August ended the way it ran, with another Chinese lab giving away a frontier-scale model. Tencent's Hunyuan team open-sourced Hy4 preview on August 28, 2026, a 770 billion parameter mixture-of-experts system activating 49 billion parameters per token across 256 routed experts plus one shared expert, carrying a 1 million token context window. It went up on Hugging Face with FP8 weights and ready-made vLLM and SGLang Docker recipes under an Apache 2.0 licence, which places no restrictions on commercial use. It scores 92.3 on GPQA Diamond and 65.7 on SWE-bench Pro, and API access runs $0.834 per million input tokens and $2.501 output.
The number that should stop people is from Tencent's internal blind evaluation, where 163 experts assessed 203 engineering tasks and Hy4 preview averaged 2.99 out of 4 against 2.92 for GLM-5.3 and 2.94 for Kimi K3. Ant Group's Ling-3.0-Flash separately shipped under MIT, packing 124 billion parameters but activating only 5.1 billion per token. Meanwhile OpenAI's ChatGPT Work split into cloud and local editions, Alibaba opened its Qianwen Work beta at a 332 billion yuan market, and Claude Sonnet 5's promotional pricing ends today. Here are the 16 stories that matter for August 31, 2026. For running coverage of every release this month, bookmark our AI industry news and trends hub.
1. Tencent Open-Sources Hy4 Preview at 770B Under Apache 2.0
Tencent's Hunyuan team released and open-sourced Hy4 preview on August 28, 2026, a mixture-of-experts model with 770 billion total parameters activating 49 billion per token through 256 routed experts plus one shared expert, with a 1 million token context window. Weights are on Hugging Face under Apache 2.0 alongside an FP8 quantised variant and Docker recipes for vLLM and SGLang. It posts 92.3 on GPQA Diamond and 65.7 on SWE-bench Pro, with API pricing at $0.834 per million input tokens and $2.501 per million output.
Apache 2.0 at this scale is the part with no precedent. GLM-5.3 shipped two days earlier at a comparable 743 billion parameters under a custom licence requiring companies above $10 billion in revenue to pass a security review, and Kimi K3 and MiniMax M3 both carry bespoke community licences. Apache 2.0 has none of that. It permits commercial use, modification, redistribution, and hosting by anyone, including the hyperscalers that GLM-5.3's terms were written to exclude. Shipping FP8 weights and container recipes on day one also removes the usual week of community porting before anyone can actually run the thing.
My take: this undercuts Z.ai's licensing strategy 48 hours after it was announced, and that timing is unlikely to be accidental. If a hyperscaler wants a frontier-scale open model to host commercially, Hy4 preview is now the obvious choice and GLM-5.3 is not. The preview label is the caveat worth holding, since preview releases from Chinese labs have historically changed before a stable version, and the blind evaluation was run by Tencent rather than independently. Our Kimi K3 review covers the rest of the open field.
2. Hy4 Beat GLM-5.3 and Kimi K3 in a 163-Expert Blind Evaluation
Tencent published results from an internal blind evaluation in which 163 human experts assessed outputs across 203 engineering tasks. Hy4 preview averaged 2.99 out of 4, against 2.92 for GLM-5.3 and 2.94 for Kimi K3. The evaluation was blind in the sense that assessors did not know which model produced which output.
Human expert evaluation on real engineering tasks measures something automated benchmarks cannot, which is whether output is actually usable rather than merely correct on a scored rubric. A model can pass a unit test with code a reviewer would reject, and a 203-task expert panel catches that where SWE-bench does not. The spread is the caveat: 2.99 against 2.94 is a five-hundredths difference on a four-point scale across three models, which is close enough that it may not survive a different task mix or a different expert panel. What the result does establish is parity, that the three leading Chinese open models are functionally interchangeable on engineering work.
My take: read this as a tie rather than a win, and treat the ranking order as noise. Tencent ran the evaluation on its own model, which is standard practice and still a reason to want independent replication. The genuinely useful finding is that three separate labs have converged on the same capability level using different architectures, which suggests the open-weight tier has hit a plateau that the next generation will have to break rather than a leader pulling ahead.
