MiniMax-M3 debuts, eclipsing GPT-5.5 and Gemini 3.1 Pro on key benchmark performance for just 5-10% of the cost

Big news in enterprise AI broke over the weekend as Chinese AI startup MiniMax released its highly anticipated M3 large language model on Sunday evening Eastern time, pairing frontier-tier coding and agentic performance with a 1-million-token context window and native multimodality for a fraction of the cost of leading proprietary models, with pricing starting at just $20 per month under its new subscription token plans.

The company's leadership also announced plans to deliver the model under an open source license including "open weights," allowing for full enterprise downloading and customizability free-of-charge, coming sometime in the next 10 days. For now, it is available via the MiniMax API at a special discounted price of $0.3 per 1 million input tokens and $1.20 per million output tokens (on fresh cache) for the next week — beating proprietary U.S. giants like Google, OpenAI and Anthropic handily on cost, while also eclipsing the performance of the latest models from the former two on selected benchmarks.

Even at its full price of $0.6/$2.40 per million input/output tokens, MiniMax-M3 remains at just 8-20% the cost of the leading, proprietary U.S. models.

The traditional matrix governing large language model development has long dictated a rigid choice: software developers can either access top-tier closed-source intelligence behind restrictive APIs, or deploy nimble, cost-effective open models that falter on multi-step reasoning, dense coding tasks, and massive data sequences. MiniMax-M3 fundamentally upends this paradigm.

By unifying these two historically separated frontier capabilities, M3 introduces a level of comprehensive utility previously restricted to expensive, closed-source ecosystems, effectively shifting the baseline of open-weights systems while drastically minimizing the operational compute footprint required to execute complex development loops.

VentureBeat Frontier AI Model API Pricing Snapshot

New MiniMax Sparse Attention (MSA) technique helps keep the model's cost low

At the core of the model's efficiency lies an architectural departure from classic Transformer networks. Standard attention mechanisms scale quadratically ($O(N^2)$), meaning computational and financial costs explode as text inputs lengthen.

To combat this "inherent flaw," the engineering team implements MiniMax Sparse Attention (MSA), a clean, extensible sparse attention blueprint.

To visualize this innovation, think of traditional full attention as an editor reading an entire library from scratch every time they need to verify a single sentence. MSA acts as an intelligent indexing clerk, using a pre-filtering phase to partition Key-Value (KV) matrices into highly precise blocks.

At the operator level, MSA uses a "KV outer gather Q" approach. The system treats KV blocks as an outer loop, dynamically aggregating only the specific queries that hit them. Because each data block is read exactly once and memory access remains strictly contiguous, hardware utilization skyrockets.

In internal trials, MSA runs more than 4x faster than alternative open-source solutions like Flash-Sparse-Attention or flash-moba.

When managing a maxed-out context length of 1 million tokens, M3’s per-token compute demand drops to just 1/20th of the previous generation model, translating into a 9x acceleration in the prefilling stage and a 15x boost during decoding.

Rather than taking a pretrained text network and fusing it with a separate vision model, MiniMax engineered M3 as a natively multimodal system from "Step Zero".

The company overhauled its data ingest machinery to blend naturally interleaved sequences of text, images, and visual components, scaling the total pretraining corpus beyond 100 trillion tokens.

This deep data alignment enables the model to translate complex visual geometries, such as programming charts or coordinate maps, into structural code without losing contextual fidelity. On standardized assessments, M3 validates this engineering path.

The model records a 59.0% on SWE-Bench Pro, an autonomous agent metric, positioning it ahead of closed models like GPT-5.5 and Gemini 3.1 Pro. It achieves a 66.0% on Terminal Bench 2.1, a 74.2% on MCP Atlas, and an 83.5 on BrowseComp—outstripping Claude Opus 4.7’s benchmark score of 79.3 in autonomous browsing and information retrieval.

However, when contrasted with Anthropic's newly released, premium frontier model, Claude Opus 4.8, from last week, the competitive ceiling of M3's efficient sparse-attention footprint becomes evident across directly comparable, tool-intensive agent benchmarks.

In the domain of pure code modification on SWE-Bench Pro, M3’s 59.0% score drops behind Opus 4.8’s leading 69.2% threshold.

A similar performance delta manifests in automated system environments via Terminal-Bench 2.1; while M3’s 66.0% terminal execution score effectively runs neck-and-neck with the previous-generation Opus 4.7 baseline of 66.1%, it trails the upgraded Opus 4.8 architecture, which achieves 74.6%.

Furthermore, evaluations tracking continuous GUI interaction on the OSWorld-Verified sandbox place M3’s automated computer use at 70.0%, compared to a higher 83.4% validation rate secured by Opus 4.8.

These standardized evaluations illustrate the structural trade-offs currently defining the ecosystem: closed-source systems like Opus 4.8 maintain absolute margin leads on hyper-complex reasoning vectors, yet M3 delivers a highly capable baseline of local, tier-one automated operation without the compounding premium of closed-door API subscription fees.

When positioned alongside the heavy-duty inference metrics of the newly minted, fellow open weights model DeepSeek-V4 Pro Max, M3 holds its ground across core agentic categories while asserting narrow advantages in specialized code synthesis.

On the software engineering matrix of SWE-Bench Pro, M3's 59.0% resolution efficiency edges past DeepSeek-V4 Pro Max’s score of 55.4%.

However, the competitive friction tightens in command-line environments; under Terminal Bench evaluations, DeepSeek-V4 Pro Max pulls slightly ahead with a 67.9% execution accuracy over M3’s 66.0% mark.

