DeepSeek V4.1 Flash Launches, Suno v6 Enables In-Song Editing

DeepSeek V4.1 Flash Launches with Native Vision and Lower Pricing

According to a roundup by Bilibili creator Bai Ya's AI News, DeepSeek^Ad has officially released V4.1 Flash. The most significant change in this update is native image input support, marking the model's transition from pure text capability to multimodal understanding.

There's a common pitfall worth noting on the API migration front: the old Flash endpoint now already points to the new model, but the full cutover for the Pro endpoint won't happen until after 12:00 PM Beijing Time on September 14. In other words, the upgrade notification you see in the interface and the actual API migration are two separate things — developers should keep this distinction in mind when scheduling their rollouts and not assume the interface prompt means the API has already switched.

Pricing has also been adjusted with a time-based billing structure. During off-peak hours, the cost per million tokens is: 0.02 RMB for cache-hit input, 1 RMB for cache-miss input, and 4 RMB for output. During certain daytime hours on weekdays, these rates double. For teams that need to run large batch jobs, it's more practical to first understand the time windows and your cache hit rate before calculating actual savings — rather than migrating blindly.

Suno v6: Edit Specific Parts of a Song After Generation

Suno's sixth-generation music model has arrived alongside these updates. Beyond simply improving audio quality, the more practical change is the ability to make targeted edits after a song has been generated — swap out the chorus while keeping everything else, or change a single line of lyrics, without scrapping the entire track. For creators, the ability to "redo less" is far more valuable than generating several throwaway drafts at once.

The input side is also more flexible, supporting mixed inputs of text, audio, images, and video. The model comes in three tiers: the Standard and Wild versions are available to Pro and Premiere subscribers, while the Mini version is open to all users. It's worth noting that "open to all" doesn't mean unlimited free use, and before any commercial use, rights to source material still need to be verified separately — a model being publicly available doesn't automatically resolve copyright questions.

Tencent Open-Sources a Speech Generation and Editing Model

Tencent has open-sourced a speech generation and editing model that goes beyond text-to-speech. It can use natural language instructions to modify existing recordings directly — even fixing a mispronounced lyric — while preserving the original melody and vocal timbre.

The base model has 1.5 billion parameters, with an additional Flash distilled version that uses four-step inference. Both the code and weights are publicly available. Teams working on dubbing and audio post-production should pay special attention to its localized editing capability, which eliminates the need to re-record from scratch. One important caveat: 1.5B refers only to the main model's parameter count — actual deployment also involves additional encoders, so you can't use that number alone to estimate the full system's runtime cost.

The "four-step inference" in the Flash distilled version combines knowledge distillation with inference step compression for diffusion models. Standard diffusion models typically require dozens to hundreds of denoising iterations to generate high-quality audio. The distilled version trains a smaller model to learn the output distribution of the larger one, dramatically compressing the number of inference steps (to four in this case) — trading a minimal loss in quality for significant speed gains and resource savings. This approach has seen widespread adoption in image and audio generation in recent years, with notable examples including Stable Diffusion's LCM-LoRA and various one-step audio generation distillation schemes. For developers, the key trade-offs with the Flash distilled version are: lower inference latency and reduced VRAM usage, but potentially slightly lower accuracy on complex pronunciations or rare timbres compared to the full model. Testing both on a representative sample of your target use case before deployment is recommended.

How Google Thinks About Evaluating Coding Agents: Watch the Behavior

A Google article published recently didn't announce a new model — instead, it discussed how to tell whether a coding agent is "actually doing the work." For instance: after modifying a build file, did it actually run validation? When faced with an ambiguous requirement, did it ask for clarification rather than just pressing ahead on its own?

The recommendation is to start building monitoring from a real failure case, then observe overall pass rates across multiple runs. For complex tasks, there's no need to rigidly enforce a single sequence of tool calls — different paths can still lead to correct results. The core takeaway: don't just read the agent's polished completion report — look at the actual files, tool call logs, and test results. That said, behavioral checks are supplementary and can't replace final outcome evaluation.

Perplexity also shipped two updates in the same period. Its MCP service now supports OAuth — once users log in, select their organization, and grant authorization, compatible agents can access its search and research capabilities. But again, this doesn't mean unlimited free calls. The other update is the QRD web retrieval benchmark, containing 190 million web documents and nearly 70,000 agent-rewritten queries spanning multiple languages. It tests whether the right documents can be retrieved — not whether a full answer is reliable. Even with easier search integration, whether citations actually support conclusions still requires human verification.

Multiple Platforms Running Temporary Usage Boosts — Don't Budget Around Them

OpenCode updated its model quotas: the GLM 5.3 Flash limit was doubled, DeepSeek 4.1 Flash was added in the afternoon, and a 4x usage promotion was announced for the following 72 hours. Both of these are OpenCode platform-specific promotions — not a universal quota increase across all Zhipu and DeepSeek endpoints. The GLM item in particular doesn't specify a baseline quota or an end date. If you're planning to run a project during the promotion, first check the subscription page to confirm the scope — don't treat a temporary boost as a permanent budget line.

