Local models, verified agents and catalog fixes reach builders

Local models can now enter ChatGPT Desktop

Ollama 0.34.0 RC1 can expose local models directly inside ChatGPT Desktop on macOS. The release also adds OpenAI-compatible tool search, response compaction and image preservation across compacted turns. It is an unusually direct bridge between a familiar cloud-agent interface and inference that stays on a user-controlled machine.

The important change is not another local chat window; it is local inference entering an interface many people already use.

Worth doing: Test the release candidate on a noncritical Mac and verify tools, image turns and compaction before relying on it for production work.

Keep in mind: The integration is a release candidate and is currently limited to macOS.

Claude agents produced a machine-checked Fermat proof

Anthropic published an Apache-licensed Lean repository that it describes as the first complete machine-checked formalization of Fermat’s Last Theorem. Claude agents worked for eleven days using theorem dependency graphs, separated compilation units, search and Lean’s kernel. The result is a rare inspectable artifact for studying long-horizon agent orchestration.

The useful engineering result is the workflow: agents produced small artifacts that a deterministic kernel could reject or accept.

Worth doing: Study the repository structure and verification loop before borrowing the orchestration pattern for other long-running agent workflows.

Keep in mind: Machine checking establishes formal consistency in the pinned environment; it does not independently measure how much human guidance the project required.

Pi repairs its SDK while adding GPT-6 Astra

Pi 0.85.1 adds GPT-6 Astra for API-key and Codex-subscription users, but the operational reason to upgrade is a packaging repair. Version 0.85.0 accidentally shipped internal dependencies that broke SDK imports. The patch also corrects long prompt-cache request handling for newer GPT Responses models.

A new model gets attention, but the SDK repair is the change that prevents a broken morning for maintainers.

Worth doing: Skip 0.85.0, upgrade to 0.85.1 and run import plus long-context cache tests in the same environment used by your agents.

Keep in mind: The release notes do not provide performance or quality measurements for Astra inside Pi.

Codex makes Astra the unpinned default

Codex 0.153.4 makes GPT-6 Astra visible in the bundled model picker and selects it from the bundled catalog when no model is pinned. The hotfix also qualifies asynchronous-question guidance so agents only attempt that behavior when the current session actually exposes the required tool.

Defaults are production behavior. A picker fix matters when an unpinned workflow silently inherits it.

Worth doing: Pin the intended model in controlled workflows and retest sessions that depend on asynchronous clarification behavior.

Keep in mind: This is a narrow hotfix; it does not claim broader changes to agent capability or safety.

OpenCode restores GPT-6 for subscription users

OpenCode 1.18.29 changes its Codex OAuth model filter to recognize integer GPT version names such as GPT-6. That small compatibility assumption had hidden Astra from the model list for OpenAI subscription users. The release restores the paid model without changing the underlying subscription or credentials.

A single parsing assumption can turn a valid entitlement into an apparently missing model.

Worth doing: Upgrade and verify the model list under the exact OAuth path used by subscription-backed development environments.

Keep in mind: The fix restores discovery and selection; it does not independently validate model behavior in OpenCode.

K2 Horizon opens six model sizes

MBZUAI released six K2 Horizon models ranging from 0.9 billion to 375 billion parameters. The family spans constrained devices through sparse enterprise inference and includes weights, code, training data and methodology. It gives builders one open architecture to compare across unusually different hardware and deployment budgets.

This is fallback coverage because it is older than the standard window, but the complete open release remains directly actionable.

Worth doing: Start with the smallest viable checkpoint and benchmark memory, throughput and task quality on the hardware you actually plan to operate.

Keep in mind: Capability claims come from the releasing institution and still need independent reproduction across the six sizes.

No specialist filler today.

Nothing beyond the general briefing cleared the bar for this role.