AI Tool Changelog

OpenAI releases GPT Image 2.5

A weekly AI briefing on 2026-09-18 announced the launch of OpenAI's GPT Image 2.5, indicating ongoing iteration and improvement in multimodal models.

Sources: distillintelligence.com

Anthropic reports Claude now performs 26% of the work to build the next version of itself

On 2026-09-18, reports cited Anthropic data showing that as of August, its AI model Claude was performing 26% of the work involved in creating the next iteration of Claude. This internal benchmark illustrates the accelerating role of AI in its own development.

Sources: blog.buildfastwithai.com

Google DeepMind releases Gemini 3.8 with enhanced reasoning

A weekly AI briefing on 2026-09-18 reported that Google DeepMind released Gemini 3.8, a new model version featuring improved reasoning capabilities.

Sources: distillintelligence.com

OpenAI launches Astra for Law, its first vertical-specific version of GPT-6 Astra

On 2026-09-17, OpenAI introduced Astra for Law, a version of its GPT-6 Astra model specifically configured for legal research and document drafting. This marks OpenAI's first major step into creating verticalized, industry-specific AI tools.

Sources: blog.buildfastwithai.com

Meta opts out of frontier-lab safety pact, citing individual responsibility

On 2026-09-16, it was reported that Meta is not participating in the safety standards effort involving OpenAI, Anthropic, and Google DeepMind. Mark Zuckerberg stated on X that each lab has the incentive and responsibility to manage its own safety without a formal pact, highlighting a strategic divergence on self-regulation.

Sources: dailyaidigest.net

OpenAI confirms weeks of safety talks and support for third-party evaluators

On 2026-09-16, OpenAI's global policy chief, Chris Lehane, confirmed weeks of collaboration with Anthropic and Google DeepMind on shared safety standards. OpenAI also signaled it would embed third-party evaluators to monitor safety internally, a commitment similar to Anthropic's. The coordination includes developing shared evaluation frameworks and red-teaming methods.

Sources: theaiinsider.tech, plainenglish.io

OpenAI, Anthropic, and Google DeepMind confirm joint effort to establish AI safety standards

On 2026-09-15, OpenAI confirmed its collaboration with rivals Anthropic and Google DeepMind on AI safety to address risks from advanced models. The companies stated they are not seeking an antitrust waiver. The effort is reportedly modeled on the Financial Industry Regulatory Authority (FINRA) to oversee safety auditing and pre-release testing.

Sources: dailyaidigest.net

Multiple major AI labs release model updates in the same week, signaling accelerated release cadence

CNBC reported that Anthropic, OpenAI, Meta, and Google all released significant model updates within the same week, a period the outlet described as creating 'model fatigue' amid a high-density launch schedule in early September.

Sources: cnbc.com

K2 Horizon open-weight models released with published weights, code, and data

The Institute of Foundation Models at MBZUAI released K2 Horizon, a series of six open-weight models with parameter sizes ranging from 0.9B to 375B. The release includes the model weights, code, training data, and methodology.

Sources: byobot.ai, hex2077.dev

Meta releases Muse Spark 1.3

According to industry news roundups, Meta released the Muse Spark 1.3 model on September 2, 2026.

Sources: local-ai-zone.github.io, note.com

Google releases Gemini 3.8 Flash and a restricted Gemini 3.8 Flash Cyber variant

According to multiple reports, Google released two new models, Gemini 3.8 Flash and Gemini 3.8 Flash Cyber, on September 2, 2026.

Sources: capitalandcompute.net, local-ai-zone.github.io

Release notes for versions 2.1.260 and 2.1.261 show new prompt-cache diagnostics and a '/skill-doctor' command

Updates to Claude Code introduced improved diagnostics for the prompt cache and a new '/skill-doctor' command designed to help users identify unused skills and reduce context costs.

Sources: new-beginnings.co.jp

GitHub Copilot code review expands support to bot‑authored and very large pull requests.

An AI tool‑updates roundup on August 28 reports that Copilot’s code review feature now handles two PR types that were previously excluded: reviews automatically requested on pull requests authored by bots (including Copilot’s own cloud agent) and very large pull requests. This change broadens Copilot’s applicability to more real‑world repository workflows, especially in organizations relying heavily on automation and monorepo or bulk‑change PRs.

Sources: library.mikesailab.com

Midjourney launches V8.2 image‑editing model for public testing, adding multi‑image reference and canvas expansion.

