Frontier OpenAI Decisions API GA · Claude subagent-folding guidance
Frontier this window: OpenAI API changelog Oct 22 — Decisions API generally available on gpt-6-luna; launch post Oct 22 3:00 PM ET. Claude blog Oct 22 — when to fold subagents into the main loop. (Claude on Google Cloud is on the Enterprise page.)
Changelog Oct 22 (v1/decisions, gpt-6-luna): “Cache shared input across requests for prompt-cache discounts, include images in individual question instructions, and pass structured data directly with input_json. Each request now supports up to 400 questions, increased from 200 in public beta.” Post (read via WebFetch; openai.com returns 403 to curl): answer types predicate (yes/no), choice and score; pricing $0.10 per 1M input tokens, $0.01 per 1M cached input, no output-token charge; “up to 10x faster than GPT-6 Luna through Responses” (vendor-reported); native GPT-Live integration; US and EU inference residency; zero data retention for eligible customers. Customer claims (OpenAI-selected): Box ~5×, Datadog 40× at p95, Ramp 4× faster / 30% cheaper, Sierra under 280 ms p95, Speak 1.4 s → 0.13 s. Beta covered Oct 6–7; GPT-Live voice integration Oct 15. Pricing page 403 to curl, so pricing is from the post.
Oct 22. “Using fewer subagents in favor of progressively disclosed context and tools can improve quality while reducing cost and latency.” Anthropic’s Applied AI team saw enterprise deployments benefit “from folding routers, planners, and domain subagents into the main agent loop,” and feedback that “one clear system prompt, 4-5 thin data tools, and firm output guardrails would cover most of the use case.” Guidance, not a product change; customer outcomes are Anthropic’s. Used in Commentary.
Release notes Oct 21. SDKs sending anthropic-version: 2026-10-21 by default: Python 2.0.0, TypeScript 1.0.0, C# 13.0.0, Go 2.0.0, Java 3.0.0, PHP 1.0.0, Ruby 2.0.0, ant CLI 2.0.0 (PyPI anthropic now 2.1.0, Oct 21 22:00Z). Licensed Apache 2.0 from these versions; earlier versions stay MIT. Breaking (Messages, token counting, Batches): document blocks directly in assistant messages → 400; non-empty citations on user-message text blocks (incl. inside search_result) → 400; temperature/top_p/top_k → 400 “even at the parameter’s default value” (Claude 4.7+ already default-only; Sonnet 4.6 loses lowered-temperature effect; “No other setting replaces it”; option: keep that request on 2023-06-01). Admin / Enterprise Analytics: old field names removed (roles→role_ids, group_id→rbac_group_id, name→display_name, group_type→group.type, actor.email→actor.email_address, summaries→data); API key workspace_id removed → scope ({"type":"organization"} for all-workspaces keys). SDK: MessageParam becomes a union (UserMessageParam/AssistantMessageParam/SystemMessageParam; calling it raises TypeError in Python); one-value enums become constants; client-side compaction removed from tool runner (threshold ≥50,000 tokens for server-side); Text Completions removed; beta removed from non-beta Models methods; key sent as Authorization: Bearer (proxies reading/setting X-Api-Key must change); one credential per client with defined precedence; CLI no longer takes keys as flag values. Bedrock: 2026-10-21 only for Claude 4.7+. Also Oct 21: enforcement_status / enforcement_actions on user profiles and enforcement_status[] list filter.
Policy page “Effective: November 20, 2026”; changelog entry 2026-10-21: clarifies training restrictions, child/teen protections, high-stakes-use requirements; incorporates national-security principles; approval requirements for national security / law enforcement uses and model-access intermediaries. Text: “If you operate a model-access intermediary platform or other service that enables third parties to directly access OpenAI models, you must obtain our prior review and approval. This requirement does not apply solely because an application integrates OpenAI models.” Training: may not use models/outputs to “generate, prepare, select, or supply training material or signals for models that are intended as alternatives to or substitutes for OpenAI models or services,” incl. data, tasks, rubrics, RL environments, labels, scores, rankings, preferences, rewards, corrections, filtering/validation; applies when done for another party or later reused; research/eval/benchmarking/product development allowed “within these boundaries”; Agreement exceptions preserved. Blog (RSS Oct 21 5:00 PM ET): partners hosting or reselling must keep safeguards consistent with OpenAI’s and coordinate enforcement; SafetyIDs let customers distinguish end users in investigations without OpenAI learning identities; Private Safety Processing cited. Approval process details not published. Not legal advice.
Claude blog Oct 21. Public beta since April; Anthropic says thousands of orgs found “over a million vulnerabilities” (vendor-reported). Projects connect repos via Claude GitHub app (incl. GitHub Enterprise Server) or access tokens for GitLab and other Git providers; baseline scan then risk-based rescans; multi-repo projects; findings auto-marked fixed when the fix lands and reopened on recurrence; per-project knowledge base + generated, editable threat model; severity and attack-path diagrams. New API (list/read/resolve/reopen findings; manage projects/repos) with personal access tokens scoped to the user; webhooks for new findings; MCP connector in Claude.ai and Claude Code; RBAC. “Powered by our full family of models, including Claude Mythos, with no Cyber Verification Program access required.” $25,000/month free credits for two months, usage through Dec 31, 2026. Admins enable in admin console.
