An anonymous model spent ten days as the most used model on OpenRouter, and the company behind it only admitted ownership after it had already been benchmarked to death.
On 17 October DeepSeek pushed out V4.1 Flash-1017, an update to the model Chinese developers call the big fat fish. Two days later MiniMax launched M3.1 Flash and claimed Space Bunny, the anonymous entry that had been attracting attention in developer communities.
On the published evaluations, M3.1 Flash ranks first in complex agent tasks on Automation-Bench, ties for first with Claude Opus 5 on BankerToolBench, and places second only to DeepSeek on coding cost-performance. OpenRouter disclosed that Space Bunny held the top spot for model usage on its platform for ten consecutive days and processed 56.7 trillion tokens. MiniMax says the model handled more than 200 trillion tokens across testing and use.
The test that matters more than a leaderboard
Leiphone ran a single practical task with the new model: hand it a hand-drawn floor plan and a building dimension table, and ask for an interactive three-dimensional campus built from scratch and running in a browser.
Both stacks delivered the main requirements. MiniMax Code with M3.1 Flash took 55.4 minutes. OpenCode with DeepSeek V4.1 Flash took about 121 minutes. Converted at the API pricing used in the test, both runs cost roughly RMB 6.5 in model fees.
Same bill, 65 minutes less. That gap is the whole product argument. Agent work is priced in wall-clock time as much as tokens, because someone is waiting for the output and a human is often supervising it.
Speed shows up in the invoices
A developer posting as Hassan on X set M3.1 Flash and Claude Opus 5.5 the same job, building a landing page for a hardware product launch. He preferred the MiniMax design, and the two bills came to USD 0.014 and USD 1.22 respectively. The MiniMax run also finished about four times faster.
A single developer test is not a benchmark, and Leiphone says as much. The more durable evidence is the Artificial Analysis index, where M3.1 Flash scored 43 against 29 for the previous M3, an improvement of close to 50 per cent, while the task-weighted average model cost fell from USD 0.51 to USD 0.22.
That combination is unusual. Scores rose and the cost of finishing a task more than halved, which is the opposite of how frontier model generations have usually been priced.
What the Flash tier is actually for
DeepSeek and MiniMax are now running a two-week cadence on the cheap end of the market, and the competitive axis has moved. The question is no longer which Chinese model is closest to the American frontier. It is which one finishes an agent task faster at a price that makes high-volume automation viable.
Two days separate these releases. Ten days of platform telemetry decided which one developers actually reached for, and neither number appears in a benchmark table.
The model that wins the agent era will not be the smartest one in the room. It will be the one cheap enough to leave running.
Editor’s note: This is an adapted translation of the original Leiphone report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.leiphone.com/category/yanxishe/IMOE7FfuhC0qGxZk.html.