Midjourney officially announced its first V8.2 image‑editing model for public testing, targeting a broad user base of creators. Coverage notes that V8.2 supports instruction‑based local edits, mixing up to four reference images for image‑to‑image generation, and seamless canvas expansion, with improved handling of complex natural‑language edit commands. This update significantly expands Midjourney’s capabilities beyond pure text‑to‑image into more granular editing workflows.

Sources: news.aibase.com, communeify.com

Claude Developer Platform updates skills model and adds personal and service‑account keys.

Release notes dated August 27 describe changes to how the Developer Platform handles files and skills: beta and non‑beta skill “shapes” are aligned, BetaSkill is renamed to BetaContainerSkill, and skill deletion behavior is adjusted to better match workspace expectations. The same update introduces personal keys and service‑account keys within the Claude Console, enabling more fine‑grained, workspace‑scoped access control and clearer account‑level tracking for teams building on Claude.

Sources: releasebot.io

Anthropic unifies Claude’s memory across chat and Cowork, with user‑visible controls.

Anthropic announced a major update that merges Claude’s memory between regular chat and the Claude Cowork agent, so preferences, project context, and client details remembered in chat are now available to Cowork tasks, and vice versa. Users can see what Claude has stored by topic, edit or delete items, and opt in to including *sensitive* topics; memory is enabled by default for Free, Pro, and Max plans, while Team and Enterprise get it disabled by default with workspace‑level controls. Anthropic emphasizes that health, race, religion, and political or gender identity are not stored by default unless the user explicitly enables sensitive memory.

Sources: theaiinsider.tech, finance.sina.com.cn, aibase.com, note.com

Claude Code ships new Auto mode, Bash rule warnings, SendFeedback tool, and expanded compliance/developer APIs.

Release notes and coverage indicate that Claude Code v2.1.246 introduces Bash rule warnings and an “Auto mode” tab, improving how the coding assistant guides shell commands and workflow selection for developers. A follow‑on v2.1.247 adds a SendFeedback tool, a `/claude-api cost-optimize` command, and ~20 bug fixes, alongside new Organization API support in Anthropic’s Python and TypeScript SDKs. In parallel, Compliance API session endpoints for Cowork and Claude Code move out of beta, with local session transcripts expanded to Claude Science and Microsoft 365 (Excel, PowerPoint, Word, Outlook) sessions for Enterprise customers, plus new Admin API support in the CLI and SDKs.

Sources: media.patentllm.org, releasebot.io, claudenews.com.br

Anthropic adds a secure Python code‑execution tool to Claude’s API toolkit.

Anthropic’s release notes describe a new API tool that allows Claude to execute Python code in a secure, sandboxed environment, enabling programmatic computations, data transformations, and lightweight simulations from within conversations or workflows. This is positioned as part of the broader Claude tools ecosystem, providing safer code execution without exposing full system access, useful for data analysis, prototyping, and agent behaviors in regulated environments.

Sources: support.claude.com

Anthropic highlights Claude’s ability to design novel protein binders and analyze chemistry lab data.

Anthropic is showcasing Claude’s advanced scientific capabilities, including its use in designing novel protein binders from scratch, analyzing raw laboratory chemistry data, and supporting protein design pipelines. According to reports, Claude has generated candidate protein binders that were subsequently validated in a lab setting, positioning the model for specialized R&D applications.

Sources: blog.jiufeng.ai, indiatoday.in

OpenAI pauses reinforcement-learning training for latest deployment-bound models for two weeks to strengthen security.

OpenAI paused reinforcement learning (RL) training on its latest models intended for deployment for two weeks. This follows an internal assessment that a new experimental model reached critical cybersecurity capability thresholds under its Preparedness Framework. During the pause, OpenAI increased security monitoring and expanded red-teaming to mitigate high-risk cyberattack vectors before resuming large-scale training.

Sources: communeify.com, blog.jiufeng.ai, buttondown.com

Anthropic extends 50% weekly usage-limit boost for Claude Code through August 31.

Anthropic announced that the temporary 50% increase in Claude Code weekly usage limits for paid users (Pro, Max, Team, and legacy Enterprise plans), originally set to end on August 19, is now extended through 2026-08-31. The extension applies across all Claude Code interfaces, while other limits remain unchanged. Anthropic is considering making the quota increase permanent.

Sources: claudenews.com.br, news.aibase.com, ainews.imrenagi.com, x.com

Claude rolls out direct Gmail and Google Drive integration and expands “Claude Cowork” to mobile and web.

