8 source-linked editorial stories.
Updates
1. Amp removes its monthly-plan requirement for self-supplied models
Amp removes its monthly-plan requirement for people bringing their own compute and model subscriptions or keys. The announcement says there are no Amp bring-your-own-key token fees or limits, except on its Enterprise tier.
Proof: Amp is now free to use when you bring your own compute and model subscriptions/keys.
Impact: Teams can mix free and paid members. This does not make outside model subscriptions or API usage free, and orbs remain paid. Early access to additional model routers is currently limited to Megawatt and Gigawatt members.
Watch next: Those members can enable More AI Routers and Subscriptions, then configure Model Routing in personal or workspace settings. The date for wider availability is not specified.
Canonical host: ampcode.com
Watch from this story
Read full story
Updates
2. Anthropic commits to embedded outside safety evaluators
Anthropic chief Dario Amodei proposes slowing frontier capability advances while safety work catches up. In an essay published and announced September twelfth, he commits Anthropic to bringing in independent evaluators with ongoing, employee-like access. Sam Altman says OpenAI will do the same.
Proof: Anthropic is unilaterally committing to this step now.
Impact: The essay says they should be able to publish key findings without Anthropic editorial control. This gives the public a concrete commitment to track beyond a general promise about safety. This is a commitment, not evidence that the teams are already embedded. Access would have legal, contractual and privacy exceptions, with narrow redactions for security-sensitive and other protected information. Amodei explicitly says pacing does not mean halting model training. Wider industry and global coordination remain proposals.
Watch next: Watch for the named review teams, their actual access contract terms and published findings. The announcement gives no fixed start date.
Canonical host: darioamodei.com
Watch from this story
Read full story
Updates
3. Codex GitHub review and pull-request incident is resolved
OpenAI reports that its September thirteenth Codex GitHub review and pull-request incident has fully recovered. The final recovery update is ten fifty-three UTC, before this edition cutoff.
Proof: All impacted services have now fully recovered.
Impact: The status page attributes failed review publication and pull-request creation to an upstream GitHub disruption. It records investigation, mitigation and recovery, giving affected users a way to distinguish a service incident from their own repository change. This is a resolved incident, not a claim that all Codex services are currently down. Aggregate status also does not guarantee that every individual job completed. Separate ChatGPT and model API incidents are not part of this event.
Watch next: Check the September 13 incident status and the result of any affected review or pull-request job and the linked service status before deciding whether that failed GitHub workflow needs another attempt. The page does not promise automatic replay of every failed job.
Canonical host: status.openai.com
Watch from this story
Read full story
Updates
4. Copilot metrics separate the VS Code Agents window
GitHub has added generally available usage metrics for the dedicated VS Code Agents window. Reports include active users, session counts and user messages, across one-day and twenty-eight-day reporting periods.
Proof: These metrics cover the dedicated VS Code Agents window only.
Impact: Authorized enterprise or organization administrators can use the Copilot usage metrics API after enabling its metrics policy. These fields help distinguish adoption of the Agents window from activity elsewhere in the editor. This does not count editor-window Agent Mode as the same product surface. Optional fields can be absent or null when data is unavailable; missing data is not a zero-usage result.
Watch next: Check GitHub metrics API documentation and your account permissions, then preserve missing fields when reading these reports. The announcement does not supply performance or productivity measurements.
Canonical host: github.blog
Watch from this story
Read full story
Updates
5. Copilot rereviews close only comments already addressed
Copilot code review adds automatic resolution of its own comments when a later commit addresses the feedback and Copilot rereviews it. Feedback that remains outstanding stays open.
Proof: Comments are automatically resolved when a later commit addresses the underlying feedback.
Impact: The review agent uses Copilot SDK shell tools behind its firewall to validate changes. Automatic resolution is not a blanket dismissal of all comments or proof that every defect is fixed. GitHub also reports experiments with an ensemble in Lite reviews; those results are its own experiments, not a guarantee for each repository.
Watch next: Inspect the unresolved threads and test the actual change after rereview. Use the GitHub changelog to distinguish these review changes from a new review-request workflow; the existing request process stays the same.
Canonical host: github.blog
Watch from this story
Read full story
Updates
6. Qoder CLI opens custom model connections to free individual users
Qoder says CLI version one point one point fifty adds custom model integration for all individual users, on free or paid plans.
Proof: Starting with v1.1.50, Qoder CLI supports custom model integration for all individual users—free or paid.
Impact: The announcement does not extend this statement to enterprise accounts, promise compatibility with every endpoint, or waive the external provider bill.
Watch next: Check the installed CLI version and verify the connection with the intended model. The release announcement does not provide a complete endpoint compatibility list.
Canonical host: x.com
Watch from this story
Read full story
Updates
7. Qwen nightly adds active-time and turn limits to Goals
The September twelfth Qwen Code nightly adds turn and active-time limits for Goals. The linked implementation describes settings named model dot goal Max Turns and model dot goal Max Active Minutes.
Proof: feat(goal): stop a Goal at a turn or an active-time budget by @qqqys in #11457.
Impact: Both limits are off by default. When a configured ceiling is reached, the Goal gets a wind-down turn and waits. Resuming authorizes another window for the limit that ran out. This is a prerelease. Ceilings are checked between turns, so a running turn can finish past the threshold. Active time excludes time the process was down; it is not a strict wall-clock deadline.
Watch next: Before trying the nightly, review the release settings and the separate command-hook timeout change. That change reads values below one thousand as seconds, preserves larger millisecond values for compatibility, and keeps the sixty-second default.
Canonical host: github.com
Watch from this story
Read full story
Updates
8. LangChain can record the model returned by a gateway
LangChain core version one point six point three adds an override for model-name and provider tracing metadata based on a gateway response.
Proof: release(core): 1.6.3 ( #40407 ) feat(core): Allow model name and provider tracing metadata override based on gateway response ( #40406 ) test(core): cover the deprecated.text() access path ( #40243 ) docs(core): remove stale Args/Raises entries from FileCallbackHandler._write and ChatGeneration.set_text ( #40211 ).
Impact: The release note does not show a configuration example or claim that every gateway supplies the necessary metadata. This is an SDK tracing change, not evidence of a model-quality improvement.
Watch next: Review the linked change before upgrading and verify what your gateway actually returns. A broader gateway compatibility list is not provided in the release note.
Canonical host: github.com
Watch from this story
Read full story