MiniMax M3.1 Flash: Pricing, Benchmarks and Mac Setup

MiniMax released M3.1 Flash into general availability on October 19, 2026 under the production model identifier MiniMax-M3.1-Flash across the MiniMax pay-as-you-go API, the M Plan subscription, and MiniMax Code ¹². Published API list pricing sits at a flat $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached input tokens across the entire 1,000,000-token context window ³. The rollout simultaneously initiates the retirement of the M2.5 and M2.7 model families, whose cloud endpoints shut down on November 18, 2026 ⁴. October 23, 2026 update: MiniMax has now released the full 428B/23B MoE weights on Hugging Face. The checkpoint is roughly 252 GB, with an early MLX Text 4bit conversion also available. Self-hosting is possible in principle but remains impractical on typical Macs because of the memory footprint ⁵⁶⁷.

Technical Specifications and Provider Identifiers

With the transition from the September 27, 2026 preview build to general availability on October 19, 2026, MiniMax retired the MiniMax-M3.1-Flash-Preview identifier ²⁸. While M Plan routes legacy preview calls to MiniMax-M3.1-Flash automatically, direct pay-as-you-go API integrations and local CLI configurations should be updated explicitly to the production string ⁸. Developers routing traffic through third-party aggregators must use provider-specific strings; OpenRouter lists its rate-limited free endpoint under minimax/minimax-m3.1-flash:free ⁹.

Specification	Production Status (Verified October 23, 2026)	Source
Official Model Name	MiniMax M3.1 Flash	1
Native MiniMax API ID	MiniMax-M3.1-Flash (replaces MiniMax-M3.1-Flash-Preview)	28
OpenRouter Model ID	minimax/minimax-m3.1-flash:free (subject to free-tier rate limits)	9
General Availability Date	October 19, 2026 (previewed September 27, 2026)	1
Context Window	1,000,000 tokens combined (input + output; max output ~524K)	210
Supported Modalities	Input: Text, Image, Video · Output: Text	211
Media Payload Limits	Images up to 10 MB; Videos up to 50 MB (URL/Base64) or 512 MB (Files API); Request body up to 64 MB	11
Reasoning Mode	Always active (Interleaved Thinking); cannot be disabled (HTTP 400, error 2013)	1112
Thinking Depth (effort)	low, medium, high, xhigh, max (defaults to max when omitted)	211
Vendor-Reported Speed	200+ tokens/s peak, 170+ tokens/s median (P50)	1
Pay-as-you-go Rate Limits	200 requests per minute (RPM), 10,000,000 tokens per minute (TPM)	13
Priority Service Tier	Not supported (service_tier field unavailable on M3.1 Flash)	11
Official Open Weights	Released October 23, 2026; 48 shards, DSpark draft head, ~252 GB repository	5
Parameters & Formats	~428B total / ~23B active MoE; NVFP4 and MXFP8	514
License	MiniMax Community License with conditional commercial use	6
MLX	Third-party Text 4bit checkpoint exists, no in-house runtime test	7

Prior to the October 19 announcement, an early build of M3.1 Flash operated anonymously on public routing platforms under the stealth alias Space Bunny, processing more than 200 trillion tokens according to MiniMax ¹. Architecturally, the model uses aggressive sparsity to nearly triple online inference throughput compared with its predecessor ¹. The 1,000,000-token context window caps the combined sum of input, reasoning, and final output tokens within a single API call ²¹¹.

For multimodal inputs on the Anthropic-compatible interface, the model accepts JPEG, PNG, GIF, and WEBP images alongside MP4, AVI, MOV, and MKV video streams ¹¹. Single-image token consumption scales with resolution and the selected detail level: low consumes up to roughly 600 tokens, default typically uses 1,000 to 5,000 tokens, and high can exceed 15,000 tokens ¹¹. Calling POST /anthropic/v1/messages/count_tokens returns exact input token counts prior to generation ¹¹.

Architectural Changes Compared with MiniMax M3

While MiniMax-M3 and MiniMax-M3.1-Flash share a 1,000,000-token context window and multimodal image/video ingestion, M3.1 Flash alters two core mechanics: reasoning enforcement and long-context pricing ²³¹¹.