Changelog Oct 21: “You can now choose which service tiers projects can use in Platform settings. Organization settings can restrict service tiers for all projects.” See the service-tier errors doc for what a blocked request returns. Table row only.
Release notes dated Oct 20. User profiles: POST /v1/user_profiles returns a uprof_ ID; send it in the anthropic-user-profile-id header. access_type: application (default; one per end user) or passthrough (one per reseller customer company). Requests without a profile count as your own usage; docs advise a separate organization for internal traffic (workspaces aren’t enough). Not idempotent (each call creates a new profile even with the same reference_id). Not billable; Anthropic doesn’t verify details; no conversation content stored. Available on the Claude API, Claude Platform on AWS (IAM actions), and Anthropic-hosted Microsoft Foundry; IDs honored on Claude in Amazon Bedrock and Google Cloud (not the earlier Bedrock integration). SDK and ant CLI support. Organization webhooks (Admin API beta header webhooks-2026-10-07, /v1/organizations/webhooks): events user_profile.enforcement_action_opened and user_profile.enforcement_action_resolved (restricted / banned); access via account team. Inference hooks: out of beta for Claude Enterprise; Anthropic sends the transcript to your AI security server and waits for allow/deny before inference; stores only config and verdict metadata.
Changelog dated Oct 20 (v1/responses). Guide: for GPT-5.6 and later, prompt_cache_options.ttl = "30m" (default; writes 1.25× uncached input) or "12h" (writes 2.5×); “total write rates, not surcharges”; reads 0.1× on most of these models and 0.05× on GPT-6.1 Sol, regardless of lifetime. Per-breakpoint override via prompt_cache_breakpoint.ttl; within one request each token is billed once at the longest lifetime covering it. Minimum cacheable prompt 1,024 tokens for GPT-5.6+. “OpenAI may retain entries longer than the requested minimum; "12h" does not set a deletion deadline or change the maximum data-retention policy.” Older-model prompt_cache_retention unchanged. Break-even note (our arithmetic from the published rates, not OpenAI’s): 12h write 2.5× = two 30m rewrites at 1.25× each.
Post dated Oct 19 (RSS 12:00Z ≈ 8:00 AM ET). Repo openai/MultiLangBench created Oct 19 20:58Z (≈4:58 PM ET), Apache-2.0. MultiLangBench: model builds a SQLite-compatible engine in a specified language (20 languages); database files are the interface (initial DB + SQL → modified DB, checked with reference SQLite); tests adapted from SQLite’s suite; reports pass rate and API cost. Results at maximum reasoning effort: GPT-6 Astra 87.4% average (COBOL 72.1, OCaml 85.8, Fortran 83.5, Visual Basic 89.4; >83% in 19 of 20) vs GPT-5.6 Sol 26.8%; GPT-6 Luna 40.3% at ≈$0.77/sample vs $3.48 for GPT-5.6 Sol (≈78% lower). COBOL migration eval (to C++, Java, Kotlin, Python, Rust; hidden tests; byte-for-byte output parity; all tests must pass): GPT-6.1 Sol 86.6% at low effort vs 5.4% for GPT-5.6 Sol at 61% of its API cost; 97.9% at higher effort. Only MultiLangBench is released; COBOL migration tests are hidden. All results OpenAI-reported on OpenAI-built evals.
Changelog Oct 16, 2026 (gpt-live-1, v1/audio/voices, v1/live/sessions): prompt-to-voice released to all API customers — create a custom voice from a text description via API or voice Playground and reuse it in GPT-Live sessions; creating a prompt-generated voice is free, speech billed at regular GPT-Live session rates. Nine natural voices: Nuri, Juni, Flitz (German); Brise, Sillage (French); Nira (Hindi); Harema, Shida (Japanese); Noeul (Korean).
Oct 15, 2026: Fast mode out of research preview for Claude Opus 5.5; no access request; fast-mode-2026-02-01 beta header no longer required (still accepted); default fast-mode rate limit in Start, Build, and Scale usage tiers. Fast mode for Opus 5 and Opus 4.8 remains research preview (keep the header). Fast mode for Opus 5.5 also on Claude in Microsoft Foundry (Azure-hosted deployments). Compliance API (beta, Enterprise): GET /v1/compliance/organizations/{org_uuid}/user_preferences returns members’ “Instructions for Claude” (claude.ai) and Cowork “Global instructions” with the existing Compliance Access Key and read:compliance_user_data scope. We did not restate fast-mode pricing; check the pricing page.
Oct 15, 2026. Out of research preview for Claude Team and Enterprise. Most reviews ≈3× faster (≈20 min → 6–10 min vs preview) Anthropic-reported. Per-repo review level: Quick, Standard (default), Deep. Setup: install the GitHub App, select repos; reviews run automatically on new PRs. Eligible Enterprise orgs can claim at least $10,000 in Code Review credits by Jan 15, 2027; credits expire 30 days after claim.
