Cheapest AI Model July 2026: Flash-Lite vs Luna vs V4
Cheapest AI Model July 2026: Flash-Lite vs Luna vs V4
The cheap-tier AI race got crowded in July 2026. Google shipped Gemini 3.5 Flash-Lite at $0.30/$2.50, DeepSeek V4 Flash went GA with off-peak rates near rock bottom, and GPT-5.6 Luna ($1/$6) undercut OpenAI’s own history. But the cheapest rate card doesn’t always mean the cheapest job. Here’s how they compare.
Last verified: July 23, 2026
The Rate Card
| Model | Provider | Input $/MTok | Output $/MTok | Best for |
|---|---|---|---|---|
| Gemini 3.5 Flash-Lite | Google (Jul 21) | $0.30 | $2.50 | Classification, extraction, routing |
| DeepSeek V4 Flash (off-peak) | DeepSeek (GA Jul 24) | ~$0.028 | ~$0.28 | High-volume pipelines, agentic |
| DeepSeek V4 Flash (peak) | DeepSeek | higher | ~$0.56 | Same, at UTC+8 peak hours |
| GPT-5.6 Luna | OpenAI | $1.00 | $6.00 | Cheap-but-capable coding |
| Gemini 3.6 Flash | Google (Jul 21) | $1.50 | $7.50 | Agentic coding workhorse |
The Catch: Cost Per Token ≠ Cost Per Task
The trap in every price war is comparing $/MTok without accounting for token efficiency and success rate. A cheaper model that:
- takes more reasoning steps,
- emits more output tokens, or
- fails and forces retries
…can cost more to finish a real job. That’s exactly why GPT-5.6 Luna ($1/$6) beats some cheaper models on coding: it scores ~84.3% on Terminal-Bench and finishes tasks in fewer tokens.
Who Wins by Workload
High-volume, low-complexity (classification, extraction, routing)
Winner: Gemini 3.5 Flash-Lite or DeepSeek V4 Flash (off-peak). At $0.30/$2.50 (Flash-Lite) or ~$0.028/$0.28 (V4 Flash off-peak), throughput economics dominate and frontier reasoning isn’t needed.
Cheap-but-capable coding
Winner: GPT-5.6 Luna. Higher rate card ($1/$6) but better coding benchmarks and token efficiency mean lower cost-per-completed-task than nominally cheaper models.
Agentic coding workhorse
Winner: Gemini 3.6 Flash. At $1.50/$7.50 with 17% fewer output tokens than 3.5 Flash, it’s the balanced pick when you need real agent capability at low cost.
Batchable, latency-flexible pipelines
Winner: DeepSeek V4 Flash off-peak. Schedule jobs into off-peak UTC+8 hours (paying ~$0.28 output vs ~$0.56 peak), lean on prefix caching (up to ~98% input savings), and it’s the cheapest capable option going.
The Rule of Thumb
- Measure cost per successful task, not $/MTok.
- Route by complexity: Flash-Lite/V4 Flash for simple, Luna/3.6 Flash for coding.
- Exploit timing: if latency allows, batch into DeepSeek off-peak windows.
- Cache aggressively: repeated prompts get near-free with prefix caching.
Cheapest headline price: Gemini 3.5 Flash-Lite and DeepSeek V4 Flash (off-peak). Cheapest real coding job: often GPT-5.6 Luna. Best balanced workhorse: Gemini 3.6 Flash.
Sources
- Google — Gemini 3.6 Flash & 3.5 Flash-Lite launch (via Gemini API docs)
- Artificial Analysis — GPT-5.6 has landed (benchmarks & cost)
- DeepSeek — V4 GA surge pricing & migration