Mercury 3 vs MiniMax M3.1 Flash vs DeepSeek V4.1 Flash
The short answer
Three cheap, fast models launched or updated within three days (October 17–19, 2026), and they suit different jobs.
- Mercury 3 (Inception) — the fastest: a diffusion LLM at ~1,068 tokens/s, $0.10 / $0.35 per million tokens at launch. For latency-bound voice, chat and browser agents.
- MiniMax M3.1 Flash — the multimodal agent workhorse: $0.15 / $0.60, image and video input, 200+ tokens/s, strong computer-use scores.
- DeepSeek V4.1 Flash-1017 — the open-weight coder: MIT weights, $0.15 / $0.60 off-peak, highest vendor-reported DeepSWE score of the three.
Model-vs-model scores below are each vendor’s own; independent evaluations were not yet complete on October 21, 2026.
The comparison
| Mercury 3 | MiniMax M3.1 Flash | DeepSeek V4.1 Flash-1017 | |
|---|---|---|---|
| Released | Oct 19, 2026 | Oct 19, 2026 | Oct 17, 2026 (update to Sep 10 model) |
| Architecture | Diffusion LLM | MoE (sparse), interleaved thinking | MoE, 552B total |
| Input / output price | $0.10 / $0.35 launch (list $0.20 / $0.70) | $0.15 / $0.60 | $0.15 / $0.60 off-peak · $0.30 / $1.20 peak |
| Cached input | 90% off input ($0.01 at launch) | $0.015 | $0.003 off-peak |
| Context / max output | 1M / 131K | 1M | 1M / 384K |
| Input modalities | Text | Text, image, video | Text (vision supported per DeepSeek) |
| Speed (vendor) | 1,068 tok/s | 200+ tok/s, P50 170+ | — |
| Weights | Closed | ”Later” | MIT, on Hugging Face |
| Headline benchmark (vendor) | “On par or better than GPT-6 Luna and Claude Haiku 5.5” | DeepSWE v1.1 73.7%, OSWorld V2 61.7, CyberGym 82% | DeepSWE v1.1 77.0%, Terminal-Bench 4.0 42.3% |
| Where | Inception API, OpenRouter, Baseten, Vercel | MiniMax API, M Plan, MiniMax Code | DeepSeek API (deepseek-flash), resellers at 2x |
Prices are per million tokens from each vendor’s page on October 21, 2026; see current API prices for the wider table.
Mercury 3 — when every millisecond counts
Inception says Mercury 3 is more than 40 times larger than Mercury 1 and the scaling laws hold for diffusion. Its partners use it for voice agents, tutoring and real-time search; one reports median generation of 0.9 seconds and end-to-end response of 2.2 seconds on a smart-glasses search flow. It is OpenAI-API compatible, supports tools and structured outputs, and new API accounts get 100 million free tokens. Weakness: text-only input, closed weights, and fewer independent coding results than the other two.
MiniMax M3.1 Flash — cheap agents that can see
MiniMax revealed that the anonymous “Space Bunny” model, which processed more than 200 trillion tokens, was an early M3.1 Flash. MiniMax claims 70% lower agent-task cost than its predecessor and nearly 3x faster inference, with 76.2 on BankerToolBench, 61.7 on OSWorld V2 and 82% on CyberGym. Image and video input make it the only one of the three suited to computer-use and screenshot-driven agents. Weights are promised but not yet out.
DeepSeek V4.1 Flash-1017 — open weights for coding agents
DeepSeek’s October 17 update keeps the price and raises its own DeepSWE v1.1 score to 77.0 (from 74.2) and Terminal-Bench 4.0 to 42.3 (from 31.2), with shorter outputs. Legacy deepseek-v4-flash names are now served by it. Watch the clock: peak hours double the price, and resellers such as Fireworks and Together charge the peak-level rate all day. New Search and Fetch APIs cost $0.75 per 1,000 calls each.
Which one to pick
- Voice or chat where latency is visible → Mercury 3.
- Agents that read screenshots, images or video → MiniMax M3.1 Flash.
- Coding agents, self-hosting or EU/US hosting of weights → DeepSeek V4.1 Flash.
- Batch jobs in US/EU hours → DeepSeek off-peak or Mercury at launch price.
- Need a known safety and support posture → compare with Claude Haiku 5.5 or GPT-6 Luna ($0.10 / $0.50) before switching.
Last verified: October 21, 2026.