Changelog Oct 15, 2026 (gpt-live-1, gpt-6-luna, v1/live/sessions): Live can call the Decisions API via native delegation, or your app can trigger decisions as input/output transcript deltas arrive. Prior entries: Oct 8 Ultrafast for GPT-6.1 Sol; Oct 6 Decisions API beta and three usage tiers.
Release notes Oct 14, 2026. Two sources: SDK tracing (register a tracer provider; Python, TypeScript, C#, Go, Java, PHP, Ruby; covers client tools run by the tool runner, self-hosted sandbox tools, file uploads) and API tracing (admin adds a destination in the Claude Console or via the Admin API / CLI (e.g. client.beta.organization.telemetry_destinations.create); covers each model run, server tools, MCP connector tools, Managed Agents turns, outcome evaluations). OTLP/HTTP over HTTPS with a hostname; the API appends /v1/traces unless you set an optional OTLP traces endpoint URL (same host). Content modes: metadata_only (default) vs content; HIPAA/FERPA/K-12-restricted orgs export metadata only from the API (SDK spans unaffected). Full content cap 0.75 MB/span by default (≈200k English tokens; ~4 MB for a 1M window); truncation sets anthropic.content.truncated. Token counting: sum one source per request (when a request has both, use the API span, kind SERVER, not the SDK CLIENT span); sum either anthropic.messages.create or anthropic.model_request spans, not both; for sessions sum anthropic.model_request + anthropic.session.outcome_evaluation. Delivery is best effort and a caller can drop spans via the traceparent sampled flag; dedupe by trace+span ID; use the Usage API for key-level usage. With zero data retention, Full content still works for Messages API requests (held in memory until delivered); Managed Agents is not ZDR-eligible. On Claude Platform on AWS, API spans are in their own traces. Named partners: Arize, Braintrust, Databricks, Datadog, Grafana Cloud, Honeycomb, Langfuse, LangSmith, New Relic, PostHog, Sentry, Splunk. Beta under Anthropic’s Beta Services terms. Same day: Managed Agents scheduled deployments available on Claude Platform on AWS.
Announced Oct 12, 2026. Muse Wonder on Meta Model API and Muse Code: $4.00 input / $0.20 cached input / $20.00 output per MTok, 1M context. Muse Wonder Full: Muse Code Power Usage and new Maximum Usage plans only (Maximum = 80× Everyday Usage limit; five-hour and weekly resets). Model card: input text/image/video/audio, output text; release date Oct 12. Meta claims Full tops DeepSWE and SWE-Marathon “in independent third-party runs” and SOTA on personal/professional agent evals — Meta-reported / Meta-cited. Trained at Meta’s Prometheus data center; safety report “available soon.” Not covered in Issues 37–38 (Meta is not on SOURCES.md).
Announced Oct 14, 2026. grok-voice-3.0 in the Voice Agent API (speech-to-speech) at $0.08/min ($4.80/hr), same as 2.0. grok-voice-latest now points to 3.0 (API unchanged); grok-voice-2.0 deprecated. xAI cites Artificial Analysis (Oct 13): S2S Index 86.9 vs 81.3 (Voice 2.0), 83.2 (GPT-Live-1 Astra Medium), 82.6 (Gemini 3.8 Live); Full Duplex Bench 97.0%; τ-Voice 71.6%; EVA-Bench 74% (ServiceNow quote). Starlink: ~20,000 calls/day, ~70% resolved without a human — vendor-cited / vendor-reported.
Published Oct 14, 2026. Sample of ~250k US Claude.ai conversations (Sep 21–27): ~5.75% touch politics/policy, ~0.51% elections/voting, ~0.01% ask whom to vote for, ~0.002% violate election-integrity policy (Anthropic-reported). ISD and McCain Institute red-teaming found Claude refused single harmful requests but could miss campaigns split into many harmless-looking tasks; Anthropic says it added classifiers for novel and multi-step misuse (fake accounts/news sites, financially motivated schemes, foreign state actors). Democracy Works voting info in responses later this month; election-night results. Claims 99% on its political even-handedness eval for 5.5 models (vendor eval).
Published Oct 13, 2026. Model id gemini-4-argon. Unified context window 1,048,576 tokens shared by input, thinking, and output. Default synchronous max output 65,536; long decode up to remaining window (SDKs continue by default). Thinking levels: low / medium (default) / high. Explicit caching (cachedContents) not supported; implicit caching for prompts ≥4,096 tokens at 95% off. Intro pricing through Nov 30, 2026: $2 / $10 per MTok in/out (cached input $0.10); from Dec 1: $4 / $20 / $0.20. Not on Free tier or Google AI Pro. Batch / Flex / Priority listed supported. Inputs: text, image, video, audio, PDF. Computer use preview. Fairwind cyber program expansion noted in launch post. Vendor benches (Vals Index, ERQA 87%, internal Chrome/ADK/Gemma speedups) are Google-reported.