On MiniMax-M3, thinking blocks are off by default and must be opted into via thinking: {"type": "adaptive"} ¹¹. On MiniMax-M3.1-Flash, adaptive reasoning is permanently locked on ¹¹. The model performs interleaved thinking prior to every tool invocation, evaluating environment state and previous tool outputs between execution steps ¹². Developers tune reasoning depth across five levels (low, medium, high, xhigh, and max). Omitting the effort parameter defaults the request to max, maximizing multi-step accuracy while increasing latency and billed output tokens ²¹¹.

Additionally, MiniMax-M3.1-Flash omits support for the service_tier: "priority" parameter, which on MiniMax-M3 provides priority queue admission at 1.5 times the standard per-token rate ³¹¹.

API Pricing, Prompt Caching, and Cost Scenarios

On the pay-as-you-go schedule, M3.1 Flash halves standard M3 input and output rates for prompts up to 512,000 input tokens and cuts cached input reads by 75 percent ³. Above 512,000 input tokens, M3 doubles its per-token rates across all three categories, whereas M3.1 Flash maintains a flat rate up to the full 1,000,000-token limit ³.

Model and Service Tier	Input / 1M Tokens	Output (incl. Thinking) / 1M Tokens	Cache Read / 1M Tokens	Separate Cache Write Rate
MiniMax-M3.1-Flash (Full 1M Context)	$0.15	$0.60	$0.015	$0.00 (billed as standard input)
MiniMax-M3 Standard (<= 512K Input)	$0.30	$1.20	$0.060	See API pricing schedule
MiniMax-M3 Standard (> 512K Input)	$0.60	$2.40	$0.120	See API pricing schedule
MiniMax-M3 Priority (<= 512K Input)	$0.45	$1.80	$0.090	See API pricing schedule
MiniMax-M2.7 Legacy (Until Nov 18, 2026)	$0.30	$1.20	$0.060	$0.375

U.S. Billing Note: All pay-as-you-go API rates are listed in U.S. dollars excluding applicable U.S. state and local sales taxes ³⁴.

Prompt caching activates on requests containing at least 512 input tokens, supporting both automatic prefix matching and explicit cache_control breakpoints on the Anthropic-compatible endpoint ¹⁵. Cache prefixes are constructed strictly in the hierarchy tools -> system -> messages and carry a 5-minute time-to-live (TTL) that refreshes automatically at no extra charge whenever hit ¹⁵. Keeping tool definitions and system prompts static at the start of the payload reduces the rate on cached prefix tokens from $0.15 to $0.015 per million tokens ³¹⁵.

Worked API Cost Scenarios

Total API charges for text and coding requests follow the standard billing formula:

Total = fresh_input_M * rate_input + cache_write_M * rate_write + cache_read_M * rate_read + output_M * rate_output

Under automatic prompt caching on M3.1 Flash, newly cached prefix tokens are billed at the standard uncached input rate ($0.15/1M, so rate_write = 0 as a separate surcharge), while all generated reasoning tokens count toward output_M at $0.60/1M ³¹⁵. Excluding optional server-side tools (web_search at $0.01/call and API-vlm at $0.01/call) and U.S. sales tax, three aggregated monthly usage tiers calculate as follows ³:

Workload Tier	Fresh Input ($0.15/M)	Output incl. Thinking ($0.60/M)	Cache Read ($0.015/M)	Total API Cost (USD Net)
Occasional (Targeted bug triage)	2M ($0.30)	0.5M ($0.30)	1M ($0.015)	$0.615
Regular (Daily coding agent)	20M ($3.00)	5.0M ($3.00)	10M ($0.150)	$6.150
Intensive (Multi-step repo automation)	200M ($30.00)	50.0M ($30.00)	100M ($1.500)	$61.500

For an identical illustrative token mix of 10 million fresh input tokens, 10 million cache-read tokens, and 2 million output tokens, MiniMax-M3.1-Flash costs $2.85 ($1.50 + $0.15 + $1.20). The same token volume on MiniMax-M3 costs $6.00 ($3.00 + $0.60 + $2.40) when every request stays at or below 512K input tokens, and $12.00 above 512K input tokens — a 52.5 percent and 76.25 percent reduction in per-token charges, respectively ³.

