0167B938Public record case file · Updated Oct 1, 2026
Accountability file
A public-record case file for tracking AI model release, access, safety, pricing, status, and operator-workflow records. It separates official OpenAI and Anthropic source records from Hugin field notes, social excitement, benchmark claims, and release-date speculation.
Original editorial artwork generated for Hugin.
0167 B938 CB23Keep roles, events, and source records here to pick up later. Your list stays in this browser and can be lost if you clear site data. Copy your saved links to keep a backup in your notes.
Your reading list is empty. Use a save button on an event, source, or role dossier to add it here.
The latest dated events in this file. Open an event for its cited sources and evidence limits.
Read three days after the date, OpenAI's deprecations ledger still lists gpt-3.5-turbo-instruct, babbage-002, davinci-002 and gpt-3.5-turbo-1106 with shutdown date 2026-09-28, unchanged, and now names gpt-5.6-terra as the replacement for all four. Their section has moved from 'Upcoming deprecations' to 'Past deprecations', but its paragraph still reads 'Access to these models will be shut down on the dates below', while the section immediately beneath it, announced the same day, was rewritten to say access 'was shut down'. The two headings now sort by how recent the announcement is rather than by whether the shutdown has happened: the 2026-04-22 group sits under 'Past deprecations' with seventeen rows whose shutdown date, October 23, has not arrived, and the Sora 2 and Videos API rows have moved to Past still carrying an em dash where a replacement would go. The ledger cannot show that any endpoint stopped answering, and this desk called none.
Evidence limits. A reading of a provider's own page. The two inconsistencies are recorded as the page states them and are not resolved here; nothing in this row shows a model answering or refusing a request.
Read September 28, the Gemini API pricing page lists paid Standard rates for gemini-3.8-flash, gemini-3.7-flash and gemini-3.6-flash at $0.75 input and $3.75 output per million tokens 'through December 31, 2026', and $1.50 and $7.50 'starting January 1, 2027'. Free-tier cells read 'Free of charge'. The page footer is dated September 24. The step is not new: archived captures show it on the 3.6 and 3.7 Flash rows by mid-August and on 3.8 Flash from the first capture that lists that model, and this desk's record finds 3.6 Flash listed at $1.50 and $7.50 with no end date before mid-August, so January returns it to its launch price. Google's changelog names neither price.
Evidence limits. A published price schedule, not an observed bill; Google can change the page before the date.
Amazon's Bedrock legacy table, read September 28, lists Nova Canvas and both Nova Reel versions for end of life on September 30, 2026 with no public extended access date; its rules say requests fail on or soon after that date unless privately arranged. It lists Claude Opus 4.1 entering public extended access, at higher provider-set pricing, on October 8, 2026 with end of life January 8, 2027, and Claude Sonnet 4 ending October 14, 2026, and says its dates are specific to Bedrock. Anthropic's deprecations page, read the same day, lists both Claude models as retired on Anthropic-operated platforms, Sonnet 4 on June 15 and Opus 4.1 on August 5, and says partner-operated platforms such as Bedrock set their own schedules.
Evidence limits. Amazon's published schedule; the desk cannot observe whether the models stop answering on the stated dates.
These questions stay open in the case file.
This view starts with the roles that have the most source-record links in this file. Explore Role Web and Timeline for the full set. A line represents a shared source record; it does not establish conduct or a direct relationship.
These are the standards every anchor on this file must clear.
A model, provider, operator note, or status event appearing in a shared timeline is context only and does not prove broad availability, benchmark superiority, or product failure.
Every public-record role in this file, grouped by kind. Open a role to read its dossier — its public-record status, what it is, and the source records behind it. A shared record linking two roles is not a finding about conduct beyond that record.
10 roles linked to 98 of 173 source records. 10 records connect more than one role. Open a role to follow its records into the timeline.
10 of 10 roles · Counts describe records in this file.
Publishes the Bedrock model lifecycle tables used to verify end-of-life and public-extended-access dates for Amazon's own models and for partner models hosted on Bedrock.
1 source record · 1 linked eventRead role →Publishes Fable/Mythos restoration, model overview, pricing, and status records used to verify Fable access and terms.