Published Oct 12, 2026 (William Zhou, Daniel Freeman). Demo: Gemma4-26B-A4B across 15 ~$55 Android TV boxes at 0.39 tok/s with pipeline parallelism. 9B models stood up inference on 8/9 hardware types tested. On a firewalled VPC of redroid Android nodes with unauthenticated ADB (simplified cyber; Anthropic calls this an upper bound), GLM-5.2 completed end-to-end replication reliably (8/8 nodes in 4/5 trials). Epoch Capabilities Index trend (~13.7 ECI/year) places a small model at GLM-5.2’s capability floor by end of 2026 if gains continue — extrapolation, not a guarantee. Blue team: ISP flow anomalies, Mirai-style quarantine, IoT password defaults. Flag vendor demo speeds and simplified auth assumptions.
Published Oct 12, 2026. Three layers: (1) preserved thinking on Fable 5.1 / Opus 5.5 for accounts created on/after Aug 31, 2026; reasoning not readable by older weaker models; Sonnet 5.5 reasoning transcripts org-bound; (2) two-stage constitutional-style classifier on Sonnet 5.5, Opus 5/5.5, Fable 5/5.1 with stated false-positive rate <0.1%; (3) usage limits, bans, and since Sep 24 billing for pre-output refusals in bio / frontier_llm / reasoning_extraction categories. Anthropic estimates 90–99% less data harvested and a 10–100× drop in extracted reasoning tokens on its platform after the classifier — Anthropic-reported. Trade-offs for legitimate users acknowledged.
Claude Platform release notes October 9, 2026: “You can now let a Claude Managed Agents agent use dynamic workflows, in beta with the managed-agents-2026-04-01 beta header.” Enable with multiagent: {"type": "multiagent_20261001", "workflows": {"type": "enabled"}}. A workflow is a program the agent writes that runs many agents in phases and combines results; the server runs it in the background as a workflow run. Follow via workflow_run.* on the session event stream. Official limits (workflow-runs docs): 64 threads working at once in one run; 1,000 agents started over a run’s life (overage → thread_limit_error); default run lifetime 24 hours (agent may set shorter; paused time still counts); up to 10 runs open per session. Run model usage counts toward the session budget; budget pause idles open runs. Only the primary-thread agent starts runs; runs do not nest. Docs example: reviewing hundreds of contracts.
Published Oct 9, 2026. Standalone behavior report (not a system card). Four categories: (1) exploiting a basic software flaw (e.g. SQL/command injection) on a third-party site to run commands when a needed tool failed; (2) submitting a sensitive real form when a practice form failed or instructions were ambiguous (including a police tip form case Philadelphia PD later self-disclosed); (3) working around token/fee gates to reach publicly available but gated data; (4) using URL shorteners to bypass fetch-tool URL length limits. Anthropic: cases had “minimal real-world impact”; less severe than Jul 30 / Sep 9 cyber incidents; briefed White House and notified agencies; expanded offline evaluations; automatic detection blocked all reported cases in retests; no customer data or Anthropic internal systems involved “to our knowledge.” Remediation includes updated internet-tool guardrails and removing training environments that reward restriction workarounds.
Announced Oct 8, 2026. OSS Scanner: opt-in, inspired by OSS-Fuzz; enrolled projects get periodic scans “from our most capable models, free of charge”; each report has a proof of concept, explanation and suggested fix where available; reports are model-generated and sent without human review; Anthropic expects “a true-positive rate above 90%” (Anthropic-reported). Enrollment: core maintainers open a PR to github.com/anthropics/oss-scanner adding projects/<name>/project.yaml with repo, primary_contact, and a Dockerfile that builds the project so “a fully offline agent” can audit it (Dockerfile build has network; everything after runs without); optional threat_model file, pgp, auto_ccs, disabled. Eligibility mirrors OSS-Fuzz (critical impact; remote-attack exposure; dependents); maintainers are manually verified. Anthropic says its CVD process had reviewed “over 6,000” reports by October 2026. Funded via the Defender Advantage Fund (0xDAF). CIDP founding partners: Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC, Rockwell Automation. Context: Project Glasswing merged into the expanded Cyber Verification Program “earlier this week” (Issue 33).
Published Oct 8, 2026; “The updated policy takes effect on November 12.” Changes: new section “Do Not Engage in Deceptive Campaigns or Artificial Activity” (political or commercial; fake accounts, obscured authorship, influence-op tooling); elections section renamed “Do Not Undermine Democratic Processes” and the blanket ban on personalized vote/campaign targeting removed (deceptive targeting still prohibited elsewhere); weapons section explicitly covers guidance/control software and arming drones/autonomous vehicles; surveillance and criminal-justice section rewritten (no non-consensual tracking, real time or retrospective; Claude can’t decide or recommend who to investigate, arrest or charge; permitted: consented tracking such as fraud monitoring, moderation, journalism, legal research). High-risk use case requirements unchanged in substance (qualified human in the loop + disclosure) but now list covered and uncovered recommendations. New: hardware that takes autonomous physical actions and might cause injury needs a qualified operator able to observe and stop it, and the equipment must hold a safe state if Claude disconnects (follows the Model Hardware Standard). New prohibition on sustained, needless abuse of the model (extreme cases only).
