GPT-6 Luna vs Gemini 3.8 Flash vs DeepSeek V4.1 Flash
The short answer
GPT-6 Luna is now the cheapest frontier-lab API at $0.10/$0.50 per million tokens, and it is the first time OpenAI has held that title. Released September 22, 2026, Luna halves GPT-5.6 Luna’s price, undercuts Gemini 3.8 Flash’s introductory rate by 7.5x, and matches DeepSeek V4.1 Flash’s off-peak input price while beating it on output. DeepSeek keeps two advantages: a $0.003 cache-hit price that nothing else approaches, and MIT-licensed weights you can self-host. Gemini 3.8 Flash keeps one: native audio and video input. For bulk text extraction and classification, route to Luna; for cache-heavy agent loops, DeepSeek off-peak; for multimodal input, Gemini.
Side by side
| GPT-6 Luna | Gemini 3.8 Flash | DeepSeek V4.1 Flash | |
|---|---|---|---|
| Released | September 22, 2026 | September 2, 2026 | September 10, 2026 |
| Input / output (per MTok) | $0.10 / $0.50 | $0.75 / $3.75 (intro) | $0.15 / $0.60 off-peak; $0.30 / $1.20 peak |
| Cached input | $0.01 | Standard Gemini caching discount | $0.003 off-peak; $0.006 peak |
| Price change ahead | None announced | Doubles to $1.50 / $7.50 on Jan 1, 2027 | Peak/off-peak by clock, every weekday |
| Context / max output | 1.05M / 128K | 1M / 64K | 1M / 384K |
| Long-prompt surcharge | >272K input: 2x in, 1.5x out | None | None |
| Modalities in | Text, image | Text, image, audio, video | Text (vision variant separate) |
| Knowledge cutoff | May 18, 2026 | March 2026 | Not stated |
| Open weights | No | No | Yes (MIT, Hugging Face) |
| Batch / off-peak discounts | Batch and Flex 50% | Batch and Flex 50%, Priority ~1.8x | Off-peak is the discount |
| Consumer access | Free and Go ChatGPT users in desktop app | Gemini app, AI Mode | DeepSeek app |
| Known benchmarks | OpenAI: “half as many mistakes” as GPT-5.6 Luna on factuality | OSWorld 2.0 59.0%; AA Index 58.7 | DeepSWE v1.1 74.2; AutomationBench 54.8; Terminal-Bench 4.0 31.2 |
Cost on a reference task
Using the standard 30K-input / 5K-output task (30 × input/1000 + 5 × output/1000):
| Model | Cost per task | Relative |
|---|---|---|
| GPT-6 Luna | $0.0055 | 1.0x |
| DeepSeek V4.1 Flash, off-peak | $0.0075 | 1.4x |
| DeepSeek V4.1 Flash, peak | $0.015 | 2.7x |
| Gemini 3.8 Flash, intro | $0.041 | 7.5x |
| Gemini 3.8 Flash, from Jan 2027 | $0.083 | 15x |
Two caveats change the ranking in practice. First, cache hits: an agent loop that resends a 50K-token system context every turn pays $0.0005 per turn on Luna ($0.01/MTok) but $0.00015 on DeepSeek off-peak, so DeepSeek wins on heavily cached workloads despite the higher headline. Second, token efficiency: Artificial Analysis measured Gemini 3.8 Flash at roughly 48,000 output tokens per Intelligence Index task in September 2026, one of the most verbose models it tracks. If Luna and DeepSeek finish tasks in fewer tokens, the per-task gap widens further; if they need more retries, it narrows. Measure your own workload.
What each is actually for
GPT-6 Luna is OpenAI’s stated “high-volume clerical” tier: extraction, classification, summarization, form-filling, and the long tail of document work where a 1.05M-token context and a $0.10 input price matter more than the last few points of reasoning. OpenAI’s headline capability claim is that it makes half as many factuality mistakes as GPT-5.6 Luna. The 272K-token repricing is the trap: a single 500K-token request bills at $0.20/$0.75 for the whole call, so chunk or cache instead of stuffing the window. See GPT-6 Sol vs Luna vs Astra for where Luna sits in OpenAI’s lineup.
Gemini 3.8 Flash is the multimodal workhorse. It is the only model of the three that takes audio and video natively, it scores 59.0% on OSWorld 2.0 (a computer-use benchmark the other two do not publish), and it runs inside Google Cloud’s compliance surface. You pay for that: 7.5x Luna now, 15x from January 1, 2027. The intro rate expiring is the single most important fact for anyone budgeting a Gemini deployment past year end. Details in Gemini 3.8 Flash vs 3.7 Flash vs Opus 5.
DeepSeek V4.1 Flash is the open-weight option and the cache king. Its Causal Encoder-Decoder architecture activates 8B parameters during prefill and 16B during decode from a 552B MoE, cuts KV cache to about 890 bytes per token, and supports a 384K max output, three times Luna’s. On DeepSeek’s own numbers it is competitive with frontier models on DeepSWE (74.2 vs 74.0 for Claude Opus 5) but far behind on Terminal-Bench 4.0 (31.2 vs 66.4 for Claude Opus 5.5). The peak window (Mon–Fri 01:00–04:00 and 06:00–10:00 UTC) lands on Asia-Pacific business hours; US and EU workloads mostly run off-peak. Full breakdown in DeepSeek V4.1 Flash vs V4 Pro vs Gemini 3.8 Flash.
Decision rule
- Bulk text, lowest bill, no self-hosting: GPT-6 Luna.
- Cache-heavy agent loops, or you want the weights: DeepSeek V4.1 Flash, scheduled off-peak.
- Audio or video input, Google Cloud residency, or computer use: Gemini 3.8 Flash, budgeted at the 2027 price.
- Anything needing real reasoning: none of these. Step up to GPT-6 Sol at $2/$10 or Claude Opus 5.5 at $4/$20, both released the same day as Luna.
The larger story is that the sub-$1 tier now includes OpenAI. Through mid-2026 the cheapest OpenAI model cost more than the cheapest Google or DeepSeek model; as of September 22, 2026 it costs less than either at list. Google has not yet responded, and Gemini 3.8 Flash’s scheduled doubling looks harder to execute every day it stands.
Last verified: September 24, 2026.