Sunday's digest covers a default model switch at Anthropic that comes with a prompting etiquette change, an unusual public release of hundreds of AI-generated mathematical results, Meta buying its way into coding agent orchestration, and a memory engine that halves DRAM requirements while keeping 99 percent of server performance. Two developer-facing launches round out the day: cloud simulators built for agents, and a no-code platform for voice agents.
1. Anthropic makes Claude Opus 5.5 the default model on all paid plans
Anthropic has switched Claude Opus 5.5 to be the default model across its Pro, Max, Ultra, and Team plans, and alongside the switch published a knowledge-work guide arguing that many users are holding the model back with prompting habits carried over from earlier generations. The top recommendation is to stop using phrases like "think hard" or "think carefully". According to Anthropic, Opus 5.5 reasons continuously on its own and scales its reasoning depth to the task and the selected effort level, so explicit nudges toward deeper thinking are dead weight; in internal tests, responses without those phrases started faster with no measurable drop in quality.
The second recommendation is to start at Medium effort, the new default. Anthropic says Opus 5.5 at Medium matched the previous Opus 5 running at High in its internal tests, and that higher settings cost more without always producing better results. The guide also pushes teams to rethink the common split of using the cheaper Sonnet for daily work and reserving Opus for complex projects: a more capable model can finish a task in fewer steps, and Anthropic notes that anyone running Sonnet at Extra High or Max effort ends up paying about as much as with Opus 5.5.
The rest of the advice treats the model as something to delegate to rather than guide turn by turn. Describe the goal, output format, and checkpoints in one message instead of drip-feeding instructions; attach diagrams and tables as images rather than transcribing them (Anthropic claims Opus 5.5 reads dense charts more accurately at its lowest effort level than Opus 5 did at its highest); request the finished document directly since the model often goes from request to final artifact in a single pass; name unwanted design styles specifically instead of vaguely banning "generic design"; and connect files and systems through connectors so the model can pull what it needs. All the benchmark comparisons come from Anthropic's own testing, and no scores were published alongside the guide, so teams weighing a switch should measure token use on their own workloads.
— Anthropic · The Decoder
🔗 Anthropic · The Decoder
2. OpenAI releases 372 mathematical result families produced by an unreleased model
OpenAI has posted 372 mathematical result families and 722 manuscripts to a public GitHub repository, claiming significant progress on some of the field's most enduring open problems. The results were produced by an unreleased internal frontier model, and the headline claims include a solution to the four-dimensional Kakeya conjecture and progress toward the Riemann hypothesis. According to OpenAI, the average result required roughly three hours of compute equivalent to a ChatGPT Pro subscription, and many of the proofs were checked with Lean, the formal proof language that verifies an argument step by step.
The methodology is the part mathematicians are watching. This release follows an earlier OpenAI claim of a Navier-Stokes result produced by an "agentic swarm" of 10,000 agents at a cost of millions of dollars. In the new batch, nearly every paper is attributed to a single agent responding to a single prompt, which, if verified, would mark a genuine shift in how mathematical discovery gets done. OpenAI formed an Advisory Group on Mathematics and Artificial Intelligence (AGMAI) to set publishing guidelines for AI-generated mathematics, though the company has not fully followed every recommendation, including disclosing the specific prompts and compute times for each problem.
The reception is cautious rather than celebratory. Daniel Litt of the University of Toronto has argued there is no reason to keep the underlying mechanisms secret if the results are correct, and Andrew Sutherland of MIT says claims about single-agent problem solving should be treated as unverified until OpenAI releases the model and the community can replicate the work independently. The practical test now runs through the repository itself: 722 manuscripts for working mathematicians to sort into genuine results, tightened restatements, and errors.
— OpenAI · Engadget
🔗 GitHub repository · Engadget
3. Meta acquires Conductor, a startup for orchestrating multiple coding agents
Meta confirmed it has acquired Conductor, an AI coding startup whose platform lets developers deploy and manage multiple coding agents in parallel. Conductor was founded in 2024 and is led by CEO Charlie Holtz and CTO Jackson de Campos. Its product gives each agent its own compartmentalized workspace, running across cloud, local, or mobile environments, so engineers can push different parts of a project forward at the same time without the agents trampling each other. Financial terms were not disclosed.
The deal fits a clear pattern at Meta. The company launched Muse Code in August 2026 and previously acquired the agent platform Moltbook, and buying an orchestration layer now points at the same goal: controlling the tools developers use every day as a way to capture enterprise spending and gather the feedback loop that improves the underlying models. Meta has also been explicit that coding and mathematics proficiency are levers for frontier capability, so an orchestration system that shows how models interact with real codebases is both a product and a training signal.
The broader consolidation context matters too. On the same day as the Conductor announcement came reports that OpenAI plans to acquire Pierre Computer, following its June acquisition of Ona, and earlier this month OpenAI closed the Astral deal that brought uv and Ruff into its toolchain. Holtz has promised "big, crazy plans for the next interface" without sharing a roadmap, and Meta has not yet detailed what happens to Conductor's standalone cloud service or its existing users during the transition.
— Meta · Axios
4. AMD details MEXT, a memory engine that halves DRAM while keeping 99 percent of performance
AMD has shared details of MEXT, a predictive memory engine built to attack one of the datacenter's most expensive bottlenecks: DRAM capacity. The system uses transformer models and flash storage to predict which data a server will need, keeping hot data in DRAM and moving the rest to flash. AMD says servers running MEXT maintain up to 99 percent of standard performance while using half the DRAM, and the technology is already active on 145 servers, with a deployment plan that scales to 2,000 servers by 2027.
