Limited access now, a migration deadline in November
Nine documented changes from OpenAI, Anthropic and xAI, checked against the vendors' own pages on September 7, 2026. Most are gated, dated, or both.
Key takeaways
- OpenAI introduced GPT-6 Astra on September 3, 2026 to a limited set of organizations, and says Astra is not yet generally available.
- Anthropic launched Claude Fable 5.1 on September 1, 2026 with a 1M token context window, 128k max output tokens, and $10 / $50 USD per MTok pricing.
- Anthropic says tool_choice types any and tool return a 400 error on Claude Fable 5.1 and Claude Mythos 5.1, so forced tool calls must be rewritten before migration.
- xAI retires grok-imagine-image-quality on November 2, 2026, after which requests are served by grok-imagine-image-2.0 at low quality with no change to the request or response shape.
- Anthropic's ant CLI 1.30.0, released September 3, 2026, manages agents, environments, skills, memory stores and deployments from repository files and writes a claude-lock.json lockfile.
Review
Nothing this week changes a default you already rely on. Almost everything published by OpenAI, Anthropic and xAI in the first days of September was gated behind eligibility, marked beta, or dated for a future switch, which makes this a week for checking calendars and access lists rather than migrating anything. OpenAI's GPT-6 Astra is the headline and the clearest example: it went to a limited set of organizations, it is not generally available, and OpenAI says its safety monitoring can pause or stop a conversation it thinks misread your instructions. That last detail matters more than the capability claims for anyone planning unattended work. Anthropic's release is the one with real migration cost attached. Claude Fable 5.1 arrives at the same headline price as Fable 5 with cheaper cache reads, so the model looks like a straight upgrade until you read the two entries underneath it: tool_choice types any and tool now return a 400 error, and thinking blocks cannot be replayed into older models. Both are the kind of change that passes a demo and fails in production at 2am. Audit your tool-calling code before you point traffic at it. xAI supplied the only hard deadline. On November 2 the grok-imagine-image-quality slug retires and requests get served by grok-imagine-image-2.0 at low quality, with the request and response shape unchanged. Silent substitutions are the dangerous kind, because no error surfaces and no alert fires; your images simply come back different. Put the date in the calendar this week. The pattern worth naming is that vendors are shipping controls rather than capability: named external viewers on ChatGPT Sites, read-only healthcare apps that stay away from patient charts, resource files and a lockfile for Claude deployments. That is what a market looks like when buyers have started asking who approved this and what happens when it goes wrong. Our read: spend the hour on the November date and the tool_choice audit, and leave Astra alone until it is generally available.
Dated events and sources
OpenAI introduced GPT-6 Astra to a limited set of organizations
Fact: On September 3, 2026, OpenAI introduced GPT-6 Astra for coding, research, computer use, and complex multi-step work, with access rolling out to a limited set of organizations and Astra not yet generally available.
Nova3 operational read: Treat GPT-6 Astra as a gated evaluation path, not a production baseline. If your organization gets access, test document, spreadsheet, presentation, and computer-use workflows with review steps, because OpenAI says a conversation may be paused or stopped when its safety monitoring detects a possible misreading of your instructions.
Uncertainty: Pricing, the list of eligible organizations, region limits, and the general availability date are not established.
OpenAI ChatGPT release notes: Introducing GPT-6 AstraOpenAI added Zendesk and OneNote plugins in beta
Fact: On September 3, 2026, OpenAI added the Zendesk and OneNote plugins to the Plugin directory for ChatGPT and Codex, with Zendesk using each member's own Zendesk account and OneNote covering note search, summaries, decisions, action items, and supported create or update actions.
Nova3 operational read: Support teams can trial Zendesk workflows without widening anyone's existing Zendesk access. Knowledge teams should confirm which OneNote actions are supported before allowing create or update use.
Uncertainty: Pricing, region limits, tenant eligibility, and whether the beta label changes production support terms are not established.
OpenAI ChatGPT release notes: Zendesk & OneNote Plugins in ChatGPT and Codex [Beta]OpenAI shipped Healthcare Public Data for clinicians in the US
Fact: On September 1, 2026, OpenAI made Healthcare Public Data in ChatGPT available to eligible ChatGPT for Clinicians users in the United States, with nine read-only apps for public healthcare sources that do not access patient charts.
Nova3 operational read: Treat the plugin as a public-source lookup tool, not a patient-record tool. Write usage guidance that repeats OpenAI's own instruction to keep protected health information out of searches sent to public sources.
Uncertainty: Pricing, availability outside the United States, the exact eligibility criteria, and whether the feature is preview are not established.
OpenAI ChatGPT release notes: Healthcare Public Data in ChatGPT for CliniciansAnthropic's ant CLI 1.30.0 added resource files and a lockfile
Fact: On September 3, 2026, Anthropic released version 1.30.0 of the ant CLI with ant apply, which creates and updates agents, environments, skills, memory stores, and deployments from files in a repository and writes a claude-lock.json lockfile.