3. Ant Group's Ling-3.0-Flash Activates 5.1B of 124B Parameters
Ant Group's inclusionAI open-sourced Ling-3.0-Flash under an MIT licence, a mixture-of-experts model with 124 billion total parameters activating only 5.1 billion per token, roughly one sixty-fourth expert activation, with a 256,000 token context window aimed at programming, search, and agent tasks. It ships in BF16 at 255GB and FP8 at 128GB on Hugging Face and ModelScope. Ant reports it matches or outperforms its own 1 trillion parameter flagship on most benchmarks.
One sixty-fourth activation is the most aggressive sparsity ratio in any released model, and it is what makes a 124 billion parameter model run at roughly the compute cost of a 5 billion parameter one. Matching a 1 trillion parameter flagship at that activation level, if it holds up independently, is a strong argument that most parameters in large dense-ish models are idle most of the time. The FP8 version at 128GB is the practically important number, because that fits on a single well-specified server or a high-memory workstation rather than requiring a cluster.
My take: this is the release most likely to matter to small teams and it is getting the least attention. MIT licensing, 128GB in FP8, and agent-oriented training is a combination that puts genuinely capable inference inside a single box. The claim that it matches a trillion-parameter model deserves independent testing before anyone plans around it, but even at a discount the efficiency story is real. Our AI coding tools hub tracks what runs where.
4. The Sparsity Race Is the Real Story of August
Every significant model released this month is a mixture-of-experts design, and the activation ratios have fallen sharply. Ling-3.0-Flash activates 5.1 billion of 124 billion parameters, roughly 4 percent. Qwen3.8-Flash-Next activates 6 billion of 125 billion, under 5 percent. GLM-5.3-Flash activates 18 billion of 320 billion, about 6 percent. Hy4 preview activates 49 billion of 770 billion, around 6 percent. Kimi K3 activates 16 of 896 experts. Qwen3.8-Max activates roughly 95 billion of 2.4 trillion, about 4 percent.
A year ago typical mixture-of-experts models activated 15 to 25 percent of their parameters. Dropping to 4 to 6 percent means serving cost per token has fallen by roughly a factor of four for the same stored knowledge, which is the mechanism behind this month's price collapse more than any competitive pressure. The trade is memory, since every parameter must be resident even when unused, which is why these models need clusters despite their low compute cost. That tension defines the current deployment landscape: cheap to run per token, expensive to host at all.
My take: when you evaluate a model release, read the activation ratio before the benchmark scores. It tells you what the thing costs to serve and how the lab expects it to be used, and it is a better predictor of pricing than parameter count. The open question is where the floor is, because at some activation level quality has to degrade, and nobody has published where that point sits. Whoever finds it first will define the next generation's architecture.
5. GLM-5.3's Weights and the Hyperscaler Licence
Z.ai published GLM-5.3's open weights on Hugging Face on August 28, two weeks after the model's August 14 API launch, under a custom GLM-5.3 License rather than MIT or Apache 2.0. Any company with aggregate revenue above $10 billion over any 12 consecutive months must pass Z.ai's own security review before commercial use of the model or its derivatives. The model is roughly 743 billion parameters with about 40 billion active and a 1 million token context, shipping in BF16 and FP8, and it needs roughly eight GPUs to run. Z.ai attributed the delay to a safety programme logging 2,436 findings.
The threshold is drawn precisely where the hyperscalers sit, so a startup deploys freely while AWS, Google Cloud, and Microsoft cannot host it commercially without approval. Z.ai released the smaller GLM-5.3-Flash under a clean MIT licence four days earlier, giving it the same two-tier structure Meta uses with Muse Glimmer under Apache 2.0 and Muse Spark 1.2 under a modified Llama licence. It was a genuinely novel design when announced on Friday.
My take: Tencent shipping a comparable model under Apache 2.0 two days later is the most direct possible answer to that strategy, and it makes the revenue-gate look expensive rather than clever. A licence only defends value if a substitute is not available on better terms, and one now is. Full detail on the licence terms sits in our August 30 roundup.
