Current State: Operations briefing for the week of 2026-09-28
Anthropic changes refusal billing and compliance data handling; OpenAI adds account security history and expands Voice workflows. The dated updates in this briefing precede the edition week.
Key takeaways
- Anthropic's 2026-09-24 refusal billing change means a Claude request can incur model-rate charges before producing any output for the bio, frontier_llm, and reasoning_extraction categories.
- Anthropic's 2026-09-24 Compliance API Activity Feed change requires audit integrations to handle missing names and titles on both new and historical activities.
- OpenAI's 2026-09-23 Voice update preserves existing app permissions and usage limits, and Voice in Work requires both Voice and Work access.
- Anthropic launched Claude Opus 5.5 on 2026-09-22 with a default 1M-token context window and 128k maximum output tokens, but the supplied evidence contains no comparative performance benchmarks.
Review
The supplied evidence contains no announcements explicitly dated within the week starting 2026-09-28. This edition carries forward selected OpenAI and Anthropic updates from 2026-09-22 through 2026-09-25 rather than presenting them as new announcements for the edition week. Nova3's priority is integration maintenance: Anthropic's refusal billing changes request costs, while its Compliance API Activity Feed removes names and titles even from older activities. Claude Opus 5.5 warrants a measured evaluation, not an automatic migration. Older Google entries and undated Perplexity entries are excluded; xAI's September announcement dates lack an explicit year in the supplied evidence.
Dated events and sources
Security history in ChatGPT
Fact: On 2026-09-25, OpenAI introduced Security history in ChatGPT on the web under Settings > Security and login, showing sign-ins, sign-outs, and changes to MFA, passkeys, and other security settings with time, location, and device details that may be approximate or unavailable.
Nova3 operational read: Add Security history to the account-review procedure for unfamiliar OpenAI activity. Treat location and device details as supporting evidence rather than definitive attribution.
Uncertainty: The supplied evidence does not establish plan eligibility, regional limits, or availability outside ChatGPT on the web.
OpenAI ChatGPT release notesBilling for refusals before any output
Fact: On 2026-09-24, Anthropic resumed billing across all platforms for refusals before any output when stop_details.category is bio, frontier_llm, or reasoning_extraction at the rates of the model used, while other categories before any output remain unbilled, mid-stream refusals remain billed, and fallback credit is unchanged.
Nova3 operational read: Update Claude cost reporting to distinguish billed refusal categories from unbilled categories. Review retry behavior so repeated refusals do not create overlooked charges.
Uncertainty: The refusal notice does not provide a model-by-model rate table or quantify the expected billing impact.
Anthropic release notesCompliance API local session endpoints out of beta
Fact: On 2026-09-24, Anthropic moved the Compliance API local session endpoints out of beta for Claude for Microsoft 365 sessions in Excel, PowerPoint, Word, and Outlook, covering product_surface values beginning with office_agents.
Nova3 operational read: Review Microsoft 365 audit coverage against the office_agents product surfaces. Validate local session ingestion before treating the beta exit as proof of complete session capture.
Uncertainty: The supplied evidence does not establish pricing, account eligibility, or regional availability for the Compliance API local session endpoints.
Anthropic release notesCompliance API Activity Feed removes names and titles
Fact: On 2026-09-24, Anthropic changed the Compliance API Activity Feed so filename and title fields on file, project document, and artifact activities are always empty or omitted, including historical activities, with name or title lookup by activity ID requiring a Compliance Access Key with the read:compliance_user_data scope.
Nova3 operational read: Test audit pipelines and reports against empty or omitted filename and title fields, including historical reprocessing. Approve the read:compliance_user_data scope only for integrations that need name or title lookups.
Uncertainty: The supplied evidence does not establish pricing or regional restrictions for the required lookup access.
Anthropic release notesCache diagnostics out of beta on the Claude API
Fact: On 2026-09-23, Anthropic moved Cache diagnostics out of beta on the Claude API, made the cache-diagnosis-2026-04-07 beta header optional while preserving existing header behavior, and made POST /v1/messages responses always include a diagnostics field that is null unless the request includes the diagnostics object.