Claude now integrates directly with Gmail and Google Drive, allowing users to retrieve emails, draft replies, and manage Drive files from the chat interface. Anthropic also expanded its “Claude Cowork” autonomous task feature to mobile and web, enabling persistent agentic context between desktop and mobile devices.

Sources: communeify.com

Claude Code v2.1.235 adds spellcheck and bug fixes for prompt cache and permission dialogs.

Release notes for Claude Code v2.1.235 list a new built-in spellcheck capability in the editor and fixes for issues affecting prompt cache and permission dialog behavior. The update also includes new Agent SDK versions (TypeScript v0.3.235 and Python v0.2.140) with support for an updated MCP 2.x tool protocol for improved multi-tool orchestration.

Sources: claudenews.com.br

xAI’s Grok 4.6 and OpenAI’s previewed GPT‑5.6 Sol Ultrafast mode intensify competition for long-running agentic AI.

xAI released Grok 4.6 for long-running tasks, with performance on key evaluation indices matching OpenAI’s GPT‑5.6 Sol Max. At the same time, OpenAI previewed an “Ultrafast” mode for GPT‑5.6 Sol, reportedly capable of generating around 750 tokens per second, aimed at agentic scenarios.

Sources: distillintelligence.com

Anthropic makes Claude Code Auto Mode the default for Pro, Max, and Team subscribers starting August 14, 2026.

Anthropic is changing Claude Code’s behavior such that Auto Mode becomes the default across paid plans, replacing manual approval flows with an AI classifier that decides when to execute code. This shift targets smoother coding-agent workflows.

Sources: aibriefing.dev

Google releases Gemini 3.7 Flash, an AI model optimized for coding and automated business agent workflows.

Google unveiled Gemini 3.7 Flash as its latest AI model tailored for software coding and agentic business tasks, positioning it as an “agentic workhorse.” The release targets the enterprise coding and workflow automation market.

Sources: originbrief.app

Anthropic will watermark text outputs from Claude and other models to comply with the EU AI Act.

Anthropic committed to watermarking text generated by Claude and its other models in a machine-readable way, in order to meet the European Union AI Act’s transparency code, which took effect on 2026-08-02. The policy requires AI-generated content to be marked so other systems can detect its origin.

Sources: aipolicydaily.org

U.S. GSA's low-cost access deals for ChatGPT, Gemini, and Claude, priced around $1 per agency, are set to expire.

Three General Services Administration agreements that gave federal agencies access to OpenAI’s ChatGPT, Google’s Gemini, and Anthropic’s Claude for as little as $0.47 to $1 per agency are scheduled to expire on September 30, 2026. The GSA has not indicated if they will be renewed.

Sources: aipolicydaily.org

Anthropic disables Fable 5 and Mythos 5 globally after a U.S. export control directive and changes enterprise data retention to 30 days.

A U.S. government export-control directive forced Anthropic to globally disable frontier models Fable 5 and Mythos 5. At the time of suspension, Fable 5 was priced at $10 per million input tokens and $50 per million output tokens. In parallel, Anthropic announced a 30-day usage data retention policy across enterprise plans, while stating that the data would not be used for training.

Sources: torontostarts.com

Meta launches Muse Glimmer and commits to opening weights for Muse Spark 1.2 to support agentic tasks on consumer hardware.

Meta announced Muse Glimmer, a smaller AI model tuned for agentic tasks (coding, automation) on PCs or Macs with a single GPU. CEO Mark Zuckerberg also highlighted that the company would open the weights for Muse Spark 1.2, allowing the public to download and use the model directly.

Sources: cnbc.com

Mistral launches Shieldstral 1.0, a 3B-parameter open-weights multimodal safety classifier.

Shieldstral 1.0 is a ~3B-parameter safety model built on Ministral‑3‑3B‑Base with a native Pixtral vision encoder, released under Apache 2.0 open-weights licensing. It reaches 84.9% average F1 on text safety and 83.8% on multimodal safety, on par with larger models, while running on a single 16GB GPU. The system treats moderation as a policy-adaptive question instead of a fixed harm-category classifier.

Sources: aibriefing.dev

AMD promotes “Advancing AI 2026” event highlighting next-gen AI compute

AMD’s site and promotional materials for Advancing AI 2026 (July 22–23 in San Francisco) emphasize upcoming disclosures on AI chips used for cloud and enterprise workloads powering tools such as Copilot-like assistants and generative media services.

