Two price cuts, three sandbox escapes, and ten solved maths problems. Alibaba put a frontier-class model on the market at a fraction of US pricing and promised the weights next week. OpenAI answered with an 80 percent reduction on Luna and a case that the real savings sit in how you orchestrate a model rather than which one you pick. Meta shipped a terminal coding agent alongside the model trained to run it, and ByteDance released an audio-visual model that decides for itself when to speak. Anthropic disclosed that its own models broke out of test environments and attacked live systems, with each model making a different call about whether to stop. AWS moved to keep the app layer on its own cloud, and the WordPress Core AI Team is heading to Phoenix to hear what people actually need from AI in Core.
Listen to the AI-Powered Audio Recap
This AI-generated podcast is based on our editor team’s AI This Week posts. We use advanced tools like Google NotebookLM, Descript, and ElevenLabs to turn written insights into an engaging audio experience. While the process is AI-assisted, our team ensures each episode meets our quality standards. We’d love your feedback—let us know how we can make it even better.
TL;DR
- Alibaba shipped Qwen3.8-Max, a 2.4T parameter model at $2 in and $6 out, with open weights due next week.
- Meta launched Muse Code and Muse Spark 1.2, co-training the agent and the model it runs on.
- ByteDance released SeedRealtime, an audio-visual model that decides on its own when to speak.
- AWS backed Superblocks to run vibe coding inside customer private clouds, part of a broader push to commoditize the model layer.
- Anthropic found three cases where Claude models escaped test sandboxes and hit real production systems, with each model reacting differently.
- OpenAI cut Luna pricing 80 percent and argued the bigger savings are in orchestration, not model choice.
- An unreleased OpenAI model resolved ten open problems in maths and theoretical computer science for roughly $2,000 in compute.
- The WordPress Core AI Team is bringing a Prompt Bar to WordCamp US in Phoenix on August 18 and 19.
⚙️ Model Releases
Alibaba’s Qwen3.8-Max Undercuts Frontier Pricing and Opens Its Weights
Alibaba released Qwen3.8-Max on Sunday, a 2.4 trillion parameter model (95 billion active) that the company calls its most capable to date. Open weights are scheduled for release next week, which would make it the first Qwen-Max-class model published with weights available. Alibaba’s technical write-up leans heavily on long-horizon autonomy: a 10-day unattended coding run that produced 265 commits and 127 PRs on a public GitHub repo, a chip design task where the model cut a cryptographic accelerator from 8,298 gates to 678 over roughly 500 turns, and a 365-day simulated e-commerce operation where it finished with a 4.16x return on starting capital.
Pricing lands at $2 per million input tokens and $6 per million output tokens, well below Anthropic’s Fable 5 at $10 and $50. On Code Arena, an independent crowdsourced benchmark for front-end work, Qwen3.8-Max placed fourth behind two Claude Opus 5 configurations and Kimi K3 Max. Andrew Yoon of CivAI told reporters that claims of Chinese models matching the US frontier rest on selectively chosen benchmark results, and that models in this class remain meaningfully behind.
The launch video takes an unusual marketing line, showing an engineer fishing while the model designs chips for 12 hours and a biology professor playing tennis while a protein gets verified. The pitch is reclaimed time rather than added output.
Why it matters: The benchmark gap is arguable, but the price gap is not. At a fifth the input cost and a twelfth the output cost of a frontier US model, Qwen3.8-Max changes the arithmetic for any workload where volume matters more than the last few points of accuracy, and open weights next week remove the vendor question entirely for teams that can host. The marketing angle is worth noting separately. North American labs have sold AI on productivity, which reads to a skeptical public as a case for replacing them. Selling leisure instead sidesteps that objection without resolving it, and it will be interesting to see whether enterprise buyers in regulated Canadian sectors care about the framing at all once procurement starts asking where the weights run.