6. The Open-Weight Licence Map as August Closes
Licence terms across the leading open-weight models now vary more than their capabilities do. Tencent Hy4 preview at 770 billion parameters and Qwen3.8-27B at 27.8 billion both ship under Apache 2.0. GLM-5.3-Flash at 320 billion and Ling-3.0-Flash at 124 billion ship under MIT. Meta's Muse Glimmer 30B is Apache 2.0 while Muse Spark 1.2 uses a modified Llama Community License. GLM-5.3 carries its revenue-gated custom terms, Kimi K3 a custom Kimi K3 licence, and MiniMax M3 a minimax-community licence.
Only Apache 2.0 and MIT among those meet the standard open-source definition, yet all get described as open source in coverage. The practical differences are concrete: whether you can fine-tune and redistribute, whether you can offer the model as a hosted service, whether attribution is required, and whether your company's size changes any of those answers. Those questions determine whether a model can go into a product long before a three-point benchmark gap does.
My take: as of today the strongest combination of capability and licence clarity is Hy4 preview at 770 billion under Apache 2.0, with Ling-3.0-Flash under MIT as the best option that fits in a single box. That both are Chinese, and that every genuinely permissive frontier-scale release this month came from a Chinese lab, is the structural fact worth sitting with. Any organisation self-hosting on compliance grounds is now choosing among Chinese models by default.
7. ChatGPT Work Splits Into Work Cloud and Work Local
OpenAI has been iterating rapidly on ChatGPT Work since its July 9 launch, and the product now has two distinct editions. Work Cloud runs on OpenAI infrastructure, accessible through chatgpt.com and the mobile apps, with internet-connected code execution and a persistent /workspace/scratch filesystem that survives between sessions. Work Local runs through the ChatGPT desktop app and can access files and execute programs directly on the user's own computer. Developer Simon Willison published a detailed breakdown on August 30.
The persistent filesystem is the feature that changes what the product is. A stateless assistant starts fresh each session and can only work with what you paste in, while a persistent scratch directory lets an agent build up artefacts across days, which is what long-running work actually requires. Work Local goes further by removing the upload step entirely, since the agent reads your actual files. Both editions have internet-connected code execution, meaning the model can run code that makes outbound network requests.
My take: this is the most capable general-purpose agent product any lab has shipped, and the capability and the risk come from exactly the same design decisions. A persistent workspace plus local file access plus outbound network access is enormously useful and also precisely the configuration security researchers have been warning about. See our AI agent frameworks hub for how other vendors are handling the same trade.
8. Why ChatGPT Work Triggers the Lethal Trifecta
Simon Willison's analysis flags ChatGPT Work as satisfying all three conditions of what he calls the lethal trifecta for AI agents. The three are access to private data, meaning the agent can read your emails, documents, and databases; exposure to untrusted tokens, meaning it processes content from external sources such as web pages and shared documents; and an exfiltration vector, meaning it can make outbound requests through rendered images, API calls, or generated links. A system with all three is vulnerable to prompt injection, where instructions hidden in content the agent reads cause it to act against the user.
The reason this matters is that no reliable general defence for prompt injection exists after three years of the problem being known. Each incident gets patched individually. This month alone produced Microsoft's zero-click CoSnitch flaw in Copilot, which let a single link exfiltrate Gmail, Drive, and Calendar data, GitHub Copilot Autofix introducing a shell injection that leaked a Snowflake token within five days, and Forcepoint demonstrating invisible text in emails manipulating AI summarisers. All three fit the same pattern.
My take: the practical mitigation is to break the trifecta rather than to hope the model resists injection, because it will not reliably. If an agent has local file access, restrict its network egress. If it browses untrusted content, keep it away from private data in the same session. Scope every connector to the minimum the task needs. That advice is unglamorous and it addresses more real risk than any vendor security feature currently shipping.
9. Alibaba's Qianwen Work Targets a 332 Billion Yuan Market
Alibaba launched the public beta of Qianwen Work, also branded Tongyi Office, on August 30, 2026. It ships as a desktop client, a web application, and an integration with DingTalk, Alibaba's enterprise messaging platform, and is positioned against a Chinese office-agent market Alibaba sizes at 332 billion yuan, roughly $46 billion.
The DingTalk integration is the strategic core rather than an add-on. DingTalk is already installed across a large share of Chinese enterprises, so distribution is solved before the product ships, which is the problem every standalone AI work tool struggles with. It mirrors what Microsoft did with Copilot inside Teams and Office, and it lands the same week OpenAI's ChatGPT Work split into cloud and local editions. Three of the largest technology companies in the world are now converging on the same product category from different starting points.