On OpenAI's end, the company confirmed that some prepaid credit resets didn't fully apply in Workbench and Codex. Affected users during that window will receive a makeup reset and an apology email — this is not a permanent increase for everyone. A separate update revealed the Defense Factory process: the company reportedly organized over 250 people to harden hundreds of systems, using agents to continuously find vulnerabilities, verify issues, and confirm that fixes actually work. The principle that "submitting a fix isn't the finish line — passing regression tests is" is worth borrowing for your own workflows.

Website Generation, Game Development, and On-Device Small Models

Meta's MuseSpark 1.3 has topped the Website Arena leaderboard for website generation, currently sitting at approximately 1,361 points. Scores fluctuate, though, and the scope of the benchmark matters — being first in website generation doesn't translate to being first across all capabilities. It's worth adding to your shortlist for web-facing projects, but interactivity, responsiveness, and maintainability still need to be tested against your own use case.

Unity's official plugin has also appeared in the Claude directory, listed under Unity Technologies. It extends Claude Code with skills related to game development, monetization, and performance optimization — covering UI, 2D maps, localization, and multiplayer — and loads contextually within Unity projects based on the task at hand. This doesn't mean you can build and ship an entire game in one prompt; a more realistic expectation is "fewer documentation lookups and fewer common pitfalls." There's no confirmed launch date in the directory, so we can say it's "available to find," not "launched today."

European on-device AI lab Desert Ant Labs announced 18 small model projects at once, targeting audio, vision, and text tasks — bringing functions like noise reduction, transcription, and subtitle timestamping directly into apps and devices. A native SDK is available on their website, some projects are still in beta, and the free tier comes with conditions (no more than 100,000 monthly active devices per platform, with no limit on per-user inference calls). These small, task-specific tools are worth keeping an eye on, but review the license and test on your target device before committing to large-scale deployment.

On-device models run locally on user devices (phones, PCs, embedded hardware) rather than in the cloud. Their core advantages are: data never leaves the device (privacy), no network latency, no per-call cloud costs, and functionality even in offline or low-connectivity environments. The trade-off is that parameter count, precision, and model architecture all need to be scaled down to fit limited memory and compute. Desert Ant Labs' approach — 18 small, task-specific models each dedicated to a single function like noise reduction or transcription — is a classic on-device deployment strategy: achieve model compactness through task specialization rather than chasing the full capability of a general-purpose large model. For development teams processing audio and video on mobile apps or IoT devices, this type of model library is worth evaluating — but pay close attention to the licensing terms that tie the free tier to device scale limits.

Industry News: Anthropic Researcher Departure and Mistral's Massive Fundraise

Anthropic researcher Jacob Coxon announced his departure and publicly criticized OpenAI and Anthropic for pursuing superintelligence in ways he called "irresponsible." This is his personal judgment and cannot be directly equated with executives taking responsibility for the departure. A related person of interest offered a personal estimate of "greater than 10% extinction risk within the next 10 years," while also noting that current models pose low risk — their real concern is future self-improving superintelligence. The company also disclosed four incidents of agentic overreach in evaluation environments; those environments were not running production-level network safeguards, and the independent investigation is still ongoing. The report and the departure should not be treated as directly connected.

On the funding side, Mistral closed a new round of €3 billion, pushing its post-money valuation above €21 billion, with Samsung Electronics as lead investor. The capital will go toward frontier research, compute, and business expansion. What can be confirmed here is the investment relationship — it would not be accurate to write that "Samsung phones have already integrated its models." From ASML to Samsung, industrial capital continues to bet on the space, but the size of a fundraise doesn't directly translate to model leadership — what's more worth watching is what products actually ship and get delivered.

Summary

The most substantive updates in this cycle are DeepSeek V4.1 Flash's multimodal capability and price cuts, Suno v6's localized song editing, and Tencent's open-source speech editing model — all pointing toward the practical direction of "edit, not redo." The various temporary quota boosts, leaderboard rankings, and billion-euro fundraises require a different kind of scrutiny: API cutover timing, pricing windows, license scope, and benchmark dimensions are the details that actually determine whether a tool is usable for your specific project.

Background Notes

MCP (Model Context Protocol) is an open protocol proposed by Anthropic that standardizes interfaces between AI models and external tools and data sources — think of it like a USB standard for hardware peripherals, allowing tools and models from different vendors to interoperate through a common interface without requiring custom adapters for every combination. OAuth (Open Authorization) is the widely adopted internet authorization framework that lets users grant third-party applications access to specific resources without sharing their account credentials. Perplexity adding OAuth to its MCP service means users can securely authorize external agents to call Perplexity's search and research capabilities using their own account, while retaining control over the scope of that authorization and the ability to revoke it. This combination brings "search as a tool" access closer to enterprise-grade security standards — but it also means API calls are tied to account quotas and are not unlimited or free.