Frontier review: model changes, routing, and permissions
Eight dated vendor events connect model releases and retirement with the routing defaults and permissions that shape deployed workflows.
Review
This review brings model releases and succession together with changes to runtime defaults and connector permissions. OpenAI, Anthropic, Google, and xAI report dated model changes; Anthropic also completed a previously announced API retirement. Perplexity reports new Computer model defaults and connector approval controls. Nova3's operational interpretation is to evaluate representative work before changing model defaults, confirm actual account access, and review the permissions inherited by recurring workflows. These recommendations are separate from the vendor facts below.
Dated events and sources
OpenAI released GPT-6 Astra
Fact: OpenAI's September 3, 2026 safety overview announces the release of GPT-6 Astra.
Nova3 operational read: Teams considering Astra should evaluate representative tasks and deployment controls before changing production defaults.
Uncertainty: The launch describes a staged rollout beginning with selected organizations. A release announcement does not establish access for every account; confirm availability and administrator settings for the intended deployment.
OpenAI: Safety overview: GPT-6 AstraAnthropic introduced Claude Fable 5.1
Fact: Anthropic's Fable page dates the Fable 5.1 announcement September 1, 2026 and lists availability through Claude products, the Claude Platform, and supported marketplaces.
Nova3 operational read: Teams evaluating a move should compare task quality, latency, safeguards, retention requirements, and cost on their own workload before changing a production default.
Uncertainty: Availability and pricing can vary by product, marketplace, region, and account eligibility. Check the current vendor documentation before deployment.
Anthropic: Claude FableGoogle introduced Gemini 3.8 Flash and 3.8 Flash Cyber
Fact: Google's announcement is dated September 2, 2026. It says Gemini 3.8 Flash is available through the Gemini API and selected Google products, while Flash Cyber is limited to trusted defenders through the Fairwind Program.
Nova3 operational read: The two variants have different access conditions. A deployment plan should distinguish the broadly available Flash model from the restricted cybersecurity model.
Uncertainty: The announcement describes introductory pricing and multiple product surfaces. Confirm current prices and access in the relevant Google product before purchase or migration.
Google: Introducing Gemini 3.8 Flash and 3.8 Flash CyberAnthropic introduced Claude Opus 5
Fact: Anthropic's product announcement is dated July 24, 2026 and says Opus 5 became available that day.
Nova3 operational read: This dated release is part of the model-succession record teams should use when evaluating saved model identifiers and migration plans.
Uncertainty: The vendor's performance comparisons describe its evaluations and early-access reports. Independent results can differ by task and harness.
Anthropic: Introducing Claude Opus 5Anthropic retired Claude Opus 4.1 from its API
Fact: Anthropic's model-deprecation documentation records August 5, 2026 as the retirement date for claude-opus-4-1-20250805 and names Claude Opus 4.8 as its recommended replacement.
Nova3 operational read: Teams should verify that saved model identifiers and regression tests cover the replacement before a retirement date reaches production.
Uncertainty: This statement is limited to the API model and replacement shown in Anthropic's deprecation table.
Claude Platform: Model deprecationsxAI introduced Grok 4.5
Fact: xAI's announcement is dated July 16, 2026 and describes Grok 4.5 as a model for coding, agentic tasks, and knowledge work.
Nova3 operational read: The date matters because it establishes the starting point for the later Grok 4.6 succession claim.
Uncertainty: Capability claims in the announcement are vendor claims. Production suitability still depends on workload-specific evaluation.
xAI: Introducing Grok 4.5xAI introduced Grok 4.6
Fact: xAI's announcement is dated August 12, 2026 and says Grok 4.6 builds on Grok 4.5 with a focus on long-running agents and interactive and visual work.
Nova3 operational read: A version change is a reason to rerun representative evaluations. It does not by itself prove that an existing workflow should migrate.
Uncertainty: The announcement's benchmark and capability statements are vendor-reported. Check current API and product availability separately.
xAI: Introducing Grok 4.6Perplexity updated Computer model defaults and connector approvals
Fact: Perplexity's August 24, 2026 changelog makes GPT-5.6 Terra the default for Computer subagents and GPT-5.6 Luna the primary model for scheduled automations. It also describes an Always ask connector control: users can approve one action, grant access for the thread, or deny it. Recurring runs inherit thread-level grants.
Nova3 operational read: Teams should check model routing and existing thread grants when reviewing scheduled workflows, including which connector actions require a fresh approval.
Uncertainty: This is a dated defaults and permissions update. It does not establish that every recurring run will request approval: thread-level grants can carry forward. Verify current settings for the relevant connector and thread.
Perplexity: Computer in Email, GPT-5.6 Terra and Luna, and Grok 4.6Editorial 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.