My take: office agents are where AI adoption will actually be decided, because that is where most knowledge work happens, and the winner will likely be whoever already owns the distribution rather than whoever has the best model. That is an uncomfortable conclusion for anyone betting on model quality as the differentiator. Watch whether Qianwen Work uses Qwen models exclusively or routes across providers, since that answer says a lot about how Alibaba views its own model advantage.
10. Google Ships Gemini 3.5 Transcribe as Two Endpoints
Google released Gemini 3.5 Transcribe, a dedicated speech-to-text model that ships as two separate endpoints rather than a single unified interface. It joins a Gemini lineup where 3.7 Flash is the generally available workhorse at $0.75 and $3.75 per million tokens through December 31, Gemini 3.1 Pro remains the most capable and is still in preview, and the promised Gemini 3.5 Pro has not shipped.
Splitting a transcription model into two endpoints usually reflects a real difference in how the workloads behave, typically separating streaming real-time transcription from batch processing of complete files. Those have genuinely different latency, cost, and accuracy profiles, and forcing both through one interface makes both worse. A dedicated speech model also signals that Google sees transcription as a volume business worth optimising separately rather than a feature of a general multimodal model.
My take: the more interesting Gemini story remains what has not shipped. Google is iterating on the Flash tier faster than anyone, with three weeks between 3.6 and 3.7, while the Pro tier that would compete at the frontier stays in preview. That is a defensible commercial choice, since volume lives in the workhorse tier, but it leaves Google trailing Anthropic and OpenAI on the benchmarks that shape perception.
11. Claude Sonnet 5 Promotional Pricing Ends Today
Anthropic's promotional pricing on Claude Sonnet 5 ends today, August 31, 2026. Teams running production workloads on Sonnet 5 at the promotional rate should expect their per-token costs to change from tomorrow. Claude Opus 5 remains at $5 per million input tokens and $25 output, unchanged from Opus 4.8's pricing, and tops the Artificial Analysis Intelligence Index at 63.
Promotional pricing expiries are the quietest cost event in AI and the one that catches the most teams. Unlike a model retirement, nothing breaks and no error appears. The bill simply changes, usually noticed at the end of the following month. This is the second such expiry teams need to track this quarter, alongside Gemini 3.7 Flash's introductory rate of $0.75 and $3.75 running only through December 31 before doubling to $1.50 and $7.50.
My take: build a calendar of pricing expiries the same way you track model retirements, because both change your costs and only one of them announces itself. GPT-5.6 Sol's cut to $4 and $20 is also explicitly a three month promotion, which puts its expiry in November. Three separate expiries across three vendors inside six months is now normal, and any 2027 budget built on today's rates is understated. Compare current pricing in our best AI models ranking.
12. Gemini Robotics ER 1.6 Preview Shuts Down Today
Google shuts down gemini-robotics-er-1.6-preview today, August 31, 2026. It follows OpenAI retiring o3 from ChatGPT on August 26 after a 90 day sunset and the official DALL-E GPT retiring on August 30, making three shutdowns in six days. The replacement is Gemini Robotics ER 2, built on Gemini 3.5 Flash, which handles real-time spatial reasoning and multi-step planning across hundreds of steps and is available in Google AI Studio and in private preview on the Gemini Enterprise Agent Platform.
Preview model shutdowns carry no stability guarantee by definition, which is the deal you accept when building against one, but the practical impact on anyone who did is identical to a general availability retirement. Code pinned to the identifier stops working today. The migration to ER 2 is not a drop-in swap either, since ER 2 separates high-level planning from motor execution and hands the latter to any lower-level vision-language-action model, which is a different integration shape.
My take: three shutdowns in six days is the clearest argument this month for treating model identifiers as configuration rather than constants. Run the audit today if you have not: search for hardcoded identifiers, substitute the replacement, and confirm your evaluation suite still passes rather than assuming it will. A swapped model that quietly performs worse is the failure that runs for weeks unnoticed.