19 source records · 16 linked eventsRead role →Publishes the Copilot changelog and supported-models documentation used to verify model removals and default-enablement policy dates.
4 source records · 3 linked eventsRead role →Publishes the Gemini API changelog, pricing page and deprecations table and the Agent Platform model-versions, open-model deprecations and release-note pages used to verify model retirement dates, replacements and listed prices.
9 source records · 5 linked eventsRead role →Tracks AI release clusters, official rows, and open receipt needs across the news desk.
9 source records · 9 linked eventsRead role →Publishes the Foundry model retirement schedule used to verify Azure OpenAI retirement dates and those of Microsoft's own MAI models and partner models sold through Foundry.
2 source records · 1 linked eventRead role →Publishes the changelog and models pages used to verify Labs and third-party model retirements.
2 source records · 0 linked eventsRead role →Publishes GPT-5.5, Codex, GPT-5.6 preview and general-availability, release-note, plan-guide, safety, and status records used in the model-release watch.
29 source records · 26 linked eventsRead role →Publishes the migration guides, API release notes and security FAQ used to verify API model-name retirements and new request fields.
3 source records · 2 linked eventsRead role →Public provider-status surfaces used to separate incidents, capacity, model access, and safety behavior from model-quality claims.
Hugin's July 25 desk entry joins Anthropic's Opus 5 launch record and OpenAI's reset-banking release note while holding an unsupported claim that Opus 5 itself issued a fresh reset.
Hugin's July 26 news desk separates OpenAI product records, sensitive health-product posture, and the DOJ Minnesota Medicaid guilty-plea record into different receipt lanes.
Hugin-authored news note that records the operator-stack framing and links back to primary provider receipts.
Hugin-authored news note documenting the July 5 Reddit hosted-scan recovery pass, duplicate-report cache reuse, and upstream-block posture.
Hugin-authored news note documenting the July 6 public case-index latest-activity rail and its matching content/case receipt posture.
Official Anthropic record for Fable/Mythos restored access, usage-credit transition, and safeguards context.
Help-center receipt for GPT-5.6 Sol, Terra, and Luna preview access, model IDs, pricing, caching, and availability limits.
Release-note surface for ChatGPT-facing model availability, update, and retirement language.
Official OpenAI preview post for GPT-5.6 Sol and its release-watch framing.
Dated OpenAI general-availability record for the GPT-5.6 Sol, Terra, and Luna family across ChatGPT, Codex, and the OpenAI API, with surface-specific terms and gradual rollout language.
Current OpenAI plan and workspace-controls documentation for Codex, including plan-varying usage limits, client access, plugins, and surface-specific controls.
Follow the record as it develops. Start with the latest events, or read from the beginning.
68 recorded events · Apr 23, 2026 — Oct 1, 2026
Showing 1–12 of 68 events · Latest first
Read three days after the date, OpenAI's deprecations ledger still lists gpt-3.5-turbo-instruct, babbage-002, davinci-002 and gpt-3.5-turbo-1106 with shutdown date 2026-09-28, unchanged, and now names gpt-5.6-terra as the replacement for all four. Their section has moved from 'Upcoming deprecations' to 'Past deprecations', but its paragraph still reads 'Access to these models will be shut down on the dates below', while the section immediately beneath it, announced the same day, was rewritten to say access 'was shut down'. The two headings now sort by how recent the announcement is rather than by whether the shutdown has happened: the 2026-04-22 group sits under 'Past deprecations' with seventeen rows whose shutdown date, October 23, has not arrived, and the Sora 2 and Videos API rows have moved to Past still carrying an em dash where a replacement would go. The ledger cannot show that any endpoint stopped answering, and this desk called none.
Read September 28, the Gemini API pricing page lists paid Standard rates for gemini-3.8-flash, gemini-3.7-flash and gemini-3.6-flash at $0.75 input and $3.75 output per million tokens 'through December 31, 2026', and $1.50 and $7.50 'starting January 1, 2027'. Free-tier cells read 'Free of charge'. The page footer is dated September 24. The step is not new: archived captures show it on the 3.6 and 3.7 Flash rows by mid-August and on 3.8 Flash from the first capture that lists that model, and this desk's record finds 3.6 Flash listed at $1.50 and $7.50 with no end date before mid-August, so January returns it to its launch price. Google's changelog names neither price.