Fixed-token comparisons remain illustrative price checks rather than end-to-end productivity measurements. MiniMax separately claims a 70 percent reduction in overall agent-task cost compared with its predecessor, which depends on how quickly the model converges on a working patch and which effort setting is used ¹. Furthermore, MiniMax’s own ProgramBench documentation for Codex notes that capping a tool’s context window at 512,000 tokens (triggering earlier history compaction) achieved comparable task accuracy to the full 1M window while cutting total session costs by roughly 20 percent ¹⁶.

M2.5 and M2.7 Retirement Schedule and Migration Risks

Alongside the M3.1 Flash release, MiniMax announced the retirement of the MiniMax-M2.5 and MiniMax-M2.7 model families, including their -highspeed variants ⁴. Both families closed to new accounts on October 19, 2026, and their cloud API endpoints will shut down permanently on November 18, 2026 ⁴. Background on the legacy architecture and self-hosted weights remains available in our MiniMax M2.7 on Mac guide.

Migration rules depend on how requests are authenticated:

1. M Plan and Legacy Token Plan: Starting October 19, 2026, requests authenticated with a Subscription Key that specify M2.5 or M2.7 model IDs are automatically routed to MiniMax-M3.1-Flash with unchanged plan allowances ⁴.
2. Pay-as-you-go API: Existing accounts using standard Open Platform API keys can continue calling MiniMax-M2.7 and MiniMax-M2.5 until November 18, 2026, after which requests to those model IDs will fail ⁴.
3. Self-Hosted Open Weights: Previously downloaded MiniMax-M2.7 weights running on private GPU servers are unaffected by the cloud endpoint shutdown.

Automatic routing on subscription keys introduces three breaking behavioral changes for existing M2.7 agent code:

• HTTP 400 on Disabled Thinking: Clients configured to send thinking: {"type": "disabled"} or reasoning_effort: "none" immediately fail with HTTP 400 (error code 2013) when routed to M3.1 Flash ²¹¹.
• Truncated Outputs under Tight max_tokens Caps: Because M3.1 Flash defaults to effort: "max" and counts thinking tokens toward max_tokens, legacy requests with small output caps (such as 512 or 1,024 tokens) can exhaust the entire token budget inside the reasoning block and terminate with stop_reason: "max_tokens" before emitting text or tool calls ¹¹.
• Lower Pay-as-you-go Rate Limits: Migrating pay-as-you-go workloads from M2.7 (500 RPM and 20M TPM) to M3.1 Flash (200 RPM and 10M TPM) reduces per-minute request and token ceilings by 60 percent and 50 percent, respectively ¹³.

M Plan Subscriptions vs. Pay-as-You-Go API Billing

MiniMax separates pay-as-you-go Open Platform accounts from its individual M Plan subscriptions (Go, Explore, and Build), which replaced the legacy Token Plan (Plus, Max, Ultra) for new sign-ups ⁸¹⁷⁴. Subscription Keys and pay-as-you-go API keys draw from completely separate balances and cannot be used interchangeably ¹⁸⁴. For a broader look at competing developer plans, see our AI subscription comparison.

Billing Option	Standard U.S. List Price (USD Net)	Included Models & Quota Mechanics
M Plan Go	$22/month ($220/year)	MiniMax-M3.1-Flash, image & audio models, web search, MiniMax Code, Subscription Key; no H3 video (1x baseline usage) 84
M Plan Explore	$55/month ($550/year)	All Go features plus MiniMax-H3 video generation; 3x relative usage allowance 84
M Plan Build	$132/month ($1,320/year)	All Explore features plus peak-hour priority access; 7.5x relative usage allowance 84
Legacy Token Plan	$20 / $50 / $120 per month	Closed to new purchases; grandfathered at original price and terms while auto-renewal remains enabled 17
Credit Packs (Add-on)	$5 / $25 / $100 (5k / 25k / 100k credits)	Valid for 1 year from purchase; automatically covers overages once an M Plan usage window is exhausted 184
Pay-as-you-go API	Billed per token ($0.15 / $0.60)	Prepaid Open Platform balance for production workloads; never charged by a Subscription Key 318