In web orchestration and open-world browsing simulations, the two architectures reach a virtual statistical parity, with M3 registering an 83.5% on BrowseComp compared to DeepSeek's 83.4%.

Similarly, on the MCP Atlas tool-use framework, M3 secures a narrow lead at 74.2% against DeepSeek’s 73.6%.

This close alignment demonstrates that while DeepSeek handles a massive 1.6-trillion total parameter footprint with specialized high-effort reasoning modes, MiniMax's block-filtered sparse attention mechanism yields directly competitive execution efficiencies without requiring extensive parameter activation scaling.

MiniMax Code AI agent offers Agentic Team capabilities

MiniMax translates these architectural gains into immediate utility through an updated product suite divided between standalone applications, customizable subscription tiers, and raw developer infrastructure. For end-user orchestration, the flagship implementation is MiniMax Code, an AI agent product designed to maximize M3's multi-step capabilities.

Operating via web or native desktop apps, MiniMax Code runs an "Agent Team" capable of breaking massive engineering tasks into multi-stage, concurrent workflows.

The system relies on a "Producer + Verifier" adversarial harness loop. As one agent instance generates code, a secondary verifier instance aggressively tests and reflects upon execution outputs, allowing the network to self-correct and operate autonomously for days without human oversight. Because of its native visual grounding, MiniMax Code supports direct computer use.

A developer can issue a cross-application voice prompt via their phone to have the model open a localized enterprise ERP client and batch-populate data tables directly from an open Excel spreadsheet.

For custom setups, developers can pipeline M3 directly into existing workflows using an API key (sk-cp) compatible with common alternative IDE environments like Claude Code, Cursor, Roo Code, and Cline. The API introduces a toggleable "thinking mode".

When enabled, M3 routes processing power into deep reasoning and long-horizon planning; when disabled, the model runs at minimal latency for quick text completion. The companion Token Plan models an aggressive pricing strategy structured around shared multimodal quotas. Billed annually, three options are available:

• Plus ($20/month): Supplies ~1.7B tokens per month and handles 3–4 concurrent agents.

• Max ($50/month): Supplies ~5.1B tokens per month, manages 4–5 concurrent agents, and adds 3 automated video clips per day via Hailuo 2.3.

• Ultra ($120/month): Supplies ~9.8B tokens per month, facilitates 6–7 concurrent agents, and extends video capacity to 5 daily clips.

Open weights makes M3 much more attractive for enterprise use

MinMax's pledge to release M3 under an open-weights license model—with weights and technical documentation launching on HuggingFace and GitHub within 10 days—carries significant strategic weight for enterprise infrastructure managers.

However, it is still to be determined precisely which license the weights will be available under, and whether or not it will be permissible for consumer usage, e.g. MIT, Apache 2.0 or the new OpenMDW license. If so, the calculus looks like this:

By shipping the underlying model weights directly to the community, MiniMax departs from the closed-door approach favored by major American AI labs.

For enterprise users bound by strict compliance and privacy rules, open weights mean they can run M3 locally on internal hardware.

This setup completely removes the risk of data leakage associated with public APIs. Furthermore, it permits engineering teams to run bespoke fine-tuning passes, modify internal architectures, or embed specialized system prompts deep within the model layers—transforming an off-the-shelf system into a highly targeted proprietary asset.

Initial community reactions are resoundingly positive

The developer ecosystem reacted immediately to M3’s operational benchmarks, singling out its long-horizon autonomous behavior and cost-to-performance profile.

A major focal point of discussion is a 12-hour automated verification test where M3 was tasked with reproducing an ICLR 2025 Outstanding Paper Award winner, titled "Learning Dynamics of LLM Finetuning".

As MiniMax's own researcher @MikaStars39 highlighted on X:

"M3 ran autonomously for nearly 12 hours, producing 18 commits and 23 experimental figures on its own, and got the core experiments working:

• it matched the predicted probability trends in the SFT stage

• clearly observed the squeezing effect central to the DPO experiments

• validated the Extend mitigation method proposed in the original paper."

Simultaneously, creators of developer tools highlighted the practical economic advantages of the model's new attention mechanism. The official team behind the agentic AI coding harness Cline posted an alert confirming day-one compatibility, stating:

"The new MiniMax-M3 is their first model to have 1m context, multimodal, and agentic coding capability. Congratulations to @MiniMax_AI for the breakthrough in sparse-attention architecture cutting compute & cost to 1/20th their previous generation."

This sharp drop in execution costs shifts how developers view the relationship between financial investment and capability. Tech commentator @jumperz mapped out this disruption, noting how M3 breaks a historical pattern in machine learning pricing:

By addressing context scaling limitations through fundamental attention-level optimizations rather than brute-force hardware scaling, MiniMax has established a highly efficient open-source baseline. M3 demonstrates that the next phase of agent development will not just be driven by larger datasets, but by efficient architectural choices that make frontier-level performance accessible to the broader open-source community.

For enterprises building autonomous software development or agent infrastructure, MiniMax M3 provides the ultimate "bang for the buck."

While DeepSeek-V4 Pro holds a microscopic price advantage of $0.195 per million tokens, MiniMax M3 justifies its marginal premium by delivering superior autonomous software engineering resolution rates (59.0% SWE-Bench Pro).

More importantly, because M3 is an open-weights model, the calculation extends far beyond the API chart. By deploying M3's weights locally inside private enterprise clouds, organizations completely bypass cloud data egress tracking, eliminate structural vendor lock-in, and can implement custom prefix-caching models on internal hardware. This technical approach transforms a highly efficient runtime budget into a permanent, privately owned corporate asset.