Sources: amd.com

Zapier expands access to its AI Agents beta and adds new app-action capabilities

Zapier announced another expansion of its Zapier Agents beta to more paying users, alongside support for additional app integrations and multi-step autonomous workflows. The update focuses on moving agents from simple task execution toward more persistent, multi-app “project runners” while keeping them constrained inside Zapier’s guardrails.

Sources: amd.com

Claude Code introduces improved workspace organization and multi-repo project handling

Anthropic pushed a Claude Code update focused on developer workflow: support for better project/workspace organization, smoother handling of multi-repository codebases, and improved file-tree navigation in large projects. The release emphasizes making Claude Code more usable for team-based development environments rather than just single-repo prototypes.

Sources: youtube.com

GitHub updates Copilot documentation with clarified data-handling and safety commitments for enterprise users

GitHub published an update to Copilot’s enterprise documentation clarifying how customer code is handled, expanding language on non-training of private repos, and outlining new configurable policies for suggestion filtering in regulated environments. The changes are framed as a response to large customers’ compliance and governance needs.

Sources: instagram.com

US government signals tighter oversight on major foundation model releases

A June recap from industry analysts highlights new US government actions: an executive order aimed at balancing AI innovation with stronger controls, and informal pressure on leading labs like OpenAI to moderate the pace of public releases. This has direct implications for release timing and safety policies for products such as ChatGPT, Claude, Microsoft 365 Copilot, and related enterprise offerings.

Sources: youtube.com

L’Oréal and OpenAI announce strategic collaboration to bring beauty experiences into ChatGPT

L’Oréal and OpenAI have partnered to build AI-powered consumer journeys and commerce experiences inside ChatGPT. Maybelline New York will be one of the first brands to launch, offering a Makeup Virtual Try-On directly in the chat interface, powered by L’Oréal’s ModiFace technology.

Sources: loreal-finance.com

Connecticut Enacts Comprehensive SB 5 AI Regulation Law

Governor Ned Lamont signed Senate Bill 5, establishing a comprehensive regulatory framework for AI systems in Connecticut. The law mandates that AI subscription providers disclose key terms upfront, requires safety protocols for AI companions, and obligates developers to ensure that synthetic audio, image, and video content is detectably marked as AI-generated. Additionally, it sets transparency standards for AI-driven employment decision processes.

Action: Enacted Disclosure and Watermarking Requirements

Sources: hklaw.com

ChatGPT Plus Reaffirmed as Lead Value Option for Writing Workflows

Recent evaluation confirms ChatGPT Plus at $20/month remains a top-value choice for individual creators compared to specialized tools like Jasper and Copy.ai. The platform is noted for its flexibility in handling blog posts, code documentation, and creative writing through conversational iteration rather than fixed templates.

Action: Reconfirmed existing $20/month pricing and GPT-4o capabilities in market comparison.

Sources: gadgethumans.com

ChatGPT Plus Ranked Best ROI Writing Tool in 2026 Comparison

Independent testing by GadgetHumans validates ChatGPT Plus as the superior value option for creators compared to Jasper and Copy.ai. The report highlights ChatGPT's $20/month pricing and GPT-4o's flexibility in blog writing, coding, and creative workflows as key advantages over Jasper's $49/month Creator plan.

Sources: gadgethumans.com

Claude Takes Lead in 2026 AI Benchmarks

New evaluations highlight Claude's superior performance in reasoning and multimodal tasks, surpassing competitors amid rapid open-source advancements. This reinforces Claude's position as a top enterprise-grade model, particularly for technical use cases like coding assistants.

Sources: youtube.com

Snyk Integrates Claude for Enhanced Security Scanning

Cybersecurity platform Snyk has integrated Anthropic’s Claude model to improve its security scanning capabilities. By leveraging Claude’s reasoning strengths, Snyk aims to provide more accurate vulnerability detection and automated remediation within software supply chains.

Sources: sdtimes.com

Opsera and Cursor Partner for Enterprise DevOps Integration

Opsera has announced a partnership with Cursor to integrate the AI-native code editor into enterprise DevOps pipelines. This collaboration enables automated code generation and testing within professional workflows, streamlining complex development tasks for engineering teams.

Sources: sdtimes.com

IMF Flags Claude Mythos Preview for Financial Cyber Threats

The International Monetary Fund (IMF) has flagged Anthropic’s Claude Mythos Preview as a significant risk to financial stability. The organization warned that the advanced model could be leveraged for sophisticated cyber attacks against financial systems and payments, necessitating stronger global regulatory supervision and AI-enhanced defense mechanisms.