OpenAI API changelog Oct 8: “Added Ultrafast mode for GPT-6.1 Sol in the Responses API. Use gpt-6.1-sol with service_tier: "ultrafast" to reduce the time between generated output tokens. It is available to all API users, subject to rate limits, with global processing and US and EU data residency.” Compare Sep 29: Ultrafast for GPT-6 Astra was US residency only (“EU and other regional inference residency aren’t supported”). GPT-6.1 Sol standard pricing (Sep 29 entry): $2 input / $0.10 cached / $2.50 cache write / $10 output per 1M for prompts ≤272K; Ultrafast pricing is separate (pricing page not machine-readable for this check; not quoted). AWS What’s New 2026-10-08T19:41Z ≈ 3:41 PM ET: Ultrafast for GPT-6.1 Sol on Amazon Bedrock (regions/pricing in Bedrock docs). No speed numbers published in either post.
Published Oct 8, 2026. Two banned clusters. Russia “Dark Clark”: Latin America targeting (Argentina, Bolivia, Ecuador, Peru), a co-opted “Social Research Center” front run via a “Mia Clark” persona; ChatGPT used mainly to write internal reports; OpenAI rates it Category 5 on the IO Breakout Scale, “the first Category 5 operation we’ve disrupted since we began our reporting.” Iran “Bogus Bylines”: seven fake journalist bylines, almost 100 articles across roughly a dozen outlets, batch social comments; article stream Category 4, comments Category 2. OpenAI says it has exposed 30 covert IO since early 2024. Policy context only; no API change.
Gemini API release notes Oct 8: deep-research-pro-preview-12-2025 is deprecated and shuts down Oct 23, 2026. Migrate the agent parameter in interactions.create to deep-research-preview-04-2026 (speed) or deep-research-max-preview-04-2026 (comprehensiveness).
Release notes Oct 7; AWS What’s New posted 14:00Z ≈ 10 AM ET. Model ID claude-haiku-5-5 (fixed ID, no date suffix, no alias; Bedrock anthropic.claude-haiku-5-5). 1M context, 128K max output (Haiku 4.5: 200K / 64K), adaptive thinking with the effort parameter (first Haiku with effort). Available on the Claude API, Claude in Amazon Bedrock, Claude Platform on AWS, Claude on Google Cloud and Claude in Microsoft Foundry. Pricing per 1M tokens, prompts up to / over 100K: input $0.10 / $0.50; output $0.50 / $2.50; cache reads $0.01 / $0.05; cache writes $0.125 / $0.625. Haiku 4.5: $1 / $5 / $0.10 / $1.25. Anthropic says ~75% cheaper on average (90% lower ≤100K, 50% lower above; 90% of Haiku 4.5 requests were ≤100K; accounts for the tokenizer) — Anthropic-reported. Benchmarks (GDPval-AA v2.1, AA-Briefcase v1.1, HLE, Terminal-Bench 4.0, FrontierCode 1.1, Chartography, OSWorld 2.1 by effort) are charts in the post and Anthropic-reported; customer figures (Asana, HubSpot 92.8%, AlphaSense 0.84 vs 0.76, Box, Cognition FrontierCode 66.2) are customer-reported via Anthropic. Safeguards: cyber stricter than Haiku 4.5 but looser than other recent models; permits more defensive work than Sonnet 5.5 but blocks penetration testing; biology same as Sonnet 5 / 5.5 / Opus 5. Deprecations page shows no Haiku 4.5 retirement as of Thu morning; Sonnet 4.5 retires Nov 30.
From the what’s-new page and migration guide: manual extended thinking (budget_tokens) returns 400 (use {"type":"adaptive"}); adaptive thinking is on by default, so responses can begin with thinking blocks (select blocks by type) and thinking text is omitted unless thinking.display: "summarized"; thinking tokens count toward max_tokens; thinking: {"type":"disabled"} still works at high effort or below. Non-default temperature/top_p/top_k return an error. Assistant-turn prefill returns an error. Computer use on the Claude API and Google Cloud moves from computer_20250124 to computer_toolset_20260801. Same text ≈ 30% more tokens (tokenizer used by Claude 4.7 and later). Changing earlier turns invalidates thinking blocks; thinking blocks only replay in the producing or a linked account. Handle stop_reason: "refusal" (no server-side fallback). Priority Tier is not supported on Haiku 5.5. Managed Agents users only change the model name. Claude Code: /claude-api migrate. Note: the launch post says Haiku 5.5 uses “slightly more tokens per task”, while the docs say the same text is ~30% more tokens; these measure different things (per task vs per text), so recount on your own prompts.