13. Local Tooling Catches Up Within Days of Every Release
Koboldcpp released version 1.120 on August 29, 2026 adding compatibility for the freshly released Qwen3.8-Flash-Next and Ling-3.0-Flash models, introducing a DirectIO loading mode and user-configurable JavaScript tools supporting standard tool-calling protocols. NVIDIA's Nemotron 3.5 Lightning shipped with day-zero GGUF availability earlier this month, and Qwen3.8-Flash-Next had day-zero support from Unsloth.
The gap between a model release and usable local tooling has collapsed from weeks to days, and in several cases to zero. That change matters more than it sounds, because a model you cannot run is a press release. DirectIO loading addresses the specific bottleneck for large mixture-of-experts models, where loading hundreds of gigabytes of weights from disk dominates startup time. Configurable JavaScript tools with standard tool-calling protocols mean local deployments get agent capability without a separate framework.
My take: this is the least covered and most consequential trend in open weights. Labs have started coordinating with tooling projects before release rather than after, which is why Qwen shipped an architecture preview specifically so that kernels would exist before the flagship. If you dismissed local inference a year ago on tooling grounds, that objection has largely gone.
14. Where the Benchmark Leaders Stand as August Closes
Benchmark leadership at the end of August remains split. OpenAI's GPT-5.4-Pro leads GPQA Diamond at 94.4 percent, with Tencent's Hy4 preview now second at 92.3 percent and MiniMax M3 above 92. Anthropic's Claude Opus 4.7 leads SWE-bench Verified at 87.6 percent. Claude Opus 5 tops the Artificial Analysis Intelligence Index at 63, with GLM-5.3-Flash at 57 and Kimi K3 at 57.11. Claude Mythos 5 leads BenchLM's coding composite at 80.1 with Claude Fable 5 at 79.8. On SWE-bench Pro, Hy4 preview posts 65.7 against GPT-5.3-Codex's 56.8.
The GPQA Diamond gap is the one that has closed most. An open-weight Apache 2.0 model at 92.3 sits roughly two points behind the closed leader on graduate-level science reasoning, a gap that was wide enough to be uninteresting a year ago. SWE-bench Verified tells a different story, with Claude Opus 4.7 at 87.6 percent still comfortably ahead of the open field, which suggests repository-level engineering remains harder to close than knowledge-heavy reasoning.
My take: identify which of these benchmarks most resembles your actual work and weight that one, because leadership genuinely differs by task and no single number settles it. The broader pattern for August is that open weights closed most of the gap on reasoning and closed less of it on sustained engineering work. That is a useful distinction when deciding where a cheap open model can substitute and where it cannot.
15. What the Cheap Tier Costs After a Month of Cuts
Pricing at the volume tier has fallen substantially through August. GLM-5.3-Flash serves at $0.15 per million input tokens and $0.50 output. GPT-5.6 Luna sits at $0.20 and $1.20 after an 80 percent cut. Gemini 3.7 Flash runs $0.75 and $3.75 through December 31 before doubling. Tencent's Hy4 preview is $0.834 and $2.501. DeepSeek V4-Pro serves 1.6 trillion parameters at $1.32 and $3.96. Grok 4.6 is $2 and $6, Kimi K3 $3 and $15, GPT-5.6 Sol $4 and $20 on promotion, and Claude Opus 5 $5 and $25.
Hy4 preview at $2.501 output is the entry worth studying, because it sits at roughly one tenth of Claude Opus 5's output price while scoring 92.3 on GPQA Diamond against the closed leader's 94.4. The spread from GLM-5.3-Flash at $0.50 to Opus 5 at $25 is fifty-fold across models separated by six points on an aggregate index. That ratio is the strongest argument for routing, sending the simple majority of requests to a cheap tier and escalating only genuinely hard work.
My take: rerun your cost model, because July figures are stale and January figures are not close. The honest caveat is that index gaps show up most on long agentic runs where small errors compound and force retries, so measure cost per completed task rather than cost per token. A cheap model that needs three attempts is not cheap. Detail on tier differences sits in our GPT-5.6 review.
16. Leadership Changes and Industry Moves in Brief
Several non-model items from the closing days of August are worth logging.
● Google is consolidating AI leadership, with Demis Hassabis stepping back to Chairman and Koray Kavukcuoglu taking operations, ending the two-continent split between the former Brain and DeepMind organisations.