These links follow shared source records; they do not imply conduct or a direct relationship.
Amazon's Bedrock legacy table, read September 28, lists Nova Canvas and both Nova Reel versions for end of life on September 30, 2026 with no public extended access date; its rules say requests fail on or soon after that date unless privately arranged. It lists Claude Opus 4.1 entering public extended access, at higher provider-set pricing, on October 8, 2026 with end of life January 8, 2027, and Claude Sonnet 4 ending October 14, 2026, and says its dates are specific to Bedrock. Anthropic's deprecations page, read the same day, lists both Claude models as retired on Anthropic-operated platforms, Sonnet 4 on June 15 and Opus 4.1 on August 5, and says partner-operated platforms such as Bedrock set their own schedules.
These links follow shared source records; they do not imply conduct or a direct relationship.
Read September 28, Microsoft's Foundry retirement schedule lists MAI-Image-2.5-Pro, MAI-Image-2.5-Flash and MAI-Image-2.5, all in Preview, for retirement on October 1, 2026 with an empty replacement column, 47 days after their predecessors MAI-Image-2e and MAI-Image-2 were retired in their favour on August 15. Kimi-K2.7-Code, also Preview, is listed for October 3 with no replacement. Microsoft describes Preview retirement dates as 'not sooner than' dates that are sometimes extended, and its MAI image how-to page lists two newer Preview models, MAI-Image-2.6 and MAI-Image-2.6-Flash, without calling them replacements.
These links follow shared source records; they do not imply conduct or a direct relationship.
Read September 27, OpenAI's custom GPT retirement and migration FAQ states that custom GPTs are scheduled to retire on December 11, 2026 across all ChatGPT plans, and that at retirement custom GPTs and their pages become inaccessible. Enterprise workspaces holding a qualified deferral get a later date, February 11, 2027. Migration turns a GPT into a plugin; custom actions, conversations, sharing settings and the selected model do not carry over. Against the Internet Archive's September 18 copy, the planned Enterprise cutoff for creating new custom GPTs now reads October 26 instead of September 25, the migration target September 22 instead of September 17, and a sentence about migrated GPT links redirecting is gone. The only ChatGPT release note on the subject, dated September 11, gives no date.
xAI's release notes add an entry dated September 25 introducing an opaque, developer-assigned 'safety_identifier' on Chat Completions, the Responses API, deferred completions, the Batch API and gRPC, so a usage-policy violation can be attributed to one end user rather than the developer's key. The security FAQ tells developers to send a hash, never an email address, phone number or display name, and says the field matches OpenAI's. The entry was not in a September 27 archived capture of the page.
These links follow shared source records; they do not imply conduct or a direct relationship.
GitHub introduced a global 'Default policy for new features' for Copilot Business and Enterprise, with the options Enabled, Disabled or Let organizations decide. The policy can be configured now but 'won't affect feature access for your users yet.' Starting October 22, eligible generally available features left Unconfigured follow the selected global default. Explicit prior decisions are preserved and preview features remain opt-in. GitHub's documentation, read September 27, adds: 'This policy is enabled by default. If you don't take action, unconfigured features will be enabled on October 22.'
These links follow shared source records; they do not imply conduct or a direct relationship.
Read on September 27, OpenAI's Sora 2 model page and video-generation guide both state that 'The Sora 2 models and Videos API were shut down on September 24, 2026 and are no longer available,' and the model page adds that there is no one-to-one replacement API. On the same day, the help-centre article on the Sora discontinuation still said 'The Sora API will be discontinued on September 24, 2026,' with its export window still conditional. The deprecations ledger still listed the six Sora rows, with no replacement, under 'Upcoming deprecations.' This desk's unauthenticated requests to /v1/videos that evening returned 404 with an empty body, the same answer as a route that does not exist. Microsoft's Foundry schedule lists its own Azure sora-2 deployment as retiring on October 15, 2026.