The problem it targets is structural. AI deployments have pushed memory demand faster than DRAM densities have grown, and memory now takes a large share of server cost in inference fleets. Approaches that decouple performance from raw memory volume change the economics of building AI infrastructure, because they let operators buy fewer DRAM channels and spend the difference on accelerators or power. Tom's Hardware, which first reported the deployment numbers, frames MEXT as AMD's answer to the memory squeeze that has accompanied the GPU buildout.
There are limits worth naming. A predictive engine is workload-dependent: workloads with stable access patterns should compress gracefully, while genuinely random access patterns will fall back to flash more often and eat into that 99 percent figure. And since flash latency sits well above DRAM latency, the prediction accuracy effectively becomes the performance contract. Still, if the 145-to-2,000 server rollout holds, MEXT gives cloud operators a lever they currently lack, and it lands at a moment when every dollar of infrastructure cost is under scrutiny.
— AMD · Tom's Hardware
🔗 AMD · Tom's Hardware
5. NVIDIA and Hugging Face publish a full walkthrough of an agent-run robot pick-and-place task
NVIDIA published a technical walkthrough on the Hugging Face blog showing how to build a robot pick-and-place simulation in Isaac Sim 6.2 using assets from SimReady Central, the open library of OpenUSD simulation-ready assets hosted on Hugging Face. The example assembles a UR17G arm, a Robotiq 2F-140 gripper, and a beverage can, then has an AI agent generate and execute the task code. In the published example the agent is GPT-6 Sol, and NVIDIA notes other agents can substitute as long as they can load the required skills and run Python against Isaac Sim.
The part that makes this reusable rather than a one-off demo is the validation layer. SimReady assets carry explicit simulation profiles, and the guide relies on the Robot-Body 4.0.0 and Robot-Gripper 3.0.0 validation files, which pass PhysX driven-joint checks, so developers reuse certified assets instead of re-tuning joint physics and collisions by hand. Three pre-run checks come before any motion: joint settings, a small arm movement with gripper open and close, and a collision check confirming the can rests stably on the work surface. The grasp itself follows an approach, grasp, lift, hold, transfer, release sequence, with a 2 newton-meter drive-effort cap on the gripper chosen through simulation tests.
The walkthrough is equally explicit about its boundaries: fixed starting positions, no vision, and scripted motion without replanning. That honesty is what makes it a useful baseline for the Physical AI workflow it demonstrates, where an agent handles asset assembly and code generation while the simulation layer verifies physics before anything runs. For teams building embodied systems, the deliverable is a repeatable pipeline: certified assets, profile checks, and an agent writing control code against a validated scene.
— NVIDIA · Hugging Face
🔗 Hugging Face blog · SimReady Central
6. Expo launches cloud simulators built for AI coding agents
Expo announced Expo Simulators, a cloud service giving AI coding agents what mobile development has been missing: environments they can actually verify their work in. Agents start a session with one command, install a build, and drive the app in an iOS or Android simulator to confirm their changes behave as intended. The service returns screen recordings, screenshots, and logs as proof of work, so the human reviewing an agent's pull request can see what the agent saw. It is available now on all paid plans.
The gap it fills is easy to describe. Web coding agents can run tests and spin up preview servers, but mobile agents have been stuck writing code they cannot run, because simulators live on a developer's machine behind Xcode and Android Studio. By moving the simulator into the cloud and giving agents a command-line entry point, Expo turns "verify your own change" from a manual step into part of the agent loop. A follow-up guide published the next day covers the surrounding setup: connect a GitHub repo, add a robot user, connect the agent, and give the workflow a GitHub token.
The release lands in a week full of agent infrastructure moves, from payment rails to orchestrators, and it targets the step where coding agents still fail most often on mobile: not writing the code, but knowing whether it works. For Expo's ecosystem the bet is straightforward. If agents can close the verification loop themselves, the cost of shipping a mobile app drops, and the platform that hosts that loop becomes the default place where agent-built apps get made.
— Expo
🔗 Expo blog · Agentic workflows guide
7. Fish Audio launches Fish Agent, a no-code platform for voice agents that make calls
Fish Audio launched Fish Agent, a platform for building custom AI voice agents that place and take phone calls, in public beta and open to every Fish Audio account. The pitch is deployment speed: the median time from configuration to a working agent is 32 seconds, using the Fish Live 3 full-duplex speech model that handles interruptions and natural turn-taking in conversation. Builders configure a prompt, voice, knowledge base, and tools in the console or through a REST API, and can attach their own LLM instead of accepting a bundled one.
The pricing structure shows how Fish Audio is bundling the voice stack. The agent bundle, which includes speech recognition, speech generation, knowledge retrieval, and tool calls, starts at $0.035 per minute through December 31, 2026, against a regular rate of $0.07. Standard phone calls add $0.01 per minute with higher rates for some international destinations, and a dedicated phone number costs $1.20 per month. Premium language models add their own token charges on top. For teams used to assembling STT, TTS, orchestration, and telephony from separate vendors, the single bundle is the product.
Under the hood the platform behaves like versioned software rather than a demo console. Publishing creates an immutable configuration version, sessions are created through the API with short-lived credentials, and the browser SDK handles the live connection while the platform runs recognition, synthesis, and turn-taking server-side. Voice agents have been moving from model demos to deployable infrastructure for a while now, and a platform that treats a phone agent as something you configure, publish, and version is a reasonable snapshot of where that shift has landed.
— Fish Audio
KD Agentic · AI Daily Digest