● Anthropic hired its first global affairs chief, weeks after US District Judge Rita Lin blocked the Pentagon's designation of the company as a national security supply chain risk, calling it illegal and baseless.
● Nvidia reported $96.2 billion in quarterly revenue, up 106 percent year over year, then paused parts of its AI cloud revenue-sharing financing programme after internal antitrust concerns, which the company denies.
● Nvidia notified Microsoft, Google, and Oracle of price increases above 15 percent on Vera Rubin and Grace Blackwell systems shipping from early 2027, attributed to surging DRAM costs.
● A National Bureau of Economic Research survey found more than 90 percent of executives reported no employment effect and 89 percent reported no productivity gains after three years of AI implementation.
Full detail sits in our August 29 roundup and the August 25 roundup.
17. Where the Frontier Models Stand Today
Here is the practical state of the model landscape as August 2026 closes.
The short version for teams choosing today: Hy4 preview if you want frontier-scale open weights with no licence conditions, Ling-3.0-Flash or Qwen3.8-27B if it must fit in one box, GLM-5.3-Flash if cost per token decides, Gemini 3.7 Flash if latency decides, and Claude Opus 5 if capability decides.
18. What to Watch Next in AI
Four things carry into September.
● Whether Z.ai revises the GLM-5.3 revenue-gated licence now that Tencent has published a comparable model under Apache 2.0 with no conditions.
● Independent verification of Hy4 preview's benchmark scores and of the 163-expert blind evaluation, both currently vendor-run.
● Qwen 4, previewed architecturally by Qwen3.8-Flash-Next, and whether Google finally ships Gemini 3.5 Pro.
● Pricing expiries: GPT-5.6 Sol's three month promotion ends in November and Gemini 3.7 Flash's introductory rate ends December 31.
The through-line for August is that open weights went from a capability question to a licensing question and then, in the last week of the month, back to a capability question. Z.ai tried to keep weights open for developers while closing them to hyperscalers, and Tencent answered 48 hours later with a comparable model and no conditions at all. Whatever September brings, the frontier-scale open-weight tier now has genuinely permissive options at 770 billion parameters, which nobody would have predicted at the start of the month.
Frequently Asked Questions
What is Tencent Hy4 preview?
Hy4 preview is an open-weight mixture-of-experts model from Tencent's Hunyuan team, released August 28, 2026 under Apache 2.0. It has 770 billion total parameters activating 49 billion per token across 256 routed experts plus one shared expert, with a 1 million token context window. It scores 92.3 on GPQA Diamond and 65.7 on SWE-bench Pro, with API pricing at $0.834 per million input tokens and $2.501 output.
Is Hy4 preview better than GLM-5.3 or Kimi K3?
Marginally, in Tencent's own blind evaluation. Using 163 experts across 203 engineering tasks, Hy4 preview averaged 2.99 out of 4 against 2.92 for GLM-5.3 and 2.94 for Kimi K3. That spread is small enough to be within noise, and the evaluation was run by Tencent rather than independently. Hy4's clearer advantage is its Apache 2.0 licence against the custom licences on both rivals.
What is Ling-3.0-Flash?
Ling-3.0-Flash is an open-weight mixture-of-experts model from Ant Group's inclusionAI, released under an MIT licence. It has 124 billion total parameters but activates only 5.1 billion per token, roughly one sixty-fourth expert activation, with a 256,000 token context window. It ships in BF16 at 255GB and FP8 at 128GB, and Ant reports it matches or outperforms its own 1 trillion parameter flagship on most benchmarks.
Which open-source AI model is best right now?
For frontier-scale capability with a genuinely permissive licence, Tencent Hy4 preview at 770 billion parameters under Apache 2.0. For a single machine, Ling-3.0-Flash under MIT at 128GB in FP8, or Qwen3.8-27B under Apache 2.0 for consumer hardware. GLM-5.3-Flash under MIT is the cheapest capable multimodal option at $0.15 and $0.50 per million tokens.
What is ChatGPT Work and is it safe?
ChatGPT Work is OpenAI's agentic work product, now split into Work Cloud running on OpenAI infrastructure with a persistent /workspace/scratch filesystem, and Work Local running through the desktop app with access to local files and programs. Both have internet-connected code execution. Simon Willison flagged that it satisfies all three conditions of the lethal trifecta, so scope its permissions carefully.