OpenAI announced GPT-6 Sol and GPT-6 Luna as available 'starting today' in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, with GPT-6 Luna for Free and Go users in the desktop app, a gradual rollout in ChatGPT through the day, and API ids gpt-6-sol and gpt-6-luna. It states the models are 'not yet available in Chat.' API prices are stated as 50% below GPT-5.6 promotional pricing: Sol $2/$10 and Luna $0.10/$0.50 per million input/output tokens. The pricing page, read September 27, lists those rates and states that GPT-5.6 Sol's promotional pricing, now listed at $4/$20, 'is available at least through November 21, 2026.' This file's August 7 reading of the same page had gpt-5.6-sol at $5/$30, and OpenAI's API changelog, as this desk's September 27 record reports, dates the $4/$20 price to August 21. The models page now names GPT-6 Sol and Luna, 'when available', as the replacements for GPT-5.5, which retires from ChatGPT, ChatGPT Work and Codex on October 14 but not from the API.
These links follow shared source records; they do not imply conduct or a direct relationship.
Anthropic introduced Claude Opus 5.5 (claude-opus-5-5) as 'now available on all platforms,' priced at $4/$20 per million input/output tokens and $0.20 for cache reads. It states it is increasing five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans and giving subscription users 'a rate limit reset, which you can now save and use whenever you choose,' without stating the size of the increase or the reset's expiry. Opus 5.5 ships with safeguards similar to Fable 5.1's: 'most cybersecurity tasks will be re-routed to Opus 4.8,' vetted organizations can apply to a new Life Sciences Verification Program for biology work the safeguards impede, it is no longer available with thinking switched off, and preserved thinking applies to API accounts created on or after August 31, 2026. The deprecations page lists it as Active with retirement not sooner than September 22, 2027.
These links follow shared source records; they do not imply conduct or a direct relationship.
OpenAI published Credit Score Terms on September 21 for a ChatGPT feature that shows a user's credit report and VantageScore 3.0 score, obtained through CSIdentity Corporation, which the terms call Experian. The terms say enrolling provides 'written instructions' under the Fair Credit Reporting Act for OpenAI and CSIdentity to obtain information from the user's credit profile for purposes including identity verification, credit monitoring, personalized insights, and improving the service and creating new features and products. The terms state the pull is a soft inquiry that does not affect the score. OpenAI's promise to delete underlying report data within 30 days of disconnecting is in a help article, not in the terms.
These links follow shared source records; they do not imply conduct or a direct relationship.
Google's Gemini API changelog states on September 18 that access to the 2.5 models is limited 'to users who have actively used them in the past.' It says the models are not deprecated, and new projects are pointed to 3.5 Flash-Lite or 3.8 Flash. The September 17 entry deprecates antigravity-preview-05-2026 in favor of antigravity-preview-09-2026, with renamed built-in tools and PascalCase parameters for local tool users, and states the old agent shuts down on October 5, 2026. The deprecations page lists that date and still carries the September 30 gemini-omni-flash-preview and October 2 gemini-2.5-flash-image dates, all qualified as earliest possible dates. Google Cloud's agent platform, a separate surface, lists the same three 2.5 models for retirement on October 20, 2026.
These links follow shared source records; they do not imply conduct or a direct relationship.
Read the records behind this file. Search titles, publishers, summaries, and extracted facts. Each source keeps its reading limits and links to the events and roles it supports.
Showing 1–10 of 173 sources · 173 in this file
OpenAI
OpenAI's API deprecations ledger, read October 1, 2026, three days after the 2026-09-28 shutdown date its 2025-09-26 announcement set for four legacy completions models. All four rows survived the date unchanged and now name a replacement: gpt-3.5-turbo-instruct, babbage-002, davinci-002 and gpt-3.5-turbo-1106, each shutdown date 2026-09-28, each replaced by gpt-5.6-terra. The section has moved from 'Upcoming deprecations' to 'Past deprecations', but its paragraph is still in the future tense, while the section immediately below it, announced the same day, was rewritten into the past tense. The two headings now sort by announcement recency rather than by whether a shutdown has happened: everything announced 2026-06-02 or later sits under Upcoming, everything announced 2026-05-08 or earlier under Past, which puts the 2026-04-22 group's seventeen October 23, 2026 rows — a date that has not arrived — under 'Past deprecations'.