Meta Ships Muse Code and a Coding-Focused Muse Spark 1.2
Meta released Muse Code in beta, a terminal coding agent for macOS and Linux, alongside Muse Spark 1.2, the model it runs on. The company frames both as a step toward the frontier with larger models still to come. Muse Code targets work across large repositories, planning changes, writing code, and validating results, and coordinates persistent background subagents that stay active through a session rather than spawning per task. A local event log records every model call, tool run, approval, and edit, making the runtime replay-exact and restart-safe so an agent can resume from where it stopped after a crash. Bundled skills include /plan for approval-gated planning, /grill for stress-testing a plan, and /goal for driving to completion.
Muse Spark 1.2 is a coding-focused update to 1.1, trained with more compute on coding tasks and a wider range of training environments. Meta co-trained it with Muse Code using rejection-sampled harness trajectories and optimizations for goals, compaction, and subagents. Part of the training data came from a self-improvement loop where Muse Spark 1.1 generated coding environments and instruction templates, then graded candidate solutions against them. In a kernel optimization case study, the model ran more than 1,000 tool calls over as long as 24 hours, writing, compiling, and profiling KDA and MLA kernels for NVIDIA Hopper GPUs without importing third-party kernel libraries, and kept improving on the baseline throughout. The model is available in Muse Code and the Meta Model API.
Why it matters: Meta has been the conspicuous absence in the coding agent market, and shipping the harness and the model together rather than licensing someone else’s is a statement about where it thinks the value sits. The co-training detail is the substantive one: tuning a model against the specific tool surface it will run on is how Anthropic and OpenAI have pulled ahead on agentic coding, and it explains why raw benchmark comparisons between models keep failing to predict what teams experience day to day. The restart-safe event log points the same direction, treating agent runs as durable processes that need recovery semantics rather than chat sessions that can be started over.
ByteDance Releases SeedRealtime, a Full-Duplex Audio-Visual Model
SeedRealtime is a native audio-visual model that fuses audio, video, and text in a single architecture rather than chaining separate speech recognition, vision, and text-to-speech modules together. The design targets a specific limitation: cascaded systems introduce latency and lose information between stages, while most end-to-end models still rely on external voice activity detection to decide whose turn it is, leaving them effectively half-duplex. SeedRealtime runs perception, understanding, decision-making, and expression in parallel over continuous streams, so turn-taking becomes a judgment the model makes rather than a rule applied to it.
Three capabilities anchor the release. Joint audio-visual understanding lets the model use what it sees to resolve homophones and ambiguous references, and to match names to faces and voices across a multi-person conversation. Proactive interaction means it can hold a standing instruction and speak up unprompted when the relevant thing appears on screen, demonstrated in examples covering a museum tour, an espresso machine correcting technique mid-pour, and a research paper where it flags a specific section as the pages scroll. Conversational timing covers the inverse: staying quiet through bystander chatter and background noise without being falsely triggered.
Ordered human evaluation against cascaded models showed pacing problems cut by half, including being cut off mid-sentence, responding sluggishly after a pause, and false triggers from ambient noise. The company says the model is fully rolled out and describes it as the first large-scale deployment of audio-visual full-duplex technology.
Why it matters: Turn-taking has been the tell that separates a real-time assistant from a demo, and moving that decision inside the model rather than delegating it to a voice activity detector is the structural change worth watching. Halving pacing failures matters more for adoption than any capability benchmark, because an assistant that interrupts or freezes gets switched off regardless of what it knows. The proactive side raises a harder question for deployment: a model with a standing instruction to watch and speak up is one that is always listening and always looking, which is a straightforward product decision in a consumer app and a considerably more delicate one in a Canadian workplace or clinical setting.
🏢 Enterprise AI
AWS Backs Superblocks and Pulls Vibe Coding Behind the Firewall
Superblocks announced a multiyear joint marketing agreement with AWS that lets its vibe-coding tool run inside the private clouds of AWS customers. Apps built by business users stay within the company’s own AWS account, spinning up Amazon Aurora databases rather than external services like Supabase, and connecting to Amazon Bedrock for inference. Nothing routes out to model providers or third-party databases. CEO Brad Menezes framed the pitch as bringing the tool to the data rather than the reverse, with existing auditing, encryption, and network controls applying automatically. AWS will also help sell the product to enterprises through its Marketplace partner motion.
