Skip to content
hubai.uk Ask HubAI
Menu
HUBAI NAVIGATOR

Start with the decision.

Build my AI stack
FIND YOUR FITAsk HubAITell us the job. Find tools that fit your budget and team.Start decision
HUBAI CHANGE RADAR

Know what changed.
Know what to do.

Launches, prices, access, integrations and safety updates—source-linked at source and translated into a buying decision.

PRICEWhat actually changes the bill?
RISKWhich control now needs review?
WORKFLOWWhat should buyers retest?

Latest AI product changes

Claude Dashboards and Motion: which plans get them, and what to review

Claude introduced Dashboards and Motion on 8 October. Dashboards is a beta for Pro, Max, Team and Enterprise; Motion is a Team and Enterprise beta. Docs, Slides and Design are now available across plans, including Free.

Buyer action: Check your actual plan and administrator settings; Inspect chart queries, source access and refresh dates; Review recipients and external sharing before publishing; Export one Motion animation and record correction time; Audit standalone Design links before 14 December

Official announcement — 8 October 2026 ↗

Anthropic OSS Scanner is free for accepted projects—but maintainers own triage

Anthropic launched an opt-in vulnerability-finding service on 8 October. Accepted open-source projects receive raw model reports; these are not human-validated findings. Maintainers must assess reports and candidate patches before acting.

Buyer action: Confirm that a core maintainer controls the enrolment; Name a triage owner and set review capacity; Reproduce findings before classifying severity; Test every patch against the project threat model; Confirm disclosure, pause and data-handling rules

Official announcement — 8 October 2026 ↗

Google Cloud Gemini agent: check identity, permissions and spending before delegating

Google Cloud announced Gemini agent on 8 October as an enterprise agent for delegated work. It is distinct from the consumer Gemini app and a standalone model API. Buyers need to verify account access, tool permissions and the complete bill.

Buyer action: Verify UK contracting, residency and account entitlement; Restrict identity, connector scopes and outbound tools; Require approval before irreversible actions; Test the spending cap and authorised resume path; Record model, tool, sandbox and reviewer cost

Official announcement — 8 October 2026 ↗

Microsoft releases a three-model MAI speech stack

MAI-Transcribe-2-Streaming adds incremental real-time transcription, while MAI-Voice-2.1 and Flash add multilingual speech at published character rates. Microsoft documents the Azure models as public previews without an SLA.

Buyer action: Run one consented end-to-end call flow with frozen audio and tools; measure partial and final errors, interruption recovery, accepted-call completion and every provider and review charge.

Microsoft MAI speech-model launch ↗

OpenAI releases GPT-6.1 Sol

The new Sol route keeps $2/$10 standard input/output pricing, cuts cached input to $0.10 per million tokens and launches across the API, ChatGPT Work and Codex, with a separate Copilot rollout.

Buyer action: Replay frozen GPT-6 Sol tasks with identical tools, permissions and acceptance criteria; measure accepted work, persistence after correction, review and complete billed cost.

OpenAI GPT-6.1 Sol launch ↗

OpenAI begins the Dots always-on agent rollout

Dots use GPT-6 Astra, a separate cloud computer and user-selected connections to continue recurring work and report through ChatGPT or collaboration surfaces.

Buyer action: Pilot one least-privilege read-only routine for two weeks; verify approvals, auto-review, logs, stop, recovery and complete cost before enabling consequential actions.

OpenAI Dots launch ↗

OpenAI releases GPT-6 Sol and GPT-6 Luna

The new family pairs a workhorse model with a low-cost volume model, both with published 1.05M context windows, API rates and product-surface rollouts.

Buyer action: Test Luna first on frozen tasks, escalate failures to Sol and record accepted output, review, tools, retries, context and complete cost.

OpenAI GPT-6 Sol and Luna launch ↗

Anthropic releases Claude Opus 5.5

The first Claude 5.5 model adds lower published Opus rates, a 1M context window and always-on adaptive thinking with migration changes for tool workflows.

Buyer action: Pin the new model in a test environment and compare accepted tasks, tool failures, cache use, latency, review and complete cost before migration.

Anthropic Claude Opus 5.5 launch ↗

xAI releases Grok 4.7 and GitHub begins Copilot rollout

The new model adds a 500K API context window, configurable reasoning, published standard token rates and availability across xAI, coding tools and paid Copilot plans.

