In-depth analysis · About 5 min read
MiniMax M3.1 Flash: separating agent scores, speed and API cost
MiniMax’s October 19 release reports stronger agent results and lower token prices, while model weights and Pro remain future releases.
- Source published
- Published here
- Source checked
MiniMax announced M3.1 Flash on October 19. The official release combines agent, coding, computer-use and security evaluations with a new API price table. It also distinguishes services available now from model weights and a Pro version promised later. These are vendor-reported results, not an AISubHub benchmark.
Listen to the news
Read this article aloud or keep listening to the news.
Preparing the reader…
Device voice
Speech is handled by your browser or device. AISubHub does not send the text to an audio API. Voice availability and playback with a locked screen vary by device.
Resuming may repeat a short phrase.
At a glance
- The published rates per million tokens are $0.15 for input, $0.60 for output and $0.015 for cached input. The separate claim of 70% lower overall agent-task cost is not a uniform discount on those three rates.
- MiniMax reports 73.7% on DeepSWE v1.1, 61.7 points on OSWorld V2, 29.6 points on ALE and 82% on CyberGym. These measure different tasks and should not be combined into one quality score.
What has actually become available
The release says M3.1 Flash can now be accessed through the MiniMax API, M Plan and MiniMax Code. It identifies the anonymously operated Space Bunny as an early version and says that version processed more than 200 trillion tokens. That usage statement comes from MiniMax; it does not reveal the mix of customers or independently establish reliability. Most importantly, the closing paragraphs put model weights and a more capable Pro version in the future. An API launch therefore cannot be described as a completed open-weight release, and an M Plan entry point does not establish free API credit.
What the agent and coding evaluations test
MiniMax describes AutomationBench-AA as testing enterprise workflows across SaaS systems, including sales, operations and finance. BankerToolBench, where it reports 76.2, concerns multi-step planning, document consistency, financial compliance and chained tool calls. DeepSWE v1.1, reported at 73.7%, tests software work requiring code understanding, feature development, bug fixes and verification through tests. Our reading is that these are useful categories for different workloads, not evidence that any enterprise process is ready to run unattended. Teams should identify the tools and acceptance tests corresponding to their own work before comparing headline scores.
Computer use and security require their own boundaries
The reported OSWorld V2 result is 61.7 points and the Agents’ Last Exam result is 29.6 points. MiniMax links these to computer operation, cross-application execution and professional tasks. It also reports 82% on CyberGym and describes internal expert evaluations in auditing, patching, penetration testing, malware analysis and reverse engineering. Those internal observations are not an external security certification. Our practical interpretation is to evaluate these capabilities in an authorized, contained environment, with approvals and recovery procedures appropriate to the task, rather than infer that a benchmark score removes operational risk.
Token prices and task cost answer different questions
The announcement’s flat dollar rates are per one million tokens: input $0.15, output $0.60 and cached input $0.015. Separately, MiniMax says overall usage cost for agentic tasks is 70% below its predecessor. The release does not turn that task-level comparison into a promise for every workload. Our analysis is that output length, cache eligibility, repeated tool calls and failed attempts can change the bill even with the same input. A fair pilot should record the applicable token categories and total cost per accepted task; these figures are usage prices, not the price of a monthly subscription.
Reading speed claims without promising response time
MiniMax attributes improved inference efficiency to more aggressive sparsity. It reports nearly three times faster online inference, output speeds reaching more than 200 tokens per second, and median, or P50, speeds consistently above 170 tokens per second. These are the vendor’s reported operating figures. They do not establish the delay before the first token, a tail-latency guarantee or an end-to-end agent completion time for a particular customer. Our suggested test records waiting time, generation time and tool execution separately under a fixed workload, instead of treating token throughput as the complete user experience.
A useful pilot checks delivery rather than the headline
A small evaluation could use representative tasks with expected outputs and a limited tool set, keeping the model version and scoring rules fixed. Record completion, failed attempts, review effort, total tokens and elapsed time, and compare only runs meeting the same acceptance standard. This is our evaluation advice, not a test performed for this article. The release makes a concrete case for investigating a cheaper agent model, but its broad capability claims still need workload-specific evidence. Procurement should also check actual account access and current terms rather than assume that API availability, a coding product and a subscription convey identical entitlements.
Why it matters
Developers can use the announcement to design a bounded comparison of completed tasks, latency and total spending. A lower token rate becomes useful only when the system meets the project’s acceptance criteria without disproportionate retries or manual repair.
Limits and availability
We read the full announcement but did not call the API, reproduce the benchmarks or verify paid-account entitlements. The article does not establish production latency, universal savings or the availability of downloadable weights.
Sources and evidence
Source published: Oct 19, 2026 · Source checked: Oct 20, 2026
AI-assisted editorial articles checked against linked primary sources. Vendor claims are attributed; inclusion is not endorsement.