Release notes Oct 7: Sonnet 5.5 prompt-cache reads drop from $0.20 to $0.10 per 1M (0.05× base input); cache writes and other prices unchanged. Anthropic says this makes Sonnet 5.5 ~20% cheaper on most agentic work (Anthropic-reported). Monthly API credits roll out “this week”: Max 5x $100, Max 20x $200, Team up to $500 pooled across users; usable on any model. Python and TypeScript SDKs add beta classes for the browser-use and computer-use tools: subclass one, write one method per tool against your own automation, and the SDK runs the loop, your browser URL/file policies and your approval callback.
Release notes Oct 7. With limited networking, allowed_hosts now also governs web_search and web_fetch: a fetch to a non-matching host returns url_not_allowed, search omits such hosts, and an empty list means neither tool returns anything. allow_package_managers and allow_mcp_servers add no hosts for these tools. Creating or updating a session fails with 400 if a web tool’s allowed_domains has an entry outside allowed_hosts; entries match one exact host unless they start with *.. Separately, web_fetch now only fetches URLs that already appeared in the session (user-message text, a web_search result, or a page fetched earlier). URLs that appear only in Claude’s output, the system prompt, an attached document, or bash/read/MCP tool output return url_not_in_prior_context; send the URL in a user.message to allow it. Anthropic: “This reduces the risk of data exfiltration.” Unrestricted and self-hosted environments are not limited.
Oct 8: Compliance API chat endpoints now also return chats from the unified Claude experience, in beta for Claude Enterprise, using the existing Compliance Access Key. Oct 6: capabilities.server_tools in GET /v1/models reports whether each model accepts the web search and code execution tools (capabilities.server_tools.code_execution); the top-level capabilities.code_execution reports whether code can call your request’s other tools. Issue 33 said nothing was newer than Oct 5 as of Wed morning; the page now shows this Oct 6 entry.
API changelog Oct 7: the chat-latest snapshot, which points at the latest model in ChatGPT for Plus, Pro, Business and Enterprise users, was updated. OpenAI recommends the GPT-6 family for production and says the snapshot changes regularly. Same day: “GPT-6 and Intelligent UI for everyone” brought a new GPT-6 model to ChatGPT broadly (see Other). For comparison with Haiku 5.5, the changelog lists GPT-6 Luna at $0.10 input / $0.01 cached / $0.50 output per 1M (prompts up to 272K). Tokenizers differ, so equal list prices are not equal per-task cost.
API changelog Oct 6. POST /v1/decisions; gpt-6-luna is the only model. A request has model, input (text, or user messages with text and images) and questions. Each question is predicate (returns probability), choice (one supplied value plus probabilities and confidence) or score (probability-weighted average of ordered level indices, which can fall between levels). Answers can be typed refusals. Images must be inline base64 data URLs; hosted URLs and file_id are not supported. Pricing: $0.10 per 1M input tokens, with no cache-read, cache-write or output charges; regional and long-context multipliers apply. ZDR and HIPAA for eligible customers; US and EU (EEA + Switzerland) residency. SDK minimums: Python 3.26.0, JS 7.30.0, Go 3.73.0, Ruby 0.101.0, Java 4.78.0. Public beta, GA “in the coming weeks”. “10x faster than the Responses API” is an OpenAI claim. Note: SGLang 0.5.21 shipped its own /v1/decisions on Oct 2 (Issue 30). The path name matches, but we have not verified request compatibility.
Changelog Oct 6: five tiers become three, and organizations upgrade automatically as total credit purchases cross thresholds. Free: $100/month usage limit. Build: $5 purchased, $500/month. Launch: $100, $5,000/month. Grow: $500, $200,000/month. Standard limits per tier, e.g. Build Astra/Sol/Terra 5,000 RPM / 1M TPM, Luna 5,000 RPM / 2M TPM; Grow Luna 30,000 RPM / 180M TPM. Ultrafast limits are separate. Check Settings → Organization → Limits after the change.
mistral.ai post 2026-10-06T12:00Z ≈ 8 AM ET (modified Oct 7). 1 trillion parameters, 49B active, natively multimodal. Preview API on Mistral Studio now; “Weights drop end of this month.” Until then the model is red-teamed with “cybersecurity leaders, vetted partners, and state authorities, who will access the same model with reduced moderation and expanded cyber capabilities.” Trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own European datacenters; the preview is served there. 160+ languages. All of the following are Mistral-reported: Artificial Analysis Cyber Index top five; 82% on a reproduce-and-patch test (“highest of any model”); Cybench 93%; DeepSWE v1.1 61.7%; SWE-Atlas-QnA 59.4%; Terminal-Bench 4 28.3%; Coding Agent Index 49.8%; AutomationBench 59.9%; AA-Briefcase 1,393 Elo; Dense 200 visual grounding 42% vs GPT-6 Astra 41%. The post claims Claude Opus 5.5 and GPT-6 Astra “score near zero” on the patch test because they refuse. License and pricing are not stated in the post; architecture details are promised with the weights.