Buyer action: Run the same bounded tasks with fixed tools and permissions; measure accepted output, review, cache behaviour, failures and complete billed cost.

xAI Grok 4.7 launch ↗

OpenAI launches Astra for Law for selected firms

The specialist route combines GPT‑6 Astra with a US legal search index, firm instructions, ChatGPT for Word and partner-built legal plugins.

Buyer action: Pilot anonymised UK matters against the firm's authoritative research baseline and measure citation treatment, reviewer time, permissions and complete cost.

OpenAI Astra for Law launch ↗

OpenAI launches ChatGPT for Financial Services

The separate financial-institution plan combines GPT‑6 Astra, included premium datasets, entitlement-linked sources and firm-controlled output templates.

Buyer action: Pilot one reconciled workflow, verify source rights and delays, and compare complete accepted-deliverable cost with the current data and productivity stack.

OpenAI financial services launch ↗

OpenAI releases Agents API in public beta

Developers can use the managed Codex harness for durable agent sessions with tools, optional subagents and a choice of execution environment.

Buyer action: Run one bounded multi-step task against the current stack and record success, intervention, recovery, latency and every model, tool and environment cost.

OpenAI launch announcement ↗
Filter the interactive change radar
FILTER SIGNALS
48 verified changes
12 Oct 202601
Launch OFFICIAL SOURCE

Muse Wonder and Full launch with different access routes

Wonder has a preview API; Full is available through upper Muse Code subscriptions.

BUYER IMPACT
Model identity, plan allowance and data mode change the purchase.
DO NEXT
Check Contributor settings and run the same repository task before upgrading.
13 Oct 202602
Access OFFICIAL SOURCE

Gemini 4 Argon opens public paid access

Paid API and eligible app plans open; introductory standard rates end on 30 November.

BUYER IMPACT
Use December rates and shared-context limits in the business case.
DO NEXT
Confirm the actual route and test long-decode storage and recovery.
13 Oct 202603
Integration OFFICIAL SOURCE

Cursor Origin opens public app submissions

Repository integrations can be submitted for review before marketplace listing.

BUYER IMPACT
Consent, selected repositories and active installations require separate review.
DO NEXT
Pilot one integration and prove uninstall from both systems.
8 Oct 202604
Launch OFFICIAL SOURCE

Claude Dashboards and Motion: which plans get them, and what to review

Claude introduced Dashboards and Motion on 8 October. Dashboards is a beta for Pro, Max, Team and Enterprise; Motion is a Team and Enterprise beta. Docs, Slides and Design are now available across plans, including Free.

BUYER IMPACT
The decision is whether your existing plan can finish a useful workflow. A new dashboard does not make its query correct, and animation created with code is a different tool from footage generation.
DO NEXT
Check your actual plan and administrator settings; Inspect chart queries, source access and refresh dates; Review recipients and external sharing before publishing; Export one Motion animation and record correction time; Audit standalone Design links before 14 December
8 Oct 202605
Safety OFFICIAL SOURCE

Anthropic OSS Scanner is free for accepted projects—but maintainers own triage

Anthropic launched an opt-in vulnerability-finding service on 8 October. Accepted open-source projects receive raw model reports; these are not human-validated findings. Maintainers must assess reports and candidate patches before acting.

BUYER IMPACT
Free scanning can still create a review workload. Decide who will reproduce findings, evaluate suggested changes and communicate confirmed vulnerabilities before enrolling a critical dependency.
DO NEXT
Confirm that a core maintainer controls the enrolment; Name a triage owner and set review capacity; Reproduce findings before classifying severity; Test every patch against the project threat model; Confirm disclosure, pause and data-handling rules
8 Oct 202606
Launch OFFICIAL SOURCE

Google Cloud Gemini agent: check identity, permissions and spending before delegating

Google Cloud announced Gemini agent on 8 October as an enterprise agent for delegated work. It is distinct from the consumer Gemini app and a standalone model API. Buyers need to verify account access, tool permissions and the complete bill.