Google's deprecations page, read on September 30 in the revision its footer stamps 'Last updated 2026-10-01 UTC'. Its Veo models table lists veo-3.1-lite-generate-preview, veo-3.1-generate-preview and veo-3.1-fast-generate-preview with a shutdown date of October 22, 2026 and gemini-omni-1.1-flash as the recommended replacement; the Internet Archive's September 26 copy listed all three with 'No shutdown date announced'. The Gemini Omni Flash section no longer lists gemini-omni-flash-preview, which the September 27 read found listed at September 30, 2026.
A shared source does not establish conduct or a direct personal relationship.
Mistral AI
Mistral's changelog entry dated September 29, 2026 deprecates three models: 'OCR 4.0 (mistral-ocr-4-0) is deprecated and retires on September 30, 2026. Use OCR 4.1 (mistral-ocr-4-1 or mistral-ocr-latest) instead, at the same price.'; 'Leanstral 1.5 (labs-leanstral-1-5), an experimental Labs model, is deprecated and retires on September 30, 2026.'; and 'Z.ai GLM 5.2 (zai-glm-5-2) is deprecated and retires on October 31, 2026. Use Z.ai GLM 5.3 (zai-glm-5-3) instead, at the same price.' Its September 28 entry says Z.ai GLM 5.3 'is now Generally Available.'
A shared source does not establish conduct or a direct personal relationship.
Amazon Web Services
Amazon's Bedrock legacy-models table (columns: provider, model, model ID, regions, Legacy date, EOL date, public extended access start date) and EOL rules, read September 28, 2026. It lists Nova Canvas and Nova Reel with a September 30, 2026 EOL and no extended-access date, Claude Opus 4.1 with public extended access from October 8, 2026 and EOL January 8, 2027, and Claude Sonnet 4 with EOL October 14, 2026.
A shared source does not establish conduct or a direct personal relationship.
Anthropic
Anthropic's model deprecations page, fetched September 28, 2026 (HTTP 200). It states that its dates apply to Anthropic-operated platforms, the Claude API, Claude Platform on AWS and Microsoft Foundry, and that partner-operated platforms, Amazon Bedrock and Google Cloud, set their own retirement schedules. Its status table lists claude-sonnet-4-20250514 as Retired (deprecated April 14, 2026, retired June 15, 2026) and claude-opus-4-1-20250805 as Retired (deprecated June 5, 2026, retired August 5, 2026).
A shared source does not establish conduct or a direct personal relationship.
xAI (SpaceXAI) developer docs
xAI's security FAQ, footer 'Last updated: September 28, 2026', read that day, explaining that a request violating usage policies can be attributed to the developer-assigned end-user identifier instead of the whole API key, and instructing developers to send a hash rather than personal data.
A shared source does not establish conduct or a direct personal relationship.
Hugin
Hugin's September 28 record on the credit-score feature OpenAI began rolling out in ChatGPT's Finances section on September 21 to Plus and Pro subscribers in the U.S., read against the Credit Score Terms published the same day and OpenAI's help article on disconnecting.
Hugin
Hugin's September 28 record on the Gemini API pricing page's two prices for every paid token rate on Gemini 3.6, 3.7 and 3.8 Flash, one 'through December 31, 2026' and one, double it, 'starting January 1, 2027', read against Google's changelog and Internet Archive captures.
Hugin
Hugin's September 28 record on three providers ending AI models in the same week with no successor named on their retirement pages: Amazon Bedrock's own Nova Canvas and Nova Reel and Mistral's Leanstral 1.5 on September 30, and Microsoft Foundry's own MAI-Image-2.5 models on October 1.
Anthropic
Anthropic's model deprecations page, fetched September 27 (HTTP 200). It lists claude-opus-5-5 as Active with a tentative retirement 'Not sooner than September 22, 2027.' The only model its status table marks Deprecated is claude-mythos-preview, deprecated June 9, 2026. claude-opus-4-7 is still Active, not sooner than April 16, 2027, and claude-sonnet-4-5-20250929 is Active with a floor of 'Not sooner than September 29, 2026.'
Import sources create source-run packets, document inventories, and review tokens. They do not publish accusations or graph relationships until the case-file evidence rules are satisfied.
baselineClaim rows are derived from timeline events and source anchors. They are evidence-status labels, not accusations or final truth claims.
