The Model That Ran 200 Trillion Tokens With No Name — MiniMax M3.1 Flash

As of Oct 21, 2026 · Checked directly against MiniMax's official announcement and each maker's official pricing page · Views -

Anyone who has handed work to an agent knows the first thing that comes to mind is the bill. Before asking how smart the model is, you ask what it costs to finish one job end to end. MiniMax's M3.1 Flash, released October 19, claims to answer exactly that question — and even the announcement was unusual. Before its name was public, this model had already processed more than 200 trillion tokens anonymously.

The short version

The anonymous model 'Space Bunny' was M3.1 Flash — it handled 200 trillion tokens of real traffic before getting a name

Flat pricing: $0.15 input · $0.60 output per 1M tokens. Slightly above the cheapest peers per token, but with no length cliff

Speed, cost savings, and benchmark scores are all MiniMax's own figures. No independent verification yet

What the announcement actually says

According to MiniMax's official post, the anonymously operated 'Space Bunny' was an early version of M3.1 Flash. It stacked up real usage behind a hidden name before receiving one — and that usage itself is the first piece of evidence the announcement leans on.

• M3.1 Flash is MiniMax's next-generation Flash model, pitched as a generational leap across agent work, coding, computer use, and cybersecurity (the official wording: "generational leap").
• Online inference is nearly three times faster than its predecessor, and total usage cost for agent work is 70% lower, the company says. Both figures are MiniMax's own; the announcement does not give measurement details.
• Everyday output speed is stated at 200+ tokens per second, with a median (P50) of 170+ tokens per second. Also the company's figure.
• Context is 1,000,000 tokens (confirmed in the official API docs) — sized for loading a large codebase or a bundle of long documents at once.
• It is available three ways: the MiniMax API, the subscription M Plan, and the MiniMax Code desktop coding tool.

How far to trust the benchmarks

Every score in the announcement comes from MiniMax. As of this writing, no independent body has re-measured the model under the same conditions.

• On AutomationBench-AA, a complex agent-work evaluation, it is said to rank among the top performers; on the finance evaluation BankerToolBench it scored 76.2.
• On DeepSWE v1.1, a real software-development evaluation, it recorded 73.7% — an evaluation that follows the long process of understanding code, building features, fixing bugs, and verifying with tests.
• On the computer-use evaluation OSWorld V2 it scored 61.7, and on Agents' Last Exam (ALE), which tests real professional work, 29.6 — first among Chinese models on both, the company says.
• On the cybersecurity evaluation CyberGym it recorded 82%, a vulnerability-discovery and validation figure.

The numbers look strong, but there is a fixed way to read them. Announcement scores are evidence for the company's claim that it can stand alongside top models — not a guarantee your work will see the same results. That a low absolute score like ALE's 29.6 is printed alongside is itself a signal: on hard, real work, failures are still frequent.

Prices side by side

Prices were checked directly against each maker's official pricing page and announcement (Oct 21, 2026). US dollars per 1M tokens, standard rates.

On the table alone, M3.1 Flash costs slightly more per token than Haiku 5.5 and Luna on both input and output. But MiniMax is not selling the per-token rate — it is selling total cost per job. Faster speed, fewer tokens per job, and fewer retries should pull the total bill down; the announcement does not say what job mix the 70% figure was measured on. The per-token prices are confirmed from official pricing; the per-job saving stays unconfirmed.

One more structural point. Haiku 5.5's rate jumps past a length threshold; M3.1 Flash's announced pricing has no such cliff. For agent work that feeds long inputs often, that structure can matter more to the final bill than the headline rate.

The jobs it wants

• Enterprise workflow automation: end-to-end handling across SaaS tools — sales, marketing, operations, customer support, finance, HR — named directly in the announcement.
• Software development: long, multi-step coding work — build the feature, verify it with tests — run at low cost.
• Computer use: looking at the screen and working across apps, with the OSWorld score offered as backing.
• Security assistance: code audits, vulnerability patching, penetration-test analysis, attached cheaply, alongside internal expert evaluations.

The common thread: this is not positioned as a ask-and-answer chatbot, but purely as an agent that calls tools and chains steps on its own.

What has already run in the field

• The biggest field case is the model itself. Under the anonymous name Space Bunny, it processed real service traffic past 200 trillion tokens, per the official announcement — a record from a live environment, not a lab benchmark, which is the distinction this launch keeps pushing.
• Before the formal launch, M3.1-Flash-Preview reportedly sat inside MiniMax Code from September 27 (second-hand reporting; the date does not appear in the official post).
• Independent speed measurement reportedly came in below the company's figures: a second-hand report (WinBuzzer) cites Artificial Analysis measuring 120.1 tokens per second streaming from the MiniMax API — worth setting next to the claimed P50 of 170+.

The cold water, and what is still unknown

• Every benchmark is the maker's own. AutomationBench, DeepSWE, OSWorld, CyberGym — all from MiniMax's post and its tables. Until independent verification lands, read them as a map of where the company is confident, not as rankings.
• The 70% cost saving has no published formula. Beyond "versus the predecessor," input/output mix and retry counts are undisclosed. Per-token math alone may not reproduce the saving; measure it on your own work.
• The weights are not out yet. The official post says only that "weights will be released later." An October 23 date circulates in second-hand reports and quotes of the official X post, but the blog itself carries no date — the release date stays unconfirmed.
• Parameter size is not confirmed in the official announcement. Figures like 428B total and ~23B active circulate in second-hand sources and are not stated as fact here — unconfirmed.
• There are no official Korean-language figures. How it performs on Korean agent work cannot be checked from the announcement materials — unconfirmed.

Sources

• MiniMax official blog (Oct 19, 2026 announcement — Space Bunny's 200 trillion tokens, flat $0.15/$0.60/$0.015 pricing, speed/cost/benchmark claims, weights "later")
• MiniMax official API docs (Supported Models table — MiniMax-M3.1-Flash context window 1,000,000 tokens)
• Claude official pricing (Haiku 5.5 — $0.10/$0.50 up to 100k tokens, $0.50/$2.50 above, cached input $0.01)
• OpenAI official pricing (GPT-6 Luna — standard $0.10/$0.50, cached input $0.01)

Not stated as fact in this article: the weights release date (the blog says only "later"); parameter size (428B / ~23B are second-hand figures); official Korean-language performance; the measurement conditions behind the 70% cost saving; whether GPT-6 Luna has length tiers (not found on the official pricing page on Oct 21, 2026). Prices and announcement figures were all checked against official sources on Oct 21, 2026.