BUYER IMPACT
Persistent execution moves the buying decision beyond answer quality. A workflow can touch several systems and accumulate model and execution charges while the person who assigned it is away.
DO NEXT
Verify UK contracting, residency and account entitlement; Restrict identity, connector scopes and outbound tools; Require approval before irreversible actions; Test the spending cap and authorised resume path; Record model, tool, sandbox and reviewer cost
7 Oct 202607
Access OFFICIAL SOURCE

ChatGPT starts GPT-6 Intelligent UI rollout in Chat

Paid Chat tiers start 7 October; Free and Go expansion starts 8 October. Pro reasoning does not support Intelligent UI; Work and Codex models are unchanged.

BUYER IMPACT
Plan, reasoning option, account rollout and enterprise settings decide access. Visual polish is not evidence of correct calculations.
DO NEXT
Try a known-answer task in the existing plan; check the completed answer, sources, changed inputs and workplace settings.
7 Oct 202608
Launch OFFICIAL SOURCE

Haiku 5.5 launches with prompt-length pricing and migration changes

Haiku base rates depend on the 100K prompt threshold. Sonnet 5.5 cache reads fall to $0.10/M; eligible Max and Team API credits have separate claim and expiry rules.

BUYER IMPACT
Task cost depends on prompt tokens, effort, output acceptance and credit boundaries; a lower rate is not a guaranteed workflow saving.
DO NEXT
Migrate one labelled queue, recount tokens and measure accepted-work cost with escalation and spending controls.
6 Oct 202609
Launch OFFICIAL SOURCE

EmbeddingGemma 2 adds local multimodal retrieval

Google publishes open weights under Apache 2.0 for embeddings across text, code, images, video and audio. The announced Android ML Kit service remains a future release.

BUYER IMPACT
Local inference can support a local data boundary, but the application operator still owns index permissions, deletion, data flows and complete operating cost.
DO NEXT
Pin one collection, checkpoint and encoder configuration; compare labelled retrieval with current search on the actual devices.
7 Oct 2026010
Access OFFICIAL SOURCE

Claude expands verified cyber access into three tiers

Defense, Red Team and Specialized Access define permitted work and controls. CVP now covers Mythos 5.1, Opus 5.5 and Sonnet 5.5; this is not a new model launch.

BUYER IMPACT
Approval does not establish UK entitlement, free usage, client-facing productization permission or independent safety. Identity readiness and retention determine deployment fit.
DO NEXT
Confirm the target scope, account route and retention; prepare phishing-resistant MFA and short-lived credentials, then run a bounded authorised pilot.
5 Oct 2026011
Integration OFFICIAL SOURCE

ChatGPT announces a US visual-ad test and expanded measurement

A new visual placement during image generation is planned for later in October in the US. Conversion integrations and attribution options expand; brand-suitability evaluations remain pilots.

BUYER IMPACT
Existing UK ads do not establish UK access to this format. Reporting attribution and campaign optimisation are different, and attributed orders do not establish incremental customers.
DO NEXT
Confirm market and placement eligibility, define a baseline, document event data and billing, and set a pilot cost ceiling before moving budget.
1 Oct 2026012
Access OFFICIAL SOURCE

GitHub Copilot adds desktop computer use on macOS and Windows

Local Copilot CLI and Copilot app sessions can read and operate approved desktop applications, including clicking, typing, scrolling and dragging through GUI workflows.

BUYER IMPACT
The preview reaches legacy applications without a suitable API, CLI or MCP connector, but app approval, visible screen content, saved permissions and operating-system access become part of the agent boundary.
DO NEXT
Pilot one reversible task with one-time approval; test hostile on-screen instructions, wrong-window recovery, revocation, logs and complete AI-credit and review cost before widening access.
1 Oct 2026013
Launch OFFICIAL SOURCE

Microsoft releases a three-model MAI speech stack

MAI-Transcribe-2-Streaming adds incremental real-time transcription, while MAI-Voice-2.1 and Flash add multilingual speech at published character rates. Microsoft documents the Azure models as public previews without an SLA.

BUYER IMPACT
The component rates make a controlled pilot budget possible, but the temporary transcription price, preview status, reasoning layer, telephony and human escalation still determine production fit and complete cost.
DO NEXT
Run one consented end-to-end call flow with frozen audio and tools; measure partial and final errors, interruption recovery, accepted-call completion and every provider and review charge.
30 Sep 2026014
Launch OFFICIAL SOURCE

Google announces Gemini 4 Argon with restricted Fairwind access

General paid availability is planned. Announced introductory rates change after an undated introductory period; a public API identifier and UK entitlement remain unverified.