Release requests are lawful public-record asks. They identify source gaps, review gates, and privacy constraints; they are not evidence that a specific person committed wrongdoing.
Deliberate gaps in the public record — open until the underlying records are released or reviewed, not questions we have skipped.
0167 B938 CB23Connections are navigation aids. They show where a record, event, or role is sourced; they do not convert association, correspondence, testimony requests, or document appearance into misconduct findings.
AI release desk for official lab receipts, safety records, public coverage, and draft promotion copy with no automatic publishing enabled.
Source leads and automated probes are review inventory. They become verified anchors only after Hugin can attach the public record, legal posture, and no-allegation guardrails.
These are not loose TODOs anymore. Each lane below is computed from the case file's source tiers, then joined to matching sources, timeline events, role nodes, and release-roadmap requests.
Compare provider status, pricing, model overview, release notes, and help-center rows before assigning failed runs to model quality.
Re-check the Scheduled Tasks and GPT-5.6 help records before repeating active-task caps, supported models, plan limits, or reset-time language.
Re-check Codex model docs, GPT-5.5 launch framing, and Codex changelog entries before repeating current model-default claims.
Track current Codex goal-mode docs, app behavior, validation-loop examples, and completed Hugin release receipts before describing long-running workflow reliability.
Track hosted Reddit 403/429 recoveries, duplicate-report reuse, and companion/browser recovery paths before claiming a scan failure is a reader error.
Keep Hugin field notes and social reactions clearly labeled until they are joined to provider docs, status rows, or repeatable verification.
Hugin accepts public records, not accusations. A submitted URL lands in a review lane, gets source-validated, and only becomes a case anchor after the evidence rules and privacy gates are satisfied.
Fact-check status is a review posture, not a verdict. Source-scout rows are leads until a reviewer joins them to publishable anchors; claim-ledger rows are timeline-derived and still require source validation plus no-allegation guardrails before promotion.
Paste an official or public-record URL. Hugin classifies the lane first, keeps lead-only material quarantined, and requires source validation before anything reaches the case file.
Queue a public source for review.Sanitized queue view: Hugin shows the public URL, lane, status, and why a source has not been promoted. Submitter identifiers are not shown.
A source is a public URL someone else can inspect. A claim is an interpretation. Hugin queues sources first and keeps claims out until a reviewed public record supports the exact wording.
Official agency pages, court dockets, congressional releases, civil records, state updates, and public repositories with source URLs.
Publisher, date, record family, and why the URL belongs in this file.
Source URLs first. Context is welcome, but claims do not publish from submissions alone.
Private tips, survivor identifiers, sealed or leaked material, graphic details, and private contact data.
Screenshots or social posts without a public source URL.
Misconduct labels for living people unless the exact cited public record supports that exact statement.
Audit feed for source anchors, queue decisions, privacy-review jobs, and hashes. It explains the process without turning leads into claims.
Extract document index and source metadata first.
Next: Metadata extraction includes publisher, URL, document family, and date.821830d977f6Extract document index and source metadata first.
Next: Metadata extraction includes publisher, URL, document family, and date.70b4ee7e45d7Extract document index and source metadata first.
Next: Metadata extraction includes publisher, URL, document family, and date.3e5eeffc13c4Extract document index and source metadata first.
Next: Metadata extraction includes publisher, URL, document family, and date.e1bc1c396c47Extract document index and source metadata first.
Next: Metadata extraction includes publisher, URL, document family, and date.ea542b4b6595Extract document index and source metadata first.
Next: Metadata extraction includes publisher, URL, document family, and date.cca0a8ec7433Extract document index and source metadata first.
Next: Metadata extraction includes publisher, URL, document family, and date.0b24398cd908Extract document index and source metadata first.
Next: Metadata extraction includes publisher, URL, document family, and date.79237a8f7e42Mirrors can compare this hash to verify the same case file bundle.
Next: Compare the short hash in the UI with this machine-readable ledger.0167b938cb23This binds document mirroring and privacy-triage rules.
Next: Compare the short hash in the UI with this machine-readable ledger.5cee74995ec2