Nova3 operational read: Platform teams get a change-controlled way to manage Claude resources from repository files and CI. Commit the generated lockfile so later runs update the same resources instead of creating duplicates.
Uncertainty: Pricing, region limits, account eligibility, and whether ant apply is preview are not established.
Anthropic release notes: Version 1.30.0 of the ant CLI adds ant applyAnthropic launched Claude Fable 5.1 and Claude Mythos 5.1
Fact: On September 1, 2026, Anthropic launched Claude Fable 5.1 and Claude Mythos 5.1 with a 1M token context window by default, 128k max output tokens, always-on adaptive thinking, pricing of $10 and $50 USD per MTok, and cache reads cut to $0.25 per MTok.
Nova3 operational read: Benchmark Fable 5.1 against your current prompts and cache patterns before moving long-context coding or research work onto it. Model the cost in three parts, because the cache-read price changed while input and output prices did not.
Uncertainty: General availability and region limits for Claude Mythos 5.1 outside Project Glasswing participants are not established.
Anthropic release notes: Claude Fable 5.1Two tool_choice types now return an error on the 5.1 models
Fact: On September 1, 2026, Anthropic said tool_choice types any and tool are not supported on Claude Fable 5.1 and Claude Mythos 5.1 and return a 400 error, while auto and none are unchanged.
Nova3 operational read: Audit tool-calling requests before you move traffic to either 5.1 model, because a forced tool call that works today will fail outright. Anthropic points to strict tool use or structured outputs when you need schema-conformant tool inputs.
Uncertainty: Whether Anthropic will later support any or tool on these models, and the rollout timing per platform, are not established.
Anthropic release notes: tool_choice on Claude Fable 5.1 and Claude Mythos 5.1Thinking blocks stopped being portable across older Claude models
Fact: On September 1, 2026, Anthropic said thinking blocks produced by Claude Fable 5.1 and Claude Mythos 5.1 are preserved only for the model that produced them or a newer one, and that earlier models cannot read them.
Nova3 operational read: Do not assume a stored thinking block can be replayed into an older Claude model. Cover prompt, tool, and earlier-message changes in regression tests for any workflow that replays thinking blocks.
Uncertainty: Pricing impact, platform-specific rollout timing, and long-term support for the related beta header are not established.
Anthropic release notes: Thinking blocks on Claude Fable 5.1 and Claude Mythos 5.1xAI dated the retirement of grok-imagine-image-quality
Fact: On September 2, 2026, xAI said grok-imagine-image-quality is retired on November 2, 2026, and that requests to the slug will then be served by grok-imagine-image-2.0 with quality set to low, with no change to the request or response shape and a lower per-image price.
Nova3 operational read: Inventory every call to grok-imagine-image-quality and test the output you get from grok-imagine-image-2.0 at low quality against what your product expects. The request shape surviving the switch is the reason this one gets missed: nothing breaks, the pictures just change.
Uncertainty: Exact per-image prices, region limits, and whether any compatibility period exists after November 2, 2026 are not established.
xAI release notes: grok-imagine-image-quality retirement on November 2Questions this edition answers
Should an operations team plan production workflows around OpenAI GPT-6 Astra now?
No. OpenAI said on September 3, 2026 that access is rolling out to a limited set of organizations and that Astra is not yet generally available. Teams that do have access should validate document generation, computer-use steps, and review behavior before depending on it, since OpenAI says a conversation may be paused or stopped when its safety monitoring flags a possible misreading of instructions.
What should Claude Fable 5.1 buyers check before migrating?
Check token costs, cache-read economics, output limits, tool-choice behavior, and thinking-block replay. Anthropic said on September 1, 2026 that Fable 5.1 supports a 1M token context window, 128k max output tokens, and $10 / $50 USD per MTok with cache reads at $0.25 per MTok. The same release says tool_choice types any and tool return a 400 error on Fable 5.1 and Mythos 5.1, and that thinking blocks are preserved only for the producing model or a newer one.
Does OpenAI Healthcare Public Data in ChatGPT for Clinicians access patient charts?
No. OpenAI says the nine apps are read-only and do not access patient charts. The feature became available on September 1, 2026 to eligible ChatGPT for Clinicians users in the United States, covering public sources such as biomedical research, clinical trials, medication information, Medicare data, and provider records. OpenAI also instructs users not to include protected health information in searches sent to public sources.
What is the xAI grok-imagine-image-quality migration deadline?
November 2, 2026. xAI said on September 2, 2026 that requests to the retired slug will be served by grok-imagine-image-2.0 with quality set to low, at a lower per-image price, with no change to the request or response shape. Because nothing errors, test your image output expectations and cost assumptions before the date rather than after.
Editorial method
Current State is a weekly review. Nothing appears here without a source check, an explicit event date, and an editor’s sign-off.
/current-state.md mirrors the latest reviewed edition.