BUYER IMPACT
Admission and actual route access come before procurement. A future rate card does not establish a model available today.
DO NEXT
Verify access and terms, budget both price stages, then compare accepted authorised work and full cost with the current workflow.
1 Oct 2026015
Launch OFFICIAL SOURCE

OpenAI releases GPT-6.1 Sol

The new Sol route keeps $2/$10 standard input/output pricing, cuts cached input to $0.10 per million tokens and launches across the API, ChatGPT Work and Codex, with a separate Copilot rollout.

BUYER IMPACT
Teams can test the new route without a higher headline token tier, but product-surface access, tools, agent persistence, review and complete cost still need migration evidence.
DO NEXT
Replay frozen GPT-6 Sol tasks with identical tools, permissions and acceptance criteria; measure accepted work, persistence after correction, review and complete billed cost.
30 Sep 2026016
Launch OFFICIAL SOURCE

OpenAI begins the Dots always-on agent rollout

Dots use GPT-6 Astra, a separate cloud computer and user-selected connections to continue recurring work and report through ChatGPT or collaboration surfaces.

BUYER IMPACT
Pro, Business Premium and Enterprise are named for the first eligible-market rollout, but Dots-specific UK entitlement, universal limits and a separate rate card are not publicly established.
DO NEXT
Pilot one least-privilege read-only routine for two weeks; verify approvals, auto-review, logs, stop, recovery and complete cost before enabling consequential actions.
29 Sep 2026017
Safety OFFICIAL SOURCE

NVIDIA launches OpenShell and a separate Sentry reference design

OpenShell provides open-source sandbox, policy, credential and audit controls, while Sentry adds an out-of-band layer on BlueField-4 DPUs.

BUYER IMPACT
Teams can evaluate the software runtime without NVIDIA hardware, but the full hardware-backed architecture is a separate infrastructure and procurement decision.
DO NEXT
Pilot one read-only agent, prove denied writes and credential isolation, then measure whether the threat model justifies BlueField-4 and Sentry.
28 Sep 2026018
Launch OFFICIAL SOURCE

Anthropic releases Claude Sonnet 5.5 and GitHub starts Copilot rollout

The model keeps Sonnet's published $2/$10 token rates, adds a 1M context window and changes thinking, tool and computer-use behaviour that can affect existing agents.

BUYER IMPACT
Teams gain a lower-rate alternative to Opus 5.5 across major cloud and coding routes, but a silent model swap can break orchestration or hide expected tool-call text.
DO NEXT
Pin the model in a test environment and compare accepted work, tool failures, logs, cache use, latency, safeguards, review and complete cost before migration.
28 Sep 2026019
Integration OFFICIAL SOURCE

Gemini begins a 14-service Connected Apps rollout

Google names new project, database, document, design, website, housing, credit, fitness and ticket services that can be connected from Gemini settings or invoked in chat.

BUYER IMPACT
The integrations can reduce switching, but rollout, Workspace access, service subscriptions, data handling and permitted actions remain account-specific.
DO NEXT
Connect one read-only service with least privilege, inspect the data path and logs, then require review before any write, publish, booking or purchase.
27 Sep 2026020
Safety OFFICIAL SOURCE

GitHub Copilot adds local sandboxing and managed agent telemetry

The Copilot app can limit local-session filesystem, network and credential access, while enterprise-managed OpenTelemetry can export model and tool traces.

BUYER IMPACT
The controls improve prevention and investigation, but sandboxing is off by default, applies to new or restarted local sessions and does not cover remote or cloud sandboxes.
DO NEXT
Enable the sandbox for one non-production project, keep content capture minimal and verify permissions, traces, accepted changes and rollback across every intended client.
26 Sep 2026021
Integration OFFICIAL SOURCE

Claude opens a tracked plugin directory submission portal

Paid-plan developers can submit a remote MCP connector or a GitHub-hosted plugin bundle, see validation and safety-scan feedback, and inspect version-level usage after publication.

