With nobody releasing new weights, the frontier labs spent the day shipping around the model rather than shipping one. OpenAI did the most, launching a separate ChatGPT for Teens gated by an age-prediction system, widening its ad platform to 31 European markets and partnering with CodeAI to reach classrooms. Anthropic countered with a genuine capability result, showing Claude autonomously drive existing tools to design working protein binders for 14 of 15 targets. Underneath, the argument was increasingly about unit economics: Cerebras claimed 30 times GPU inference speed with its new CS-4 the same day Perplexity published a cost-per-task figure for a US-hosted DeepSeek, and Block open-sourced a workspace for running agents across models. The through-line is that when the model stops being the differentiator, the fight moves to the surface where work starts and the speed and price of running it.
ChatGPT for Teens arrives, and age prediction decides who gets it
This matters even if there is no teenager in your house, because it is the clearest signal yet that OpenAI will segment ChatGPT by user rather than by plan. Anyone whose stated age is 13 to 17, or who the age-prediction system estimates is under 18, is routed automatically into a separate experience built around Study Mode, quizzes, learning visualisations, homework reminders and Study Hours that parents can schedule. Protections covering self-harm, eating disorders, violence and sexual content are on by default, and the updated under-18 model spec now bars romantic language and anything that encourages emotional dependence.
The practical consequence for everyone else is that age inference is now running quietly inside a product used by hundreds of millions of people. If you build on the API, or run ChatGPT in a school or family setting, it is worth understanding how that routing decision gets made.
ChatGPT Ads reaches 31 European markets next week
Six months after the United States pilot, advertising inside ChatGPT is arriving in Germany, France, Spain, Italy, the Netherlands, Austria and the Nordics among others, building on the European rollout flagged at the weekend. Ads appear only for Free and Go users, so Plus, Pro and Enterprise stay clean, which quietly turns the paid tier into an explicitly ad-free product rather than just a capability upgrade. Access runs initially through OpenAI's ads team and agency and technology partners, with self-service through Ads Manager promised later in the summer.
For marketers this is the first serious route to European users at the point where they are comparing options rather than typing keywords. For everyone else it is a reminder that the economics of free frontier AI are now underwritten by an ad platform with geo-targeting, custom audiences and a conversions API.
Claude designed working protein binders for 14 of 15 targets
Anthropic published lab results in which Claude, given a single design prompt written by a human expert, autonomously drove existing protein-design and co-folding tools to produce binders against 14 of 15 targets. Adaptyv Bio and Twist Bioscience built and tested the designs independently, and between 22 and 35 per cent bound successfully against a field-typical rate of 10 to 15 per cent, with at least four targets matching or beating the best reported affinity.
The reason this is more interesting than another benchmark score is that Claude was not a new specialist model, it was operating tools scientists already have. Anthropic also notes that life-science tasks remain blocked in its most capable model and that an access programme for scientists is being prepared, so the capability is real but deliberately gated. If you work anywhere near computational biology, the gating decision is the thing to watch rather than the hit rate.
Perplexity Computer now takes instructions over email
You can now send, forward or copy [email protected] on any email thread and the request runs as a normal Computer session, viewable on web and mobile with the same audit trail as any other task. That is a small interface change with a large practical effect: it removes the need to open the app at all, and it makes an agent addressable by anyone on a thread rather than only the account holder.
Email is still where most business context lives, so forwarding a supplier quote or a contract chain and getting a completed task back is a genuinely different workflow from pasting text into a chat window. Test it carefully before pointing it at anything sensitive, because forwarding a thread hands over everything in it.
DeepSeek V4 Pro, hosted in the US, lands in Perplexity Computer
Perplexity has added a US-hosted instance of DeepSeek V4 Pro and published its own evaluation: 0.359 on WANDR at 0.75 US dollars per task, which it says is 62 per cent cheaper than the next model on the cost-performance frontier. The hosting location is the point for most teams, because it removes the main objection that has kept DeepSeek out of Western procurement processes.
If you have been watching the open-weight cost curve but could not get past the data-residency question, this is the first easy way to try V4 Pro on real agentic work without standing up your own hosting.
Cerebras launches the CS-4 and claims 30 times GPU inference speed
The CS-4 is the first system built on Cerebras' new Nexus architecture, packing three WSE-3 Turbo wafers per rack on TSMC's 5nm process, with wafer-to-wafer interconnect latency down to roughly 2 microseconds and 50 per cent fewer components than the CS-3. Cerebras claims double the speed of the previous generation, up to ten times the throughput per watt, and more than 1,000 tokens per second on 10-trillion-parameter models.
Vendor benchmarks always deserve scepticism, but the direction of travel matters: inference speed is becoming the axis on which agentic workloads are won, because a coding agent that runs ten times faster is a different product rather than a better one. It is the same bet OpenAI made routing GPT-5.6 Sol through Cerebras for its Ultrafast mode. First shipments begin this quarter.
Block open-sources Berd, a desktop workspace for running agents
Berd is the app Block's own teams use internally, now on GitHub under Apache 2.0 with builds for macOS, Windows and Linux. It is built on Tauri 2 and React 19 rather than Electron, connects to Block's Goose framework over the Agent Client Protocol, and keeps conversation history on your machine instead of in the cloud. The genuinely useful part is that it is harness-agnostic, sitting above Goose, Claude Code and Codex rather than locking you into one of them.
Block is not accepting external pull requests, so treat it as source-available in practice, though the licence lets you fork it for internal use. If your team is juggling three agent command-line tools and losing track of sessions, this is the most credible open answer so far.
Industry themes
The frontier labs shipped around the model rather than shipping a model: teen segmentation, an ad platform across 31 markets, and an email entry point for an agent. When nobody is releasing weights, the competition moves to who owns the surface where the work starts, which is why a routing decision and an inbox address were bigger news today than any benchmark.
Inference speed and cost per task are quietly becoming the real battleground. Cerebras claimed 30 times GPU throughput on the same day Perplexity published a cost-per-task figure for a US-hosted DeepSeek V4 Pro, and both are arguments about unit economics rather than capability. For buyers, that is the more useful lens right now: not which model wins a leaderboard, but which one is cheap and fast enough to run at the scale you need.
The two most practically useful agent stories of the day, Perplexity's email entry point and its DeepSeek hosting, appeared on the company's own X account hours ahead of any press coverage, and Anthropic's protein results and Hugging Face's Hub milestone followed the same pattern. That is a reasonable argument for keeping the social scan ahead of the news sweep.