep a fallback key.** Route primary traffic here, fall back to the direct API on error. You capture the savings on every successful call without making a gateway a single point of failure.

All of it runs off one endpoint and one key, so model choice is a string:

```http
POST https://you.bot/api/v1/generate
Authorization: Bearer YOUR_API_KEY

{ "modelId": "z-image", "input": { "prompt": "..." } }
```

Change `modelId` to `gpt-image-2-text-to-image`, `claude-sonnet-5`, `kling-3-0` or `suno` and everything else stays the same. More on why in [One key, every model](/blog/one-key-every-model).

# Cheaper than going direct, per model

- [OpenRouter alternatives](/blog/vs-openrouter) — for teams already on a text router
- [Replicate alternatives](/blog/vs-replicate) — for per-second media pricing
- [fal.ai alternatives](/blog/vs-fal) — for video-heavy workloads
- [Claude API pricing](/blog/claude-api-pricing) · [Gemini API pricing](/blog/gemini-api-pricing) · [Nano Banana API](/blog/nano-banana-api) · [the Suno API](/blog/suno-api-guide)

Or browse [all models with live prices](/pricing) and the [full catalog](/market).

# FAQ

**What is the cheapest AI API in 2026?**
It depends on modality. The cheapest single call here is Z-Image image generation at $0.0046. For video it's Seedance 2.0 Mini at $0.019 per second at 480p. For text it's Gemini 2.5 Flash at $0.1574 per million input tokens. Prices checked 15 August 2026.

**Is there a cheaper way to use GPT, Claude and Gemini than going direct?**
Yes — the same models are available through a gateway below the direct rate. The largest gap is GPT Image 2 at 89% below its direct price, and GPT-5.6 Terra at 74% on text. With the $1,250 top-up pack the headline figure reaches 90%.

**What's the cheapest LLM API for high volume?**
Gemini 2.5 Flash at $0.1574 / $1.3117 per million tokens. If you need a frontier model, GPT-5.6 Terra at $0.65 / $3.90 is the best value in that band.

**Do cheap AI APIs mean lower quality?**
Not when it's the identical model. A gateway fronting the same weights returns the same output; only the price and billing terms change.

**How do I avoid paying for failed generations?**
Choose a provider that refunds them. Here, failed, errored and empty generations are refunded automatically, so you only pay for output you receive.