BUYER IMPACT
The route improves distribution and review visibility, but directory approval does not establish secure permissions, correct results or suitability for a buyer's deployment.
DO NEXT
Separate directory approval from procurement: inspect scopes, data paths, maintainer evidence and rollback, then run one least-privilege read-only pilot.
26 Sep 2026022
Launch OFFICIAL SOURCE

Microsoft reorganises Copilot around Home, Code and Autopilot

Home combines Chat, Cowork and Office work, Code builds apps and workflows, and Autopilot introduces a persistent agent with tenant-hosted identity and workspace.

BUYER IMPACT
Home and Code begin in Frontier, Autopilot remains a private preview, and long-running Cowork, Code and Autopilot work uses usage-based billing beyond the seat licence.
DO NEXT
Pilot one reversible workflow with least privilege, human approval, an explicit credit cap and complete accepted-work cost measurement.
25 Sep 2026023
Integration OFFICIAL SOURCE

Claude Tag adds personal connectors to requested Slack channel work

A requester can use their own connected tools in a channel; scheduled and self-started work still requires admin-attached shared connectors.

BUYER IMPACT
A personal source account stays scoped to the requester, but posted answers are visible to channel members. Team rollout has started; Enterprise follows.
DO NEXT
Start in review mode with a read-only connector, verify output exposure and audit logs, then test unattended routines separately.
23 Sep 2026024
Launch OFFICIAL SOURCE

OpenAI releases GPT-6 Sol and GPT-6 Luna

The new family pairs a workhorse model with a low-cost volume model, both with published 1.05M context windows, API rates and product-surface rollouts.

BUYER IMPACT
Teams gain a clear price ladder, but inputs above 272K use higher rates and provider benchmarks do not establish accepted-task cost.
DO NEXT
Test Luna first on frozen tasks, escalate failures to Sol and record accepted output, review, tools, retries, context and complete cost.
23 Sep 2026025
Launch OFFICIAL SOURCE

Anthropic releases Claude Opus 5.5

The first Claude 5.5 model adds lower published Opus rates, a 1M context window and always-on adaptive thinking with migration changes for tool workflows.

BUYER IMPACT
Long-running agent teams may reduce token rates, but forced tools, thinking blocks and older computer-use integrations need retesting.
DO NEXT
Pin the new model in a test environment and compare accepted tasks, tool failures, cache use, latency, review and complete cost before migration.
22 Sep 2026026
Launch OFFICIAL SOURCE

xAI releases Grok 4.7 and GitHub begins Copilot rollout

The new model adds a 500K API context window, configurable reasoning, published standard token rates and availability across xAI, coding tools and paid Copilot plans.

BUYER IMPACT
Teams gain another long-context coding and agent candidate, but provider benchmarks, route-specific prices and staged access do not establish lower accepted-task cost.
DO NEXT
Run the same bounded tasks with fixed tools and permissions; measure accepted output, review, cache behaviour, failures and complete billed cost.
21 Sep 2026027
Launch OFFICIAL SOURCE

Anthropic redesigns Claude Code Projects around parallel threads

One coordinating conversation can delegate work to separate Claude Code cloud sessions, share project memory and assemble the result.

BUYER IMPACT
Selected Pro and Max users can test parallel repository work, but every thread uses plan allowance and overlapping edits can create merge conflicts.
DO NEXT
Pilot one separable repository task and measure accepted changes, conflicts, review minutes, permissions and complete usage.
20 Sep 2026028
Launch OFFICIAL SOURCE

OpenAI launches Astra for Law for selected firms

The specialist route combines GPT‑6 Astra with a US legal search index, firm instructions, ChatGPT for Word and partner-built legal plugins.

BUYER IMPACT
Eligible firms gain a more integrated legal workflow, but UK authority coverage is not established, access and pricing are controlled, and the API is still coming soon.
DO NEXT
Pilot anonymised UK matters against the firm's authoritative research baseline and measure citation treatment, reviewer time, permissions and complete cost.
19 Sep 2026029
Launch OFFICIAL SOURCE

Google releases Gemini 3.8 Live and Extended Thinking

The new real-time models add asynchronous tool calls, visual context and a separate background-reasoning route for more complex live work.

BUYER IMPACT
Voice-agent teams now have a latency-led default and a deeper-reasoning option at published API rates, but complete resolved-call cost and quality still require a same-flow test.
DO NEXT
Pilot both model codes on identical calls and record completion, interruption, tool failures, escalation, latency and complete cost.
15 Sep 2026030
Access OFFICIAL SOURCE

Gemini Windows app: launch reviewed, account boundaries checked

Google launched Windows access on 10 September. HubAI reviewed the official download, requirements and platform differences on 15 September.