An M Plan subscription is not an unlimited flat rate for autonomous agents. Every connected client — including MiniMax Code, Claude Code, Codex, Cursor, OpenCode, OpenClaw, and Hermes Agent — draws from a single shared pool governed by both a rolling 5-hour window and a 7-day weekly window ⁸¹⁸. A 5-hour window reset does not restore a depleted weekly allowance ¹⁸. Even when quota remains available, high concurrency or burst traffic can trigger dynamic rate limiting (HTTP 429, error code 1002) ¹⁸.

Grandfathered Token Plan subscribers retain their legacy pricing and perks (such as historical 150% weekly limits) only while automatic renewal stays active; upgrading to M Plan applies a prorated credit for the remaining billing cycle but permanently forfeits the ability to return to Token Plan ¹⁷.

Mac users can inspect their remaining M Plan allowance directly from the terminal via the MiniMax CLI or REST endpoint ⁴:

# Option 1: Query via the official MiniMax CLI
mmx quota

# Option 2: Direct HTTP check using an M Plan Subscription Key
curl --fail-with-body -sS "https://www.minimax.io/v1/token_plan/remains" \
  -H "Authorization: Bearer ${MINIMAX_SUBSCRIPTION_KEY}" \
  -H "Content-Type: application/json"

Configuring Hermes Agent, Claude Code, and Cursor on macOS

Because MiniMax exposes both an Anthropic-compatible endpoint (https://api.minimax.io/anthropic) and an OpenAI-compatible endpoint (https://api.minimax.io/v1), MiniMax-M3.1-Flash connects directly to standard macOS coding clients without local translation proxies ²¹¹. MiniMax recommends the Anthropic-compatible interface for agentic workflows so that clients receive structured thinking blocks and preserve interleaved reasoning across multi-turn tool calls ²¹².

To connect Nous Research’s open-source Hermes Agent on macOS, MiniMax documents three setup paths ¹⁹:

# 1. Install Hermes Agent on macOS and verify environment health
curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash
hermes doctor

# 2a. One-click setup wizard via MiniMax CLI (writes ~/.hermes/config.yaml and ~/.hermes/.env)
npx -y mmx-cli@latest agent setup

# 2b. Alternative interactive selector (OAuth sign-in or Subscription Key on global endpoint)
hermes model

After selecting the model, edit ~/.hermes/config.yaml under the model: section to set context_length: 1000000 (or 512000 to trigger earlier context compaction and reduce token usage) ¹⁹¹⁶. When operating Hermes Agent, OpenClaw, or Claude Code against M3.1 Flash, never invoke /think off or set reasoning effort to none or minimal; use /reasoning to switch between low, medium, high, xhigh, and max ¹⁹.

Minimal Python API Connectivity Example

The Python script below demonstrates API connectivity, explicit thinking-depth configuration (output_config.effort), and separate extraction of reasoning blocks, text output, and prompt-cache counters via the Anthropic-compatible endpoint ²¹¹¹⁵. Without attached file-system tools or test runners, this minimal call verifies endpoint reachability rather than auditing a repository.

Install the Anthropic SDK on macOS (python3 -m pip install anthropic) and export your API key (export MINIMAX_API_KEY="your_key"):

import os
import anthropic

api_key = os.getenv("MINIMAX_API_KEY")
if not api_key:
    raise SystemExit("Environment variable MINIMAX_API_KEY is missing. No request sent.")

