SenseTime launches Token Plan: three models free for a limited time, 1,500 uses every 5 hours

SenseTime didn’t sell memberships this time—it just issued coupons directly.

On May 8, SenseTime SenseNova launched a subscription service called Token Plan, along with a limited-time free public beta. Developers can register and use it right away—each model gets 1,500 calls refreshed every 5 hours, with no restrictions and no charge. It covers three models: the next-generation lightweight multimodal intelligent agent model SenseNova 6.7 Flash-Lite, the native unified multimodal understanding and generation model SenseNova U1 Fast, and DeepSeek V4 Flash, resold through SenseTime’s API.

Just as Xiaomi’s MiMo distributed tokens and Zhipu’s GLM-4.7 drove prices down to one-seventh of GPT-5.5, SenseTime’s move isn’t surprising—but the configuration combo is interesting.

A free lunch for developers—but the menu deserves close attention

Let’s crunch the numbers. The core terms of the Token Plan public beta (Free tier) are as follows:

• SenseNova 6.7 Flash-Lite: 1,500 calls every 5 hours
• SenseNova U1 Fast: 1,500 calls every 5 hours
• DeepSeek V4 Flash: 150 calls every 5 hours (note: 150, not 1,500)
• Up to 20 API keys can be created
• Fully compatible with the OpenAI API format; Base URL is https://api.sensenova.cn/v1

There’s a subtle gap: DeepSeek V4 Flash has only one-tenth the quota of SenseTime’s own models. The reason is easy to guess—DeepSeek V4 Flash is a lightweight edition of DeepSeek’s newly released V4 series (May), and SenseTime resells access as a channel partner, paying upstream per use. The 150-call cap is more like a “sampling ticket” rather than an “all-you-can-eat buffet.”

But that sampling ticket still packs value. Developers have tested DeepSeek V4 Flash through SenseTime’s channel at 140 tokens/s, which is among the fastest free options available domestically for scenarios involving thinking mode + 256K context.

Role allocation among the three models

This time, SenseTime didn’t put all its chips on a single flagship model. Instead, it bundled three differently positioned models into a “workflow suite.” That’s the biggest difference between Token Plan and the free-token approaches of Xiaomi and Zhipu.

SenseNova 6.7 Flash-Lite: Lightweight foundation for office workflows

This is the “slimmed-down” version of the 6.7 series, focused on multimodal intelligent agents, with a 256K context (max input 252K, max output 64K). SenseTime’s keywords for it are Cowork-Skills, Hermes Agent, and OpenClaw—in plain language, it’s tightly integrated into SenseTime’s own Agent framework, targeting long-chain office tasks like “reading documents, filling forms, writing reports.”

Officially, SenseTime claims that multimodal-based agents can save 60% of token consumption compared with pure text models when handling long workflows. That number should be taken with quotes—it’s an average estimate and real workloads vary. Still, the logic stands: in complex tasks, repeated text-only calls can burn more tokens than a single multimodal one.

SenseNova U1 Fast: Designed for infographic generation

U1 Fast is an accelerated version of SenseNova U1, specialized for infographic generation scenarios. One key detail: it uses a dedicated POST /v1/images/generations endpoint—not Chat Completions—and does not support image input.

In other words, it’s not a generalist like GPT-4o that can “chat and draw” simultaneously. It’s a vertical tool—you provide a text description, and it outputs a text-integrated infographic. SenseTime emphasizes its “unified understanding–generation architecture” for structured business contexts like PPTs, marketing posters, and data charts.

DeepSeek V4 Flash: Long context + thinking mode

Little introduction needed. The DeepSeek V4 series launched in early May; the Flash version trades some parameters for speed, supports dual thinking/non-thinking modes, 256K context, 64K output, built-in JSON output and tool calls. Including it in Token Plan is mainly to enrich the bundle—developers might not register for SenseTime’s own models, but they will for DeepSeek V4 Flash.

Integration cost: virtually zero

Token Plan is fully OpenAI-compatible—projects already using the OpenAI SDK only need two changes:

• Change Base URL to https://api.sensenova.cn/v1
• Set Model ID to sensenova-6.7-flash-lite / sensenova-u1-fast / deepseek-v4-flash
• Obtain API keys from SenseTime’s dashboard (up to 20 keys)

Existing Agent toolchains—SenseTime’s own OpenClaw and various community frameworks built on the OpenAI SDK—can migrate easily. Since 2025, this OpenAI-format compatibility has become standard practice for domestic models. Anyone still trying to enforce proprietary protocols gets left behind—no one will play with you.

This free plan is essentially a race to capture Agent entry points

The strategy behind Token Plan is obvious: free tokens are the bait, while Hermes Agent and Cowork-Skills are the real catch.

Throughout the Free tier documentation, phrases like “native Cowork-Skills support, optimized for office workflows” and “Hermes Agent and OpenClaw quick integration” are emphasized. The workflows developers build with these free tokens will depend on this Skills ecosystem. When usage scales up, upgrading to Lite or Pro tiers becomes natural. It’s a textbook PLG (Product-Led Growth) play—smarter than the old BD-driven API sales model.

The issue is, this track is already crowded. Zhipu has ChatGLM Agent, Alibaba has Bailian, ByteDance has Coze, and DeepSeek is building its own Agent. SenseTime’s advantage lies in multimodality—especially in documents, spreadsheets, and infographics. The combo of 6.7 Flash-Lite + U1 Fast does offer differentiation. But to make developers adopt Hermes Agent as their default, free tokens aren’t enough—it needs real killer apps in its ecosystem.

Lateral comparison: a “free token arms race”

Let’s count the recent moves:

• Xiaomi MiMo: distributed tens of millions of tokens in one go
• Zhipu GLM-4.7: API pricing cut to one-seventh of GPT-5.5
• SenseTime Token Plan: refreshes every 5 hours, long-term availability

SenseTime’s approach differs from the other two. One-time token drops are consumable; price wars are attritional. “1,500 refreshes every 5 hours” is a rhythm-based kind of free—you can’t hoard it, but as long as you keep using it, it keeps coming. This design better encourages developers to embed it into real daily workflows instead of just cash-grabbing once and leaving.

For domestic developers, API options are now abundant. Platforms like OpenAI Hub are becoming increasingly practical—one key can access GPT, Claude, Gemini, DeepSeek, etc., with stable domestic connections. New models are usually added quickly too. Windows of free access like SenseTime’s Token Plan, combined with such aggregator platforms for main/backup switching, make a cost-efficient pairing for budget-sensitive projects.

Conclusion

The launch of the SenseNova Token Plan isn’t a tech bombshell—6.7 Flash-Lite and U1 Fast are derivative versions of known models. What’s truly notable is the strategic shift: from selling APIs to selling workflows, from leaderboard competition to capturing Agent entry points.

With GPT, Claude, and Gemini locking down top-tier positions, domestic AI firms’ collective pivot toward “vertical scenarios + free acquisition + Agent ecosystems” is essentially inevitable. SenseTime’s step isn’t radical, but the direction is clear.

As for how long this free period will last—SenseTime only calls it a “limited-time public beta,” with no set deadline. Developers interested in trying it out should act soon—after all, free lunches rarely last forever.

References

• SenseTime releases SenseNova 6.7 Flash-Lite multimodal model - IT Home: Launch event details and Token Plan quota information
• OpenSenseNova/SenseNova-Skills - GitHub: SenseTime Cowork-Skills open-source repository