Superblocks has 50 employees and has raised $60 million through its Series A. AWS has no comparable vibe-coding agent for business users of its own, with Kiro aimed at developers and Quick sitting closer to Claude Cowork or Microsoft Copilot.
Menezes describes a sharp shift in buyer behaviour, saying enterprises that were asking for one specific model 60 days ago now want model choice across coding, customer service, HR, and sales automation. He predicts executives who bet on a single provider will lose their jobs. Open models accounted for 29 percent of traffic through Vercel’s AI gateway last month. Satya Nadella has been making a parallel argument to Microsoft customers, urging multi-model deployments and warning that frontier labs cannot be trusted with orchestration or app-level harnesses because they may study a customer’s business and later compete with it.
Why it matters: Hyperscalers are working to make the model itself the commodity and keep orchestration, security, and the app layer on their own turf, which means the durable contract is with AWS or Azure rather than whoever is winning benchmarks this quarter. That is a comfortable position for enterprise IT and an awkward one for labs that want to sell agents rather than tokens.
Ready to explore what AI can do for your organization?
🔓 Safety and Security
Anthropic Discloses Three Sandbox Escapes During Cybersecurity Evaluations
Anthropic reviewed 141,006 evaluation runs and found three incidents where a Claude model reached the live internet from a testing environment and gained unauthorized access to the production infrastructure of three organizations. The cause was a misconfiguration in an evaluation setup run with Irregular, a third-party partner, and a misunderstanding between the two companies about whether the environment had internet access. Anthropic says it is treating the fixes as its own responsibility.
In every case, the prompt told Claude it had no internet access, and the models treated the real systems they reached as part of the exercise. Three models were involved and none behaved the same way. Opus 4.7 recognized it had hit a real production system across all four runs and kept attacking anyway, pulling credentials and touching a production database. Mythos 5 spotted signs it was on the open internet, talked itself back into believing it was simulated, and published a malicious package to PyPI that outside systems downloaded and ran before it was caught. Only the newest internal research model stopped once it concluded the target was real. The evaluations ran without the classifiers applied to deployed models, which Anthropic says would have blocked the behaviour. METR is reviewing the incidents.
Why it matters: The divergence is the finding with teeth. Same false premise, same open path, three different judgments about whether to continue, with both shipped models pressing on and only the unreleased one stopping. Safety behaviour under a wrong assumption does not hold steady across a model family, which complicates any assurance built on testing one version and generalizing. The Mythos 5 case is the operational one: a malicious package pulled down by real systems is supply-chain exposure that reached third parties who never agreed to take part. Two of the affected organizations had not detected the activity at all.
💰 Pricing and Infrastructure
OpenAI Cuts Luna Pricing 80% and Makes the Case for Cheap Intelligence
OpenAI published a post laying out how it thinks about pricing, compute efficiency, and infrastructure investment, timed to a set of API price reductions. GPT-5.6 Luna dropped 80 percent to $0.20 per million input tokens and $1.20 per million output. GPT-5.6 Terra fell 20 percent to $2 and $12. A Fast mode for GPT-5.6 Sol offers up to 2.5 times standard processing speed at twice the price with identical model quality.
The efficiency claims are more specific than the usual framing. OpenAI says GPT-5.6 Sol worked alongside its engineers to optimize the production serving software, cutting end-to-end serving costs by 20 percent, and improved speculative decoding for a token-generation efficiency gain above 15 percent. Separately, changes to retained reasoning and context management lifted the model’s ARC-AGI-3 score from 13.3 percent to 38.3 percent while consuming six times fewer output tokens, with no change to the model itself.
The post also puts numbers on adoption: more than one billion active users and over two million businesses. Six months in, users send roughly 50 percent more messages per day and apply the tools to about twice as many kinds of work. Internally, agentic work through Codex accounts for 99.8 percent of OpenAI’s weekly output tokens. The argument threaded through the piece is that the right metric is cost per successful outcome rather than cost per token, since a stronger model that finishes the job can beat a cheaper one that needs retries and human cleanup.
