Access Has Improved, but the Barriers Have Not Disappeared

This weekly review brings together projects moving in different directions: StepFun opened the weights for Step 5 Preview, Perplexity described its SPACE sandbox, Stack Overflow for Agents added a ChatGPT plugin, and Itaú plans to bring generative AI into its banking services. They do not address the same technical problem, but they meet the same engineering boundary: getting a model into a real workflow depends on deployment conditions, permission controls, and result verification—not just the model itself.

The useful way to read these cases is as a survey of what comes after access. Open weights let teams deploy a model themselves, but they also bring hardware budgets and license terms into the decision. An open-source sandbox offers inspectable isolation, but still requires suitable hosts and ongoing maintenance. A plugin shortens the path to a product, but cannot prove retention or business value. Access friction is falling; responsibility for deployment is not.

Open Weights Still Come With a Concrete Deployment Bill

Step 5 Preview is a 600-billion-parameter sparse MoE, activating 27 billion parameters per token and supporting a one-million-token context. Sparse activation and a long context window may look like entries on a capability sheet, but the self-hosting requirements bring the question back to infrastructure: the supplied material says FP8 requires at least four B300s or eight H200s, with higher requirements for BF16. “Open” does not mean that most teams can run it in their own services cheaply.

A benchmark score cannot stand in for a deployment decision either. The reported Terminal-Bench 4.0 score is 33.3%, but StepFun used the High reasoning setting while competitors used Max, so the comparison is not like-for-like. The license also restricts some commercial uses. For a technical leader, the useful question is not whether to chase the score, but whether the hardware and license fit the intended workflow, followed by an evaluation of quality, latency, and cost under comparable tasks and reasoning settings.

A Sandbox Depends on Egress Control as Much as Isolation

SPACE places isolation outside the model. It is an Apache 2.0-licensed sandbox platform written in Rust, with each task running in its own Firecracker microVM, complete with a separate kernel and writable filesystem. Credentials stay out of the guest, while a host-side gateway controls all outbound connections. The design turns “what can the agent do?” into rules enforced by infrastructure, rather than relying on the model to obey a prompt every time.

The more instructive part for engineering teams is how the project responded to policy bypasses. Red-team testing found that DNS spoofing and shared CDN IPs could evade some network policies. Perplexity then added source-address validation, terminated TLS at the gateway, and checked both SNI and the HTTP Host header. The supplied figures show production creation latency falling from a median of 185 milliseconds to 60, and from 447 to 89 milliseconds at P90. The reported escape testing found zero instances in which host secrets were reached across 108 runs, but the evaluation covered only nine models, 216 runs, two tasks, and four settings. That is a result under specific conditions, not a security certification. Self-hosting also requires Linux x86_64, KVM, and btrfs, and does not support macOS.

A Smoother Entry Point Does Not Prove Demand

Stack Overflow for Agents makes reviewed organizational knowledge available to agents. When an agent cannot find an answer, it can draft a question, a TIL, or a reusable design pattern for human review and publication; feedback also records the conditions under which an answer applies. The ChatGPT plugin lets users discover and set up the product through an existing interface, reducing steps compared with traditional MCP configuration. The design is not simply about making an agent say more. It is about retaining review and reuse processes after knowledge becomes available to the agent.

The growth figures support a narrower claim: adoption friction may have fallen. The material says sign-ups tripled overnight after the plugin launched, but gives neither the absolute number of sign-ups nor retention. The estimate that agent creation rose by as much as tenfold comes from an interviewee and measures something different from registrations. A similar boundary applies to Itaú: an initial group of roughly 300,000 customers and a goal of reaching the full customer base by year-end are expansion plans, not evidence that broad adoption has already occurred. Teams should distinguish entry-point conversion, ongoing use, and verifiable work outcomes rather than treating one growth figure as proof of the whole chain.

Integration Means Encoding Constraints Into the Workflow

The practical lesson for technical leaders is to change the unit of evaluation from “which model should we choose?” to “which part of the workflow can we constrain and inspect?” For an open-weight model, check hardware requirements, license terms, and benchmarks run under comparable settings. For an execution environment, keep credentials outside the guest, enforce egress at the gateway, and test whether network policies can be bypassed. For a knowledge agent or conversational interface, track actual use, human review, and whether the intended business outcome is delivered as separate measures.

This does not make model capability or openness irrelevant. It means neither can answer production questions on its own. Different tasks impose different requirements on cost, latency, reasoning quality, and accountability, and the supplied material does not provide like-for-like comparisons across these projects. A practical decision is to begin with a narrow task and a clearly accountable owner, define reproducible verification conditions and a stopping rule, and only then consider expansion. Connection is the starting point; auditability, verification, and responsibility for failure are what make it part of a workflow.