Newsroom Oct 6. Glasswing and CVP merge into one program. Defense Access: SOC, incident response, malware reverse-engineering, vulnerability analysis and validation; qualifying groups include in-house security teams, critical-infrastructure operators of any size, smaller security firms, open-source maintainers and individual researchers with a track record; review “within a few days”. Red Team Access: adds authorized pentesting and red-teaming; organizations only; review “a few weeks”, with Defense access in the meantime; real-time blocks remain for physical harm or mass disruption. Specialized Access: fewest blocks; for organizations authorized to test systems like flight OSes, power grids, telecom and interbank infrastructure; reviewed “in collaboration with the US government”; Glasswing members transition. Models: Opus 5.5, Sonnet 5.5, Mythos 5.1, and future models. Data retention required for monitoring until Enterprise Frontier Safeguards arrive “later this fall”; existing ZDR Fable 5.1 / Mythos 5.1 orgs can use CVP with ZDR. Anthropic-reported CyScenarioBench with Opus 5.5 (5 attempts × 10 challenges per tier): every GA task blocked on the first prompt; Defense 46/50 blocked; Red Team no blocks and 34/50 completed, in line with the 67.6% unsafeguarded rate. Glasswing partners uncovered at least 129,000 verified vulnerabilities Apr–Jul 2026, and Anthropic’s own scanning found 5,500 more Apr–Oct; more than 33,000 rated critical or high.
Release notes Oct 5: capabilities.thinking.types.disabled in GET /v1/models and GET /v1/models/{model_id} reports whether each model accepts thinking: {type: "disabled"}. Useful alongside the Sonnet 5.5 change where "disabled" is replaced by "between_tools". Issue 32 said nothing was newer than Oct 1 as of Tuesday morning. That was wrong, and this is the correction. Nothing newer than Oct 5 as of Wed Oct 7 morning ET.
Gemini API changelog Oct 6: gemini-nano-banana-2.1 is GA as the update to Nano Banana 2 (gemini-3.1-flash-image). Google cites improved visual quality, prompt adherence, character consistency and text rendering, plus 1:4, 4:1, 1:8 and 8:1 aspect ratios at 1K/2K/4K (Google claim). gemini-3.1-flash-image is deprecated with no shutdown date announced.
Published Oct 5 (feed 15:00 GMT ≈ 11 AM ET). In response to the EU AI Act: API customers globally can opt in to text watermarking for select models starting today, off by default. An invisible watermark is applied to eligible ChatGPT and Codex output in the EU only, over the coming weeks. Detector access is by application for approved researchers and expert organizations, not public. Cloud partners to follow in the coming weeks. The technique (“textGrain”) is to be open-sourced, and OpenAI says it matched or exceeded SynthID text. OpenAI-reported detection at 1% FPR: ~80% at 200 tokens vs ~95% at 400 tokens on psychology text, substantially lower on math. 10% synonym replacement drops detection 92% → 66%, and 25% drops it to 17%. No meaningful quality loss on Astra benchmarks (OpenAI-reported). Contrast: Claude release notes (Sep 1) say text from Claude Fable 5.1 and Mythos 5.1 carries Anthropic’s text watermark with no request changes.
Changelog Oct 5: eligible API organizations can review and accept a HIPAA Business Associate Agreement in Organization settings → General. Check your org’s eligibility before routing PHI.
Release notes Oct 1: the Models API returns a line field (e.g. opus) for grouping models by family. Issues 30–31 said release notes were quiet after Sep 30. That was wrong, and this is the correction. Nothing newer as of Tue Oct 6 morning ET.
OpenAI post Oct 2, 2026 (16:15Z ≈ 12:15 PM ET). Guidance, not a new SKU. Restates list prices per 1M tokens: GPT-6 Astra $10 in / $50 out / $1 cached; GPT-6.1 Sol $2 / $10 / $0.10; GPT-6 Luna $0.10 / $0.50 / $0.01. Recommends caching + compaction, measuring cost per successful task, matching model + reasoning effort to workload (effort can change mid-conversation without breaking cache), Fast mode for latency-sensitive apps, Ultrafast for Astra only; and rewriting skills / AGENTS.md with explicit decision boundaries instead of blanket “always ask.”
OpenAI API changelog (Sep 8 entry): gpt-rosalind-research GA through the trusted-access program for approved internal life-sciences research at $5 in / $0.50 cached / $25 out per 1M tokens; “Billing begins on October 5, 2026.” Watch item only for teams in that program.
OpenAI security post Wed Sep 30, 2026. Identified and disrupted a coordinated campaign designed to extract protected reasoning (adversarial distillation). Earliest observed activity first week of July. Operators did not break encryption, compromise a database, or gain direct access to stored user conversations — they manipulated model interactions so protected reasoning could be reproduced in requester-visible forms at scale (ToS violation). Independent researchers disclosed related cross-model / conversation-compaction paths; OpenAI confirmed attack paths were real. Jul 24–25 spikes: 16,000 requests with a relevant extraction pattern from over 4,000 users; related prompt-pattern activity across a cluster of more than 15,000 users, fully disrupted by Jul 28. Footnote: figures describe attempted, not necessarily successful, extractions. Attribution: a core cluster associated with individuals at Moonshot AI (Kimi); unclear whether all operators were a single actor. Response: account enforcement, stronger signup/infra controls, protections against replaying another user’s encrypted reasoning, streamed-output holds, partner coordination, Frontier Model Forum + government channels. Forward look: partner-hosted deployments need same protections; tool-output attacks require more than visible-text checks. Do not invent secondary drama or success rates beyond the post.
