AI Daily · Oct 20, 2026: MiniMax M3.1 Flash free across platforms, tested 88x cheaper than Opus 5.5
MiniMax M3.1 Flash is free for a limited time on OpenRouter, Cline, Ollama Cloud and Command Code; a developer test cost $0.014 vs $1.22 for Opus 5.5. DeepSeek open-sourced its DeepInfer inference stack.
Pick a date
-
Developer test finds M3.1 Flash 88x cheaper than Claude Opus 5.5
In a landing-page build test, MiniMax M3.1 Flash cost $0.014 versus $1.22 for Claude Opus 5.5, about 88x cheaper, while running 4x faster and using fewer tokens.
-
MiniMax M3.1 Flash launches free on OpenRouter for 72 hours
MiniMax's final M3.1 Flash model, previously the stealth model Space Bunny that topped OpenRouter for 10 days and processed 56.7T tokens, is now live and free on OpenRouter for 72 hours; it is also free for a limited time in Cline, on Ollama Cloud (US-hosted, zero data retention, 1M-token context) and for all Command Code plans.
-
DeepSeek open-sources DeepInfer inference framework under Apache-2.0
DeepSeek released DeepInfer, the production inference stack behind DeepSeek-V4.1-Flash, under Apache-2.0, with optimized kernels, Attention DP/balanced CP, MoE EP with EPLB, DSpark verification and async scheduling; it also open-sourced deepseek-recipe for token encoding/decoding matching the official API.
-
Command Code offers $100 DeepSeek V4.1 Flash credits for 24 hours
Command Code is raising DeepSeek V4.1 Flash 1017 usage credits from $10 to $100 for the next 24 hours, after which the allowance returns to $80.
-
Inception AI's Mercury 3 diffusion model launches on OpenRouter
Mercury 3 generates over 1,000 tokens/s with a 1M context window, 4x that of Mercury 2.5, and Inception reports it beats GPT-6 Luna and Haiku 5.5 on IFBench, GPQA Diamond and SWE-Bench Atlas. Launch pricing is 50% off at $0.10/M input and $0.35/M output tokens.
-
Perplexity Computer runs fully local on Mac for Pro, Max, Enterprise
Perplexity Computer can now run the model and agent entirely on a Mac, using PPLX Pro (a post-trained Qwen 3.8) for local tasks without consuming Computer credits; it is available to Pro, Max and Enterprise subscribers on supported Apple silicon Macs with 36 GB unified memory recommended.
-
Tencent open-sources Hy Memory with 3B active model and quantized builds
Tencent released Hy Memory, a memory layer for chat, work and code that shrinks from 60 GB to 18 GB and runs on a single llama.cpp binary instead of vLLM; it scores 58.8% (Q6_K) and 56.5% (Q4_K_M) on the LoCoMo eval, with data kept on local disk.
-
Meta publishes technical report on 30B open-weight Muse Glimmer model
Meta released a technical report for Muse Glimmer, a 30B open-weight model for agentic tasks that runs at 4-bit precision on a single consumer GPU in a Mac or PC and leads similarly sized open models on tool use and multi-step planning.
-
Requesty adds per-key agent visibility and one-click leaked key revocation
Requesty's gateway now shows which agents (Claude Code, Cursor, Codex, curl) are using each API key, lets teams detect leaked keys and revoke them in one click.
-
Qoder launches $10/month Lite plan with 1,000 credits
Qoder Lite costs $10/month for 1,000 credits; everyday tasks run on Flash models at 0.1x, hard jobs route to Qwen3.8-Max, DeepSeek-V4-Pro or Kimi-K3, and first subscribers get 1,000 bonus credits. The plan includes the full agent harness, one shared credit pool across Desktop, IDE, JetBrains and CLI, and BYOK.