MiniMax M3 is worth a controlled pilot when your work needs long context, image or video input, and an agent that can iterate through tools. It is not a blanket replacement for M2.7: simpler work may prefer M2.7's smaller context and faster economics, while a difficult debugging loop may still need escalation.
The decision anchor is a capacity-versus-quota split. MiniMax's API documentation, checked September 20, 2026, lists 1,000,000 total input-plus-output tokens for M3 versus 204,800 for M2.7. The official Token Plan and community reports show why a larger context window is not the same thing as a larger monthly allowance or lower completed-task cost.
Key takeaways
M3 combines open-weight distribution, a 1M total context limit, native image/video input and agentic coding claims.
Official API docs list approximate output speed above 100 tokens/s for M3 versus about 60 for M2.7, but provider latency and harness behavior can differ.
MiniMax Code, Token Plan, pay-as-you-go API, third-party providers and Tabbit are different access and billing layers.
MiniMax's release reports 59.0% SWE-Bench Pro, 66.0% Terminal-Bench 2.1 and long autonomous coding examples; these are vendor evidence.
Community reports split between “excellent workhorse” and quota or hard-debugging frustration. Treat both as workflow signals, not universal rankings.
For source-level resources, see the MiniMax M3 model page, its prompt library and review library.
MiniMax M3 at a glance
| Question | Current source snapshot | Boundary |
|---|---|---|
| Model and release | MiniMax-M3; launch page dated June 1, 2026 | A hosted route may expose different controls |
| Context | 1,000,000 total input plus output tokens | Not 1M input plus unlimited output |
| Modalities | Text with native image and video input; desktop-operation claim | Tool permissions depend on the client |
| API access | HTTP, Anthropic SDK and OpenAI SDK | API key and Token Plan key are separate |
| API pricing | Standard input-length rate up to 512K; higher long-context rate above 512K | Exact current dollar table was not visible in the opened docs |
| Token Plan | Plus $20/1.7B, Max $50/5.1B, Ultra $120/9.8B approximate monthly M3 tokens | Shared text, image, speech and music pool |
| Tabbit | Public homepage lists MiniMax M3 | No authenticated task, exact allowance or Provider was verified |
The API docs also list M2.7 at 204,800 total context and approximately 60 tokens/s. Maximum token count means input plus output, so a long source package leaves less room for the answer than the headline context number suggests.
Access is four different economic systems
API pay-as-you-go
The API is billed per call and separates API keys from Token Plan subscription keys. MiniMax says requests up to 512K input tokens use the standard rate and longer inputs use a higher long-context rate. The opened official pages did not show a complete current dollar table for M3, so use the live MiniMax pricing overview before budgeting. Do not convert Token Plan tokens into dollars or assume the API and subscription count usage the same way.
MiniMax Token Plan
The launch snapshot lists Plus at $20/month with about 1.7B M3 tokens, Max at $50 with about 5.1B and Ultra at $120 with about 9.8B. MiniMax says modalities share one pool. These are subscription allowances, not API rates. Thinking can be toggled, but a task can still consume more allowance through long context, tool calls and output.
Third-party Providers
A Provider may add its own markup, cache rules, privacy terms, rate limits, region and harness. One Reddit commenter described an OpenCode offer as about $60 of M3-rate usage bundled into a $10 subscription, but that is a provider/account report, not MiniMax billing. Compare the provider invoice and data policy separately from the model's advertised capability.
Tabbit
Tabbit's public homepage lists MiniMax M3 among supported models and describes a free Trial allowance, a no-cost Standard tier with more use after setting Tabbit as default, and an optional paid Pro tier. It does not show exact M3 quota, per-model cost or provider routing on the public page. The Tabbit overview and AI browser comparison explain the client layer; neither substitutes for an authenticated M3 test.
What changed from M2.7?
| Dimension | M2.7 | M3 | What to verify |
|---|---|---|---|
| Total context | 204,800 tokens | 1,000,000 tokens | Measure usable context after system and tool messages |
| Approximate output speed | About 60 tokens/s | Above 100 tokens/s | Provider and concurrency change observed speed |
| Modality | M-series text route | Native image/video input in release description | Confirm client tools and file limits |
| Thinking | Route-dependent | API toggle; Token Plan docs say on by default in Claude Code | Log the setting in every comparison |
| Agent thesis | Strong agentic model | Long-horizon coding, tool use and multimodal collaboration | A harness, not just the model, creates tool behavior |
| Distribution | Hosted M2.7 routes | Open-weight claim plus MiniMax Code/API | Hosted weights and local weights may differ |
The MiniMax API documentation is the source for the context and speed comparison. The Claude Code setup guide says to clear conflicting Anthropic environment variables, configure the route, then verify /status and /model shows MiniMax-M3. That is a useful access check, not a performance guarantee.