BUYER IMPACT
A faster access route does not unlock every paid feature or every native Mac capability; account permissions still determine approved business use.
DO NEXT
Compare one accepted task with the browser baseline, then check account-specific UK pricing and data rules before expanding use.
13 Sep 2026031
Launch OFFICIAL SOURCE

Cognition brings Fusion to Devin Desktop and CLI

A frontier lead plans and reviews while a lower-cost sidekick executes delegated work in its own persistent context.

BUYER IMPACT
Teams can test multi-model coding without building a router, but reported savings vary by task and do not prove lower accepted-change cost on their repositories.
DO NEXT
Benchmark representative repository tasks against the current single-agent baseline, including retries, review, tests and recovery.
13 Sep 2026032
Launch OFFICIAL SOURCE

Cognition releases SWE-2 for agentic coding

The coding model adds configurable effort and is available in Devin Desktop and CLI, including as a recommended Fusion sidekick.

BUYER IMPACT
Paid Devin users can evaluate SWE-2 under a time-limited promotion, but future terms and independent task performance remain unverified.
DO NEXT
Record the plan, effort level, promotion boundary and complete cost per accepted repository change.
12 Sep 2026033
Launch OFFICIAL SOURCE

OpenAI launches ChatGPT for Financial Services

The separate financial-institution plan combines GPT‑6 Astra, included premium datasets, entitlement-linked sources and firm-controlled output templates.

BUYER IMPACT
Investment-banking and equity-research teams can move from source-linked research to editable models and client materials in one governed workspace, but price is private and dataset coverage is conditional.
DO NEXT
Pilot one reconciled workflow, verify source rights and delays, and compare complete accepted-deliverable cost with the current data and productivity stack.
12 Sep 2026034
Access OFFICIAL SOURCE

GPT‑Rosalind leaves research preview with expanded trusted access

Eligible organisations can now request global access to OpenAI's specialist life-sciences model through ChatGPT, Codex and the API.

BUYER IMPACT
Research teams have a published API price and broader route to evaluation, but access remains controlled and Rosalind Workbench is still in research preview.
DO NEXT
Apply only with a bounded internal-research workflow, expert acceptance rule, controlled data path and complete-cost measurement plan.
11 Sep 2026035
Launch OFFICIAL SOURCE

OpenAI releases Agents API in public beta

Developers can use the managed Codex harness for durable agent sessions with tools, optional subagents and a choice of execution environment.

BUYER IMPACT
Teams can outsource more orchestration infrastructure, but complete cost, permissions, recovery and beta change risk need direct testing.
DO NEXT
Run one bounded multi-step task against the current stack and record success, intervention, recovery, latency and every model, tool and environment cost.
11 Sep 2026036
Launch OFFICIAL SOURCE

OpenAI releases GPT-Live-1 in the API

The full-duplex voice layer can listen and speak together, handle interruptions and delegate deeper work to a backend model.

BUYER IMPACT
The $0.05-per-minute voice rate is only one part of total resolved-call cost.
DO NEXT
Compare identical call flows on completion, interruption errors, latency, escalation and complete cost.
10 Sep 2026037
Launch OFFICIAL SOURCE

OpenAI launches the Data agent in ChatGPT Work

The new plugin investigates approved company data, creates interactive dashboards and can prepare actions through connected tools.

BUYER IMPACT
Business users can question governed data without writing queries, but useful results still depend on semantic definitions, source permissions and publication controls.
DO NEXT
Pilot one recurring question against an analyst-approved baseline and measure errors, correction time, refresh reliability and complete cost.
10 Sep 2026038
Safety OFFICIAL SOURCE

GitHub Copilot adds centrally managed agent permissions

Business and Enterprise administrators can classify supported shell, file and network operations as denied, approval-required or allowed.

BUYER IMPACT
A company can enforce a minimum agent permission boundary that local user settings and saved approvals cannot weaken.
DO NEXT
Start with destructive commands, sensitive paths and unapproved domains; widen access only after a bounded repository pilot.
10 Sep 2026039
Safety OFFICIAL SOURCE

OpenAI details Astra controls for consequential enterprise work

Administrators can restrict approved websites and desktop apps, manage uploads, downloads and browsing history, and require confirmation for consequential actions.