client = anthropic.Anthropic(
    base_url="https://api.minimax.io/anthropic",
    api_key=api_key,
)

response = client.messages.create(
    model="MiniMax-M3.1-Flash",
    max_tokens=4096,
    output_config={"effort": "medium"},
    system="Provide concise, technically accurate answers.",
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "type": "text",
                    "text": "List three breaking changes when migrating an agent from MiniMax-M2.7 to MiniMax-M3.1-Flash.",
                }
            ],
        }
    ],
)

for block in response.content:
    if block.type == "thinking":
        print(f"[Reasoning Block]\n{block.thinking}\n")
    elif block.type == "text":
        print(f"[Response Text]\n{block.text}\n")

print(
    "Token Usage:",
    {
        "fresh_input_tokens": getattr(response.usage, "input_tokens", 0),
        "cache_read_tokens": getattr(response.usage, "cache_read_input_tokens", 0),
        "cache_write_tokens": getattr(response.usage, "cache_creation_input_tokens", 0),
        "output_tokens_incl_thinking": getattr(response.usage, "output_tokens", 0),
    },
)

In multi-turn tool-calling loops, append the entire response.content array — including all thinking and tool_use blocks — back into the message history unmodified on subsequent turns to maintain reasoning continuity ¹¹¹². On the OpenAI-compatible interface (https://api.minimax.io/v1), pass reasoning_effort instead of output_config.effort and read the reasoning trace from reasoning_content ²¹².

Vendor Benchmarks and Harness Context

In its October 19, 2026 release documentation, MiniMax published benchmark comparisons pitting M3.1 Flash against DeepSeek V4.1 Flash-1017, DeepSeek V4.1 Flash, Kimi K3, GPT-5.6 Sol, and Claude Opus 5. Values were cross-checked on October 23 against the newly available NVIDIA model card ¹¹⁰¹⁴. All scores originate from MiniMax’s launch report and scorecard footnotes, which specify varying evaluation harnesses (including Claude Code 2.1.263 / 2.1.285, OpenCode 1.17.18, and mini-SWE-agent) and context windows ranging from 200K to 1M tokens ¹¹⁰.

Benchmark	Task Domain & Documented Evaluation Setup	M3.1 Flash	DeepSeek V4.1 Flash-1017	GPT-5.6 Sol	Claude Opus 5
Automation-Bench (v1.0.6)	Multi-step enterprise SaaS workflow automation	62.8	60.5	45.8	50.3
BankerToolBench	Financial compliance & chained tool calls (OpenCode 1.17.18, M2.7 judge)	76.2	n/a	74.6	76.2
DeepSWE v1.1	Software engineering & issue patching (mini-SWE-agent, effort: max)	73.7%	77.0%	73.0%	74.0%
Terminal-Bench 4.0	Command-line execution & systems administration	38.1%	42.3%	39.9%	51.8%
NL2Repo-Bench	End-to-end repository generation from specifications	61.2%	68.2%	56.8%	75.3%
OSWorld 2.0	Cross-application desktop computer operation	61.7	n/a	65.7	70.6
Agents’ Last Exam (ALE)	Complex professional tasks with economic value	29.6	n/a	30.6	32.2
CyberGym	Vulnerability discovery & security patching (Claude Code 2.1.263, 1M)	82.2%	n/a	83.6%	n/a
HLE (Text-only, no tools)	Academic reasoning (Humanity’s Last Exam)	43.4%	42.6%	49.5%	55.6%
BabyVision (with tools)	Visual reasoning (Claude Code 2.1.285, 1,500 px image side)	91.0%	92.5%	n/a	n/a

October 23 source-version note: The NVIDIA model card reports 62.8 on Automation-Bench and 92.5% for DeepSeek V4.1 Flash-1017 on BabyVision. Earlier launch-day tables sometimes displayed 62.2 and 93.0% respectively. We use the currently accessible NVIDIA release figures here; these remain vendor-reported, not independently reproduced ¹⁴.