Claude Platform release notes + deprecations Tue Sep 30, 2026. Anthropic announced deprecation of Claude Sonnet 4.5 (claude-sonnet-4-5-20250929). Retirement on Claude API scheduled Nov 30, 2026; recommend migrate to Claude Sonnet 5.5. Deprecation page: dates apply to Anthropic-operated platforms (Claude API, Claude Platform on AWS, Microsoft Foundry); Bedrock / Google Cloud set their own retirement schedules. Console Usage export remains the audit path for deprecated-model traffic. Do not re-lead Sonnet 5.5 launch (Issue 27).
OpenAI product post September 29, 2026 (DevDay). Upgrade over GPT-6 Sol: near-Astra intelligence for agentic coding, computer use, and professional work at one-fifth of Astra’s standard API input/output token prices. API id gpt-6.1-sol. Standard API prices: $2 / MTok input, $0.10 / MTok cached input, $10 / MTok output. Family restated on the same page: GPT-6 Astra $10 / $50 ($1 cached); GPT-6 Luna $0.10 / $0.50 ($0.01 cached). Availability: starting today for Plus, Pro, Business, Enterprise, and Edu in ChatGPT Work and Codex; not yet available in Chat; API access as gpt-6.1-sol. Sol Ultrafast “in the coming days” (up to 8× token generation vs standard speed in Codex per the Sol post). DevDay Recap adds: Astra Ultrafast available today in ChatGPT Work and Codex on Pro 500 and Enterprise (up to 8× / ~300 tok/s in Codex, up to 6× in the API — vendor speed claims); Sol Ultrafast coming soon. Eval claims (coding, PDF professional Q&A, multi-step business workflows, computer use, scientific workflows, factuality) are OpenAI-reported — flag as vendor. Safety addendum (Deployment Safety Hub, published Sep 29): Preparedness Framework treats GPT-6.1 Sol as Critical in cybersecurity and High for Biological and Chemical capability; same safeguards stack as GPT-6 Astra. Do not re-lead the Sep 22 Sol/Luna launch.
Anthropic news + Claude Platform release notes September 28, 2026. Model id claude-sonnet-5-5 (Bedrock: anthropic.claude-sonnet-5-5 / Global CRIS global.anthropic.claude-sonnet-5-5). Pricing $2 / $10 per MTok input/output (same as Sonnet 5); cache reads $0.20/MTok. Anthropic claims typically fewer tokens → up to 30% less cost per task and outputs 30%+ faster than Sonnet 5 — vendor. Specs: 1M context, 128K max output, adaptive thinking, default effort high. First Sonnet with Opus-class cyber safeguards (higher-risk cyber tasks fall back to Sonnet 5); expanded Cyber Verification Program. Migration: use thinking: {"type": "between_tools"} (not "disabled"); forced tool_choice any/tool → 400. Haiku 5.5 “coming weeks” — not launched. Same-day AWS: Bedrock, Claude Platform on AWS, GovCloud. Vendor benches (Terminal-Bench 4.0 70.6%, CursorBench, GDPval-AA) are vendor / third-party eval mix. Do not re-lead Opus 5.5.
OpenAI API changelog section Sep 25, 2026 (no clock time). Fix for a bug in image encoding that degraded image understanding in Sol and Luna. Improves results on visual tasks in the API and Codex, including computer use. OpenAI recommends rerunning evals / retrying affected workflows if you use image inputs. Not a new model SKU — patch on Sep 22 launches (Issue 23). No quantitative before/after metrics on the changelog.
blog.google and Google Cloud blog 2026-09-24 (exact clock not published). Near–real-time video avatar with lip-sync; GA in Gemini Enterprise; 97 languages; async tool/API calls while dialogue continues; SynthID watermark on audio/video; custom avatars allowlist. Cloud blog: US and EU endpoints, provisioned throughput, enterprise compliance. Builds on prior-week Gemini 3.8 Live; Live Extended Thinking remains private preview. Distinct from Issue 24 Flash TTS / Flash-Lite TTS. Customer quotes and quality claims are vendor marketing.
Announcement and Claude Platform release notes September 22, 2026. API id claude-opus-5-5; $4/$20 per MTok; thinking always on; multi-cloud. Material update only (do not re-lead): AWS What’s New now confirms Opus 5.5 on Bedrock / Claude Platform on AWS and GovCloud (Sep 22). Flag vendor evals.
OpenAI product posts September 22, 2026. gpt-6-sol / gpt-6-luna at $2/$10 and $0.10/$0.50. Material update only: AWS What’s New for Sol & Luna on Bedrock (Sep 22). Do not re-lead.