BUYER IMPACT
Astra's deployment envelope can be narrowed before teams give a computer-use agent broader access.
DO NEXT
Confirm the exact admin controls in the contracted product and test every consequential action path before expansion.
9 Sep 2026040
Launch OFFICIAL SOURCE

OpenAI releases Images 2.5 with Flare and Sunburst API routes

The image system adds more precise editing and separates a speed-led Flare route from a quality-led Sunburst route.

BUYER IMPACT
Teams can match model choice to volume or precision, but both routes need complete-task cost and acceptance testing.
DO NEXT
Freeze one production brief, run both variants and compare approvals, retries, latency and billed tokens.
7 Sep 2026041
Launch OFFICIAL SOURCE

Google releases Lyria 3.5 in Gemini and the Gemini API

The model generates structured, vocal full songs from text or image references at a published API price of $0.08 per request.

BUYER IMPACT
Low request pricing makes API-scale music generation easier to test, but revision count and rights review determine the real production cost.
DO NEXT
Benchmark the same brief against Suno and Udio and verify commercial-use terms before distribution.
5 Sep 2026042
Launch OFFICIAL SOURCE

GitHub opens the HydraFusion research preview

Copilot CLI can test single-model, cascade and separate-critic workflows through experimental settings.

BUYER IMPACT
One coding request may now use more than one model, changing both cost and latency.
DO NEXT
Benchmark every invoked model and the complete accepted-code cost before production use.
5 Sep 2026043
Integration OFFICIAL SOURCE

GPT-6 Astra begins its GitHub Copilot rollout

Eligible Copilot plans gain Astra across IDE, CLI, coding-agent, web and mobile surfaces through a gradual release.

BUYER IMPACT
Astra can be tested inside real repositories without building a separate API integration.
DO NEXT
Check plan eligibility, administrator policy and usage billing before changing the default model.
5 Sep 2026044
Pricing OFFICIAL SOURCE

Claude Fable 5.1 cuts the cost of cached context

Headline token pricing remains premium while published cache-read pricing falls to $0.25 per million tokens.

BUYER IMPACT
Persistent agents and large-codebase workflows may have different economics from one-shot prompts.
DO NEXT
Measure cache-hit rate and total cost per accepted task instead of comparing headline token prices.
4 Sep 2026045
Access OFFICIAL SOURCE

GPT-6 Astra starts a staged paid-plan release

OpenAI announced staged ChatGPT and API access for a frontier model focused on computer use and tool-heavy work.

BUYER IMPACT
Availability depends on plan, organisation settings and deployment route rather than one universal launch state.
DO NEXT
Confirm the exact account route and model identifier before planning migration work.
4 Sep 2026046
Safety OFFICIAL SOURCE

Astra reaches OpenAI’s critical cyber capability threshold

The release carries stronger isolation, monitoring and access-control requirements alongside its new capabilities.

BUYER IMPACT
Security behaviour and operational controls become part of the buying decision, not an appendix.
DO NEXT
Test legitimate security workflows, interruption behaviour and administrator controls before approval.
4 Sep 2026047
Safety OFFICIAL SOURCE

Anthropic adds enterprise frontier safeguards

The announced enterprise layer combines customer-controlled infrastructure choices with misuse-monitoring controls.

BUYER IMPACT
Regulated buyers can evaluate who holds evidence, encryption keys and review authority.
DO NEXT
Confirm eligibility, rollout phase, storage cost and the exact zero-data-retention arrangement.
2 Sep 2026048
Launch OFFICIAL SOURCE

Gemini 3.8 Flash moves into general availability

Google positions the model for long-context coding, agents and tunable reasoning with introductory API pricing.

BUYER IMPACT
A lower headline price may still produce a higher task bill when long agent runs consume more tokens.
DO NEXT
Measure total tokens, retries, latency and accepted-task rate against the current production model.
EDITORIAL BOUNDARY

A changelog is not a recommendation.

HubAI links to the primary announcement, states the buyer impact and withholds a score when independent evidence is missing. Sponsored visibility cannot create or rewrite a change record.

Read the evidence standard →