Two technical conclusions emerge from the vendor dataset:

1. Competitive Tool Chaining and Issue Patching: M3.1 Flash leads or ties the vendor table on structured business and financial tool workflows (Automation-Bench at 62.8 and BankerToolBench at 76.2). On DeepSWE v1.1, scaling reasoning effort from high (68.1%) and xhigh (72.1%) to max (73.7%) brings the model within 0.3 percentage points of Claude Opus 5 (74.0%) and 3.3 points of DeepSeek V4.1 Flash-1017 (77.0%) at a fraction of frontier token pricing ¹¹⁰.
2. Performance Gaps on Terminal and Whole-Repo Tasks: When tasks require sustained shell exploration (Terminal-Bench 4.0: 38.1% vs. 51.8% for Claude Opus 5) or building an entire repository from scratch (NL2Repo-Bench: 61.2% vs. 75.3%), M3.1 Flash trails frontier models by 13 to 14 percentage points ¹¹⁰. Similarly, MiniMax’s claim of ranking first on OSWorld 2.0 (61.7) and Agents' Last Exam (29.6) applies specifically among Chinese models, whereas Claude Opus 5 (70.6 and 32.2) and GPT-5.6 Sol (65.7 and 30.6) post higher overall scores ¹.

Official Open Weights, MLX on Apple Silicon, and Privacy

October 23, 2026 update: MiniMax has released the official M3.1 Flash checkpoint at MiniMaxAI/MiniMax-M3.1-Flash. The public repository includes 48 Safetensors shards, configuration and tokenizer files, plus a DSpark speculative-decoding draft head, totaling approximately 252 GB. Its mixture-of-experts network comprises approximately 428 billion total parameters with 23 billion active per token. Claims that M3.1 Flash remains a cloud-only model are now outdated ⁵¹⁴.

Architecture and actual hardware requirements

Routed experts use NVFP4 (W4A4) quantization, while other components use formats including MXFP8. M3.1 extends sparse attention and uses DSpark speculative decoding rather than the older MTP path. 23B active parameters describes compute per token, not model size in memory ⁵¹⁴.

At an idealized four bits per parameter, 428B parameters would require 214 billion bytes (about 199 GiB) for values alone, before mixed-precision components, metadata, KV cache, inference buffers and macOS. This is a lower-bound illustration, not a measured runtime requirement. Even 256 GB Unified Memory does not guarantee a functioning deployment or a usable 1M context. Typical 16–128 GB Macs cannot reasonably keep the full model in memory.

Download, MLX conversion, and serving

Download the official checkpoint with the Hugging Face CLI:

hf download MiniMaxAI/MiniMax-M3.1-Flash --local-dir MiniMax-M3.1-Flash

Official documentation focuses on SGLang and vLLM on Linux/NVIDIA accelerator servers. Download availability is not proof of a working Apple Silicon deployment ⁵¹⁴.

Route	Verified as of October 23, 2026	Limitation
Official MiniMax weights	Public download, approximately 252 GB	Requires substantial memory; not a small 23B model
NVIDIA servers with SGLang / vLLM	Documented by MiniMax and NVIDIA	Hardware and usable context length depend on configuration
MLX Community Text 4bit	Early Apple Silicon third-party conversion	Not an official release; text-oriented, untested here, no confirmed full image/video parity
API / M Plan / OpenRouter	Convenient access from Mac	Hosted inference sends prompts and optional media off-device

MiniMax Community License and commercial restrictions

These weights use the MiniMax Community License, not MIT or Apache 2.0. Its initial grant covers noncommercial purposes. Commercial products/services must prominently display Built with MiniMax M3.1-Flash. For products/services generating up to $20 million annual revenue, a one-time notice to MiniMax is required; above $20 million, prior written permission is required. The appendix also prohibits specified uses. Read the original license before commercial deployment ⁶.

Data residency and smaller local Mac models

Actual self-hosting can retain source code and prompts within a controlled environment if telemetry and other data flows are checked. By contrast, MiniMax API, M Plan, MiniMax Code and cloud routers execute externally. Publishing weights alone does not establish ZDR, HIPAA or other compliance guarantees.

For Macs with 16–64 GB Unified Memory, smaller quantized alternatives such as qwen3.5:9b, gpt-oss:20b, or selected 27B–30B models are more realistic. Budget for model weights, KV cache, metadata and OS reserve. See Unified Memory, Ollama on Mac mini M4, LM Studio vs. Ollama and the Hardware Matchmaker.

