DeepSeek Open-Sources 304B Vision Model: GPT-6 Rumors, API Billing Changes, and More

DeepSeek's Late-Night Release: 304B Vision Model Completes Its Multimodal Lineup

DeepSeek has struck again — this time in the dead of night. The team published the first vision model in the V4 series on Hugging Face: FlashVision EXP, released under the permissive MIT license. This marks DeepSeek's formal entry into multimodal AI, filling a critical gap in its model lineup.

The technical specs are ambitious. The model carries a total of 304B parameters, but uses a Mixture-of-Experts (MoE) architecture that activates only 6 experts at inference time. Combined with FP4 quantization, it keeps computational costs manageable without sacrificing too much on performance. The context window stretches to a remarkable 1 million tokens, opening the door for long-document understanding and complex visual reasoning tasks.

Benchmark results for multimodal agent tasks show meaningful improvements over the previous generation. That said, a word of caution: with 304B total parameters, local deployment is essentially out of reach for most users. The barrier to entry remains high for individual developers and small teams.

How Developers Can Get Started with FlashVision EXP

For those eager to try it out, the recommended first step is to check the model weights and benchmark card on Hugging Face — pay particular attention to the 384-token-per-image limit. Integrate it into your existing multimodal pipeline before drawing any conclusions. Open-source doesn't mean plug-and-play; the model's real value needs to be validated in your specific use case.

One pattern worth noting: DeepSeek continues to release its core models under the MIT license — a deliberate strategic choice within the Chinese LLM landscape. By betting on an open ecosystem, DeepSeek is trading proprietary control for developer mindshare and influence.

Is GPT-6 Coming This Week? Multiple Signals Point to an Imminent Launch

Rumors around OpenAI's next flagship model are reaching a fever pitch. Multiple sources suggest that GPT-6, codenamed "Astra," could drop as early as this Thursday — though none of this has been officially confirmed.

One curious signal: Sam Altman switched back to his old profile picture, the kind of subtle move the community often reads as a pre-launch warm-up. At the same time, the Sora model has been experiencing frequent resets and a reinstated 5-hour usage cap — which many observers interpret as OpenAI clearing compute capacity ahead of a major release.

For users who rely on these tools in their daily workflows, here's a practical approach: screenshot your benchmark outputs for key tasks right now, then rerun the same tasks after any version update. This controlled comparison method helps you cut through the noise and make a clear-headed judgment about whether the new version is actually worth switching to.

The End of "Unlimited": Platforms Are Shifting to Compute-Based Billing

The most significant industry signal in this news cycle may be the broad shift in pricing models across major platforms.

Google has officially announced that NotebookLM will switch to compute-based billing starting September 2, adding a new 5-hour window cap. The more complex your tasks, the faster you'll burn through your quota. Combined with OpenAI reinstating its own 5-hour limit, the era of "subscribe and use freely" is coming to a close.

For heavy users, the advice is straightforward: push through any batch jobs you have queued before the new policies take effect, then go into your account settings and confirm your quota multiplier. This shift reflects a broader reality — under sustained compute cost pressure, LLM providers are collectively pivoting toward cost control and monetization as the defining themes of the industry's next phase.

Security Alert: Claude Accounts Compromised by Info-Stealing Malware

Anthropic has disclosed a security incident: attackers used info-stealing malware — including Cedar and Amos — to steal Claude login credentials, then exploited those accounts to drain users' image generation credits and API quotas. In response, Anthropic force-logged out affected accounts and removed linked payment methods.

This incident highlights a sobering reality: account assets can be directly stolen, and even multi-factor authentication (MFA) may not be enough to stop session cookie hijacking. The specific security steps you should take right now: check your active login devices, sign out of all active sessions, clear your cookies — and if you're on a Mac, run a malware scan while you're at it.

On the same day, Sony's lawsuit against Anthropic resurfaced in the news, with the core dispute still centered on copyright and tightening platform policies. AI companies are now fighting on two fronts simultaneously: compliance and security.

Expert Take: Why You Feel Like "The Model Got Dumber"

A lot of users have been complaining that models seem to have gotten worse — but the root cause is often not a degraded base model.

There are three more likely culprits: the routing system quietly switched to a cheaper model variant; the inference budget was trimmed; or safety filters were tightened. Think of it like this — the same service desk, but staffed by a more cost-conscious clerk. Nothing changed on the surface, but the experience quietly degraded.

The right way to test whether a model has actually regressed: fix your input channel, fix your prompt, fix your task — then run a controlled comparison. Don't rely on gut feeling alone.

Chinese LLMs: From Racing on Models to Racing on Revenue

The commercial numbers are also worth paying attention to. Reports indicate that Zhipu AI posted revenue of 953.9 million RMB in the first half of the year — a 4x year-over-year increase — with API revenue surging a staggering 27x. MiniMax crossed $800 million ARR in August, with paid (ToP) revenue now accounting for 63% of the total.

These figures tell a clear story: China's leading LLM players are shifting from competing on model capabilities to competing on revenue. Cost reduction and monetization are the central themes of the second half. For teams sourcing Chinese LLM APIs, the recommendation is simple: pull the latest pricing from both providers, build a comparison table, and do your homework before committing.

Closing Thoughts

From DeepSeek completing its multimodal lineup and GPT-6 launch rumors, to tightening API billing and account security incidents — the stories in this cycle collectively point to the same underlying trend: the technology race is still accelerating, but cost control, monetization, and security compliance have risen to become equally central themes. For developers, staying technically sharp while keeping a close eye on cost accounting and security hygiene isn't optional — both are table stakes now.