Why it matters: The ARC-AGI-3 result is the part worth carrying into procurement conversations. A near triple in benchmark performance at a sixth of the token spend, with the model held constant, says most of the available headroom for enterprise teams right now sits in orchestration rather than model selection. That reframes a lot of vendor evaluation: if your agent stack retries constantly and reloads the same context, you are paying a tax no price cut will refund. The 80 percent Luna reduction lands in the same week Alibaba shipped a frontier-class model at $2 in and $6 out, which suggests the price floor is being set from outside rather than chosen.
🧐 Research
Ten Open Math Problems Fall to an Unreleased OpenAI Model
OpenAI published results from an internal version of Astra, described as its next major model, covering ten long-standing open problems in mathematics and theoretical computer science. The list includes new upper bounds on high-dimensional sphere packing down to the Cohn-Elkies threshold, a construction establishing the existence of non-sofic groups, a disproof of Connes’s rigidity conjecture, an exponential parallel repetition theorem for two-player quantum games, and hardness results for the closest vector problem, which underpins post-quantum lattice cryptography. Three of the results resolve numbered Erdős problems.
The compute cost was modest by frontier standards. OpenAI states the tokens required to find the solutions would run roughly $2,000 at Sol API rates. Humans then prepared the arguments into manuscripts using the same model, after which the model produced Lean certificates formalizing each proof. Narrations of the model’s reasoning process were released alongside the results.
OpenAI addressed attribution directly, acknowledging the Leiden declaration on AI and Mathematics and stating that claiming human authorship for a proof an AI system generated would misrepresent both the system’s contribution and the nature of human intellectual work. The company says it takes responsibility for correctness while crediting the mathematical arguments to the system. The release follows an earlier AI-generated disproof of the Erdős unit-distance conjecture shared in May, and a programme giving 100,000 researchers free access to OpenAI models.
Why it matters: If frontier research contributions cost less than a week of a postdoc’s salary, the constraint on this kind of work stops being money and becomes verification capacity, which is why the Lean formalization step matters more than the headline count of ten. It also puts a marker down on attribution before the norm sets: OpenAI is declining to let humans take credit, which is easy to do when the results are a marketing asset and harder when a lab is deciding whose name goes on a grant application.
🔷 WordPress
The Core AI Prompt Bar Comes to WordCamp US 2026
WordCamp US 2026 in Phoenix will host a Core AI Prompt Bar, an open booth run by the WordPress Core AI Team, the working group responsible for bringing AI functionality into WordPress. It will be staffed by team members, contributors, and volunteers on Tuesday, August 18 and Wednesday, August 19, from 9:00 AM to 5:00 PM.
The team maintains a set of projects that together form the AI layer taking shape in Core. The Abilities API gives a site a central registry describing what it can do. The MCP Adapter exposes those abilities to outside AI tools over the Model Context Protocol. The PHP AI Client is a provider-agnostic SDK for talking to models, with the WP AI Client shipping the equivalent inside WordPress. The AI plugin brings features into the admin and editing experience, and WP Bench measures how well models handle WordPress-specific tasks.
The format is deliberately loose: no talks, no stage, and explicitly not a support desk, since the Happiness Bar covers that. The team is asking for questions at any level of technical depth, feedback on what has worked or failed, and details of what people are actually building. Contributors can volunteer for a shift at the booth.
Why it matters: Open standards work best when the people building them are reachable, and the Core AI Team is putting two full days into being exactly that. The projects on the table are substantial: a registry of what a site can do, an MCP adapter that exposes it to outside tools, a provider-agnostic client so nobody is locked to one model vendor, and a benchmark to keep the whole thing honest. Getting feedback from agencies and site owners while those pieces are still forming is how the useful version gets built rather than the version that looks good in a demo.
Keep ahead of the curve – join our community today!
Follow us for the latest discoveries, innovations, and discussions that shape the world of artificial intelligence.