What the evidence supports
MiniMax reports 59.0% on SWE-Bench Pro, 66.0% on Terminal-Bench 2.1, 34.8% on SWE-fficiency, 28.8% on KernelBench Hard and 74.2% on MCP Atlas. It also describes an ICLR paper reproduction that ran nearly 12 hours with 18 commits and 23 figures, and a Hopper FP8 optimization that ran about 24 hours with 147 submissions and 1,959 tool calls, improving reported utilization from 7.6% to 71.3%. These examples support a pilot hypothesis about long feedback loops; they do not establish your success rate.
Independent signals are narrower. Joseph Perla's X post reports five M3 runs scoring 68.1 on a 100-task DRACO setup versus 65.3 for Fable 5, with a modeled $37 M3 run. A Reddit brownfield test says M3 retained context across 20-plus tool turns while K2.6 was faster on single turns. Neither is a standardized replication.
Community reports provide the trade-off. One M2.7 user says M3 quota restrictions weakened the value proposition. Another calls M3 a cheap workhorse but reports failure on an Electron asar problem that Opus solved quickly. The lesson is not “M3 is weak”; it is “keep an escalation path.”
For the broader client decision, see agentic reasoning and deep research, browser automation and best AI browsers in 2026.
Identity and workload self-check
| If this describes your work | Start with | Why |
|---|---|---|
| A repository with long logs and repeated tool calls | M3 API or MiniMax Code | The 1M ceiling and long-loop design are relevant |
| Image/video-grounded research | M3 through a route that exposes files | Native input is useful only when the client passes it |
| Short edits, routing and mechanical sub-tasks | M2.7 or a faster model | More context is not automatically better economics |
| Hard debugging with unknown failure modes | M3 plus a stronger escalation model | Community evidence shows a workhorse can still stall |
| Strict monthly quota | Measure a complete session first | Plan pool, thinking and tools affect burn |
| Privacy-sensitive or regulated work | Check provider region and retention | Open-weight does not define hosted data handling |
A sensible next step is a reversible task: a small repository issue with a known test, or a multimodal document packet with a fixed output schema. Record model ID, route, context size, thinking setting, tool calls, elapsed time, retries, quota use and accepted-result cost.
Tabbit boundary and next step
Tabbit's public page was checked during this research. It lists MiniMax M3, but the session exposed no authenticated selector, model label, transcript or exact allowance. Therefore this overview does not claim a Tabbit M3 run. Before publication, a real account should verify the selected model, run the reversible task above and capture the output and limits.
Tabbit BrowserVerdict
MiniMax M3 is a strong pilot for long-context coding, multimodal inputs and persistent agent loops. The context expansion and official long-run examples are meaningful. They do not collapse API prices, Token Plan quotas, Provider offers or Tabbit allowances into one number, and they do not remove the need for escalation on hard debugging.
Start with the route you can audit, log the full task cost and keep M2.7 for simple work. Leave this article as a draft until the authenticated Tabbit test and 4–8 community captures are complete.
Sources
FAQ
What is MiniMax M3?
MiniMax M3 is MiniMax's open-weight M-series model for coding, agentic tool use, long context and native image/video input. The official API docs list a 1,000,000-token total input-plus-output context limit.
What changed from MiniMax M2.7?
The official API table lists 1,000,000 tokens versus 204,800 for M2.7 and approximate output speed above 100 tokens per second versus about 60. M3 also adds native multimodality and a thinking toggle.
How can I access MiniMax M3?
MiniMax lists MiniMax Code, the Token Plan and the pay-as-you-go API. The API supports HTTP, the Anthropic SDK and the OpenAI SDK. Third-party provider offers and Tabbit allowances are separate routes.
How much does MiniMax M3 cost?
The launch page lists Token Plan tiers of $20, $50 and $120 per month with approximate M3 token pools. It describes API billing by input length but the opened pricing pages did not expose a complete current dollar table; do not treat Token Plan quotas as API prices.
Is MiniMax M3 available in Tabbit?
Tabbit's public page lists MiniMax M3 among supported models, but this overview did not verify an authenticated model selector or run an M3 task. Exact Tabbit allowances and Pro pricing were not shown.
Who should use MiniMax M3?
It is a good pilot for long-context coding, multimodal research and persistent agent loops. Keep M2.7 or another faster model for simple sub-tasks, and escalate hard debugging instead of assuming a larger context solves it.