Workload and Budget Decision Matrix

Choosing between MiniMax-M3.1-Flash on the cloud API, an M Plan subscription, a frontier model, or local Apple Silicon inference depends on task complexity and data sensitivity:

Mac Development Scenario	Recommended Path	Technical and Cost Rationale
High-volume agent sub-steps, tool routing, and iterative bug patching (cloud allowed)	MiniMax-M3.1-Flash via Pay-as-you-go API	Flat $0.15 input, $0.015 cache read, and $0.60 output rates with no 512K price cliff keep per-task costs low; tuning effort to medium or high curbs thinking overhead 311.
Interactive solo coding in MiniMax Code, Hermes Agent, or Cursor	MiniMax-M3.1-Flash on M Plan Go ($22/mo) or Explore ($55/mo)	Predictable monthly billing with a Subscription Key; requires monitoring the shared 5-hour and 7-day rolling quotas across all connected tools 818.
Existing M2.5 and M2.7 API integrations	Explicit migration to MiniMax-M3.1-Flash before November 18, 2026	Remove thinking: disabled and effort: none (HTTP 400), raise max_tokens to accommodate default max reasoning, and adjust concurrency for the 200 RPM / 10M TPM limit 13114.
Complex shell automation and full-repository generation	Hybrid routing with a frontier fallback (Claude Opus 5 or DeepSeek V4.1 Flash-1017)	Use M3.1 Flash for fast file discovery and initial patches, escalating hard terminal and architecture tasks where M3.1 Flash trails on Terminal-Bench 4.0 and NL2Repo-Bench 110.
Confidential code, offline work, or strict ZDR mandates	Smaller local Mac models or a high-memory, self-hosted M3.1 server	Weights are public but large and license-conditional; hosted endpoints still process prompts externally 56.

Sources

Footnotes

1. MiniMax Blog: MiniMax M3.1 Flash — Redefining Agent Performance and Cost (October 19, 2026) ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹ ↩¹² ↩¹³

2. MiniMax API Docs: Text Generation & Model Invocation ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹ ↩¹² ↩¹³ ↩¹⁴

3. MiniMax API Docs: Pricing Overview (Pay-as-you-go & Model Rates) ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹ ↩¹²

4. MiniMax M Plan Product Page & M2.5/M2.7 Model Retirement Notice ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹ ↩¹² ↩¹³ ↩¹⁴

5. Official MiniMax M3.1 Flash model card and weights (Oct 23, 2026) ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

6. MiniMax Community License for M3.1 Flash (official license text) ↩ ↩² ↩³ ↩⁴

7. MLX Community: MiniMax M3.1 Flash Text 4bit (third-party conversion) ↩ ↩²

8. MiniMax API Docs: M Plan Overview (Go, Explore, Build) ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹

9. OpenRouter: MiniMax M3.1 Flash (minimax/minimax-m3.1-flash:free) ↩ ↩²

10. LLM Stats: MiniMax M3.1 Flash Benchmarks, Pricing & Context Window (October 19, 2026) ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

11. MiniMax API Docs: Anthropic SDK Compatibility, Parameters & Thinking Control ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹ ↩¹² ↩¹³ ↩¹⁴ ↩¹⁵ ↩¹⁶ ↩¹⁷ ↩¹⁸ ↩¹⁹ ↩²⁰ ↩²¹

12. MiniMax API Docs: Tool Use & Interleaved Thinking ↩ ↩² ↩³ ↩⁴ ↩⁵

13. MiniMax API Docs: Rate Limits (RPM & TPM by Model) ↩ ↩² ↩³

14. NVIDIA: MiniMax M3.1 Flash model card, serving and architecture ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

15. MiniMax API Docs: Prompt Caching & Pricing Formula ↩ ↩² ↩³ ↩⁴ ↩⁵

16. MiniMax API Docs: Codex Context Window & 512K Cost Optimization ↩ ↩²

17. MiniMax API Docs: Existing Token Plan Subscribers Notice ↩ ↩² ↩³

18. MiniMax API Docs: M Plan Usage Rules, Quotas & Rate Limiting ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

19. MiniMax API Docs: Hermes Agent Setup & Thinking Depth ↩ ↩² ↩³