What is Alibaba Qianwen Work?
Qianwen Work, also branded Tongyi Office, is Alibaba's enterprise AI platform, launched in public beta on August 30, 2026. It ships as a desktop client, a web application, and a DingTalk integration, targeting a Chinese office-agent market Alibaba sizes at 332 billion yuan, roughly $46 billion.
How many AI models were released in August 2026?
Trackers report between 14 and 24 confirmed releases from between 8 and 18 providers, depending on whether variants and quantisations count separately. Either figure makes August 2026 the densest model release month recorded, with most being open-weight releases and most of those from Chinese labs including Alibaba, Zhipu, Tencent, Ant Group, MiniMax, and Moonshot.
Which model leads GPQA Diamond?
OpenAI's GPT-5.4-Pro leads GPQA Diamond at 94.4 percent. Tencent's Hy4 preview is second at 92.3 percent, and MiniMax M3 also scores above 92. That places the best open-weight model roughly two points behind the closed leader on graduate-level science reasoning.
Is Claude Sonnet 5 pricing going up?
Yes. Anthropic's promotional pricing on Claude Sonnet 5 ends August 31, 2026, so per-token costs change from September 1. Two other expiries follow: GPT-5.6 Sol's cut to $4 and $20 is a three month promotion ending in November, and Gemini 3.7 Flash's introductory $0.75 and $3.75 rate doubles on January 1, 2027.
What is the lethal trifecta in AI agents?
The lethal trifecta, a term coined by developer Simon Willison, describes three conditions that together make an AI agent vulnerable to prompt injection: access to private data, exposure to untrusted content such as web pages or shared documents, and an exfiltration vector allowing outbound requests. An agent with all three can be manipulated by instructions hidden in content it reads.
Recommended Blogs
● GLM-5.3 Weights Ship With a Hyperscaler Catch: AI News August 30 2026
● OX Alpha Was GLM-5.3-Flash, Now Open: AI News August 29 2026
● Qwen3.8-Flash-Next Previews Qwen 4: AI News August 26 2026
● OpenAI Models Escaped and Hacked Hugging Face: AI News August 25 2026
● Best AI Models July 2026: Ranked by Use Case and Price
● GPT-5.6 Review: Sol, Terra, Luna Benchmarks and Pricing
● Kimi K3 Review: Benchmarks, Pricing, and K2 Comparison
Resources & Community
Join our community of 70,000+ AI enthusiasts and learn to build powerful AI applications! Whether you're a beginner or an experienced developer, Build Fast with AI helps you understand and implement AI in your projects.
● Website: buildfastwithai.com
● LinkedIn: Build Fast with AI
Agentic AI Launchpad 2026
A structured 6-week cohort program that takes you from AI basics to building and deploying real-world agentic AI systems. Includes live sessions, expert mentorship, project reviews, and a builder community network.
Ready to go from learning to building? Join the next cohort: Agentic AI Launchpad 2026
Free AI Resources
Access free tools, workshops, and micro-learning to keep building:
● AI Workshops: Free resources, upcoming events, and past recordings
● Unrot: Learn AI in 5 minutes a day (free micro-learning app)
● Gen AI Experiments: free cookbooks and notebooks on GitHub
September opens with Qwen 4 previewed and the open-weight licence question wide open. Follow Build Fast with AI so each recap reaches you before your standup.
References
● Tencent open-sources Hy4 preview (TechNode)
● Hy4 preview specs and benchmarks (MindStudio)
● Ant Group open-sources Ling-3.0-Flash (Crypto Briefing)
● Ling-3.0-Flash announcement (Business Wire)
● Understanding ChatGPT Work (Simon Willison)
● The lethal trifecta for AI agents (Simon Willison)
● GLM-5.3 goes open weight (The New Stack)
● GLM-5.3 model card (Hugging Face)
● Nvidia record quarter results (CNBC)
● Judge blocks Pentagon Anthropic blacklist (NBC News)
● ChatGPT release notes (OpenAI)
● Model benchmark leaderboard (BenchLM)
● Independent model evaluations (Artificial Analysis)