Nova3 operational read: Update Messages response parsers to accept diagnostics: null and opt in on requests used for cache investigation. Remove the beta header on a normal maintenance schedule rather than treating removal as urgent.
Uncertainty: The supplied evidence does not establish separate diagnostics pricing or regional restrictions.
Anthropic release notesUse plugins in Voice and get work done by speaking
Fact: On 2026-09-23, OpenAI announced plugin support during Voice conversations on web, iOS, and Android and Voice in Work on web and mobile, with Free and Go users limited to plan-supported plugins in Chat, Voice in Work requiring both Voice and Work access, existing connections, permissions, and usage limits preserved, and unfinished Work tasks able to continue in text after a call ends.
Nova3 operational read: Check Voice and Work entitlements before assigning spoken workflows to staff. Test the transition from a Voice call to text for unfinished Work tasks, and retain existing connected-app permission reviews.
Uncertainty: The supplied evidence does not establish regional availability, numeric usage caps, or pricing for Voice and Work access.
OpenAI ChatGPT release notesClaude Opus 5.5
Fact: On 2026-09-22, Anthropic launched Claude Opus 5.5 (claude-opus-5-5) with a default 1M-token context window, 128k maximum output tokens, always-on adaptive thinking, and $4 / $20 USD per million tokens compared with $5 / $25 for Claude Opus 5, with availability on the Claude API, Claude in Amazon Bedrock, Claude Platform on AWS, Claude on Google Cloud, and Claude in Microsoft Foundry.
Nova3 operational read: Evaluate claude-opus-5-5 against current coding or knowledge-work workloads and measure actual request costs before changing defaults. Include always-on adaptive thinking in migration testing.
Uncertainty: The supplied evidence does not establish regional availability, platform-specific pricing differences, or comparative performance benchmarks.
Anthropic release notesGPT-6 Sol and Luna in Work and Codex
Fact: On 2026-09-22, OpenAI introduced GPT-6 Sol and GPT-6 Luna in ChatGPT Work and Codex as models separate from Chat models, with model availability and reasoning effort options dependent on plan and workspace settings and Codex desktop retaining the manually selected model.
Nova3 operational read: Inventory Work and Codex model settings separately from Chat settings. Check manually selected Codex desktop models before assuming staff have moved to GPT-6 Sol or GPT-6 Luna.
Uncertainty: The supplied evidence does not establish pricing, a plan-by-plan availability matrix, regional limits, or preview status.
OpenAI ChatGPT release notesQuestions this edition answers
Does Anthropic charge for Claude requests that are refused before any output?
Anthropic resumed billing on 2026-09-24 for refusals before any output in the bio, frontier_llm, and reasoning_extraction categories, at the rates of the model used. Other refusal categories before any output remain unbilled, while mid-stream refusals were already billed. The change applies on all platforms, and fallback credit is unchanged.
Why are file names and artifact titles missing from the Claude Compliance API Activity Feed?
Anthropic removed names and titles from the Compliance API Activity Feed on 2026-09-24, including activities recorded before the change. The filename and title fields are now empty or omitted. Looking up a name or title by activity ID requires a Compliance Access Key with the read:compliance_user_data scope.
Should an operations team migrate to Claude Opus 5.5 immediately?
Nova3 recommends evaluating Claude Opus 5.5 on existing workloads before migrating. Anthropic's 2026-09-22 release specifies a default 1M-token context window, 128k maximum output tokens, always-on adaptive thinking, and $4 / $20 USD per million tokens compared with $5 / $25 for Claude Opus 5. The supplied evidence does not establish comparative task performance or regional availability.
Can Free and Go users use ChatGPT Voice with plugins?
OpenAI's 2026-09-23 release says Free and Go users can use Voice in Chat with the plugins their plans support. Voice in Work requires both Voice and Work access. Existing app connections, permissions, and usage limits still apply, so the announcement does not establish unrestricted plugin access.
Editorial method
Current State is a weekly review. Nothing appears here without a first-party source and an explicit event date.
/current-state.md mirrors the latest edition.