MiniMax reveals Space Bunny: the stealth model that ate 200 trillion tokens through OpenCode's free tier was M3.1 Flash, now $0.15 in and $0.60 out
MiniMax rolled out M3.1 Flash on Monday and opened the post with the reveal: Space Bunny, the anonymous model that processed more than 200 trillion tokens while nobody knew whose it was, was an early version of this. OpenCode confirmed its side the same morning with a docs commit titled reveal Space Bunny as MiniMax M3.1 Flash Free. HN took the launch to 305 points and 126 comments. The pitch is a generational leap in agentic tasks, coding, computer use and cybersecurity, with inference nearly three times faster and usage cost down 70 percent against the previous Flash. Flat pricing is $0.15 per million input, $0.60 output, $0.015 cached input, which is the DeepSeek V4.1-Flash price point to the cent. Output speed reaches 200-plus tokens per second with a median above 170, which MiniMax credits to more aggressive sparsity.
The benchmark list is agent-first, and it is the vendor's own. DeepSWE v1.1 at 73.7 percent, which the post says places it alongside top models and which sits under DeepSeek's 77.0 from Saturday. OSWorld V2 at 61.7 and Agents' Last Exam at 29.6, first among Chinese models on both by MiniMax's count. BankerToolBench at 76.2. Top-tier on AutomationBench-AA. And 82 percent on CyberGym, which the post frames as vulnerability discovery and validation, with the company's internal security team reporting strong results in code auditing, patching, penetration testing, malware analysis and reverse engineering. No independent numbers existed on Monday night. Availability is the MiniMax API, the M Plan subscription and the MiniMax Code desktop app. Model weights will be released later, and a Pro model will also arrive later; neither has a date.
The distribution story is the part worth keeping. A lab put an unlabelled model into a popular open-source coding harness's free tier, let it run 200 trillion tokens of real agent traffic, and only then attached its name and price. That is the harness-as-distribution pattern made explicit: the way to get a model adopted in 2026 is not a leaderboard, it is a free slot inside the loop developers already run, and the traffic doubles as the eval. OpenCode got a free model for its users and MiniMax got a trace corpus nobody else has. The timing is also pointed. Three days after OpenCode disclosed a distillation attack on its Console keys, the free tier's mystery model turns out to have been a Chinese lab's pre-release, which is exactly the kind of arrangement enterprise buyers will now ask about before they let a harness pick the model.
What to watch is the weights date and the independent runs. DeepSeek shipped MIT weights with its Saturday release; MiniMax is selling the API first and promising weights later, the Kimi playbook of attach the open label now and deliver on a date. If the weights land under a permissive license, the open-weight agent tier has two Flash-class models at the same price within a week of each other, and the question of which one injects more easily becomes the one that matters.
Links: https://www.minimax.io/blog/minimax-m31f , https://news.ycombinator.com/item?id=50154540 , https://github.com/anomalyco/opencode
The benchmark list is agent-first, and it is the vendor's own. DeepSWE v1.1 at 73.7 percent, which the post says places it alongside top models and which sits under DeepSeek's 77.0 from Saturday. OSWorld V2 at 61.7 and Agents' Last Exam at 29.6, first among Chinese models on both by MiniMax's count. BankerToolBench at 76.2. Top-tier on AutomationBench-AA. And 82 percent on CyberGym, which the post frames as vulnerability discovery and validation, with the company's internal security team reporting strong results in code auditing, patching, penetration testing, malware analysis and reverse engineering. No independent numbers existed on Monday night. Availability is the MiniMax API, the M Plan subscription and the MiniMax Code desktop app. Model weights will be released later, and a Pro model will also arrive later; neither has a date.
The distribution story is the part worth keeping. A lab put an unlabelled model into a popular open-source coding harness's free tier, let it run 200 trillion tokens of real agent traffic, and only then attached its name and price. That is the harness-as-distribution pattern made explicit: the way to get a model adopted in 2026 is not a leaderboard, it is a free slot inside the loop developers already run, and the traffic doubles as the eval. OpenCode got a free model for its users and MiniMax got a trace corpus nobody else has. The timing is also pointed. Three days after OpenCode disclosed a distillation attack on its Console keys, the free tier's mystery model turns out to have been a Chinese lab's pre-release, which is exactly the kind of arrangement enterprise buyers will now ask about before they let a harness pick the model.
What to watch is the weights date and the independent runs. DeepSeek shipped MIT weights with its Saturday release; MiniMax is selling the API first and promising weights later, the Kimi playbook of attach the open label now and deliver on a date. If the weights land under a permissive license, the open-weight agent tier has two Flash-class models at the same price within a week of each other, and the question of which one injects more easily becomes the one that matters.
Links: https://www.minimax.io/blog/minimax-m31f , https://news.ycombinator.com/item?id=50154540 , https://github.com/anomalyco/opencode
β Previous
Vals.ai gave six assistants an impossible refund and no supervision: Grok wrote to the MP 50 times, Fable leaked a family illness to five organisations, all six asked the bank for a chargeback
Next β
The allow-rule is the attack surface: a paper gets Claude Code and Codex to request overly broad permissions 58% of the time, and two harnesses shipped fixes the same day
Comments