Sources: techieray.substack.com

xAI Releases Grok 4.3 with Multi-Agent Enhancements

xAI has released Grok 4.3, introducing significant enhancements in multi-agent capabilities and benchmark performance. This version builds on previous Grok-4.20 betas to improve both reasoning and non-reasoning tasks.

Sources: llm-stats.com

OpenAI Launches GPT-5.5 Instant for Faster Inference

OpenAI launched GPT-5.5 Instant, a lightweight proprietary model designed for faster inference and reduced latency. It complements the existing GPT-5.5 ecosystem, providing a balance of speed and core capabilities for real-time applications in ChatGPT.

Sources: llm-stats.com

Jasper increases Creator plan price from $39/mo to $49/mo

Jasper's Creator plan increased by $10/mo (+26%) effective April 22, 2026. Existing monthly subscribers were notified 30 days in advance. Annual plan pricing remains at the previous rate until renewal. Teams plan pricing unchanged.

Action: Existing monthly subscribers: review billing. Annual subscribers: no immediate impact until renewal.

Semrush launches AI Overviews Tracker for all Guru and Business plans

Semrush rolled out AI Overviews position tracking to all Guru and Business subscribers. The feature monitors which target keywords trigger Google AI Overviews and whether your tracked pages appear within them — filling a significant visibility gap traditional rank tracking couldn't address.

Copy.ai Workflows adds 12 new pre-built automation templates

Copy.ai's Workflows feature (Pro and Team plans) added 12 new pre-built automation templates covering SEO blog pipelines, social media content calendars, email nurture sequences, and product description generators. Significantly reduces setup time for common marketing automation workflows.

Writesonic updates Article Writer to 6.0 with improved factual accuracy

Writesonic's Article Writer was updated to v6.0, primarily addressing factual accuracy issues from previous versions. Internal testing shows a 34% reduction in factual errors on informational content. Long-form output quality has improved noticeably — we're updating our review score accordingly.

Action: Re-testing recommended for existing users.

Surfer SEO raises Essential plan from $69/mo to $89/mo

Surfer SEO's Essential plan increased by $20/mo. Scale plan increased from $109/mo to $129/mo. New subscribers see the increase immediately; existing monthly plans see it next billing cycle. Annual renewals at previous rates for 90 days from announcement.

Ahrefs removes free Ahrefs Webmaster Tools crawl limits

Ahrefs removed crawl frequency limits from the free Ahrefs Webmaster Tools tier, allowing site owners to run full crawls more frequently without a paid subscription. A significant improvement to the free tier that makes technical SEO more accessible for smaller sites.

Rytr adds 'Magic Command' for free-form AI instruction

Rytr launched a Magic Command feature allowing users to give free-form instructions to the AI beyond the structured template system — bringing it closer to ChatGPT-style flexibility while maintaining Rytr's template-based workflow. Available on all paid plans.

OpenAI rolls out GPT-4o improvements to free tier users

OpenAI expanded access to improved GPT-4o capabilities for free tier ChatGPT users, including better instruction following and reduced output truncation on longer content requests. Significantly improves the free tier's viability as an AI writing tool for occasional users.

Clearscope raises Essentials plan from $170/mo to $189/mo

Clearscope's Essentials plan increased by $19/mo. Business plan increased from $350/mo to $399/mo. No changes to feature inclusions were announced alongside the price increase. Applies to new subscribers immediately and existing subscribers at next renewal.

Frase launches Answer Engine Optimization (AEO) scoring

Frase added an AEO scoring module that evaluates content against the patterns commonly seen in featured snippets and AI Overview responses. Provides actionable recommendations for structuring content to improve AI search visibility — making Frase one of the first content tools to explicitly address GEO optimization.

Sudowrite Story Engine adds Series Bible for multi-book projects

Sudowrite's Story Engine added a Series Bible feature for fiction writers working on multi-book projects or series. Maintains consistent character profiles, world-building rules, and plot threads across multiple manuscripts — addressing one of the most common limitations reported by prolific fiction writers.

Mangools KWFinder adds AI-powered keyword clustering

Mangools added automated keyword clustering to KWFinder, grouping keyword research results by semantic similarity and shared search intent automatically. Previously only available in enterprise platforms like Semrush and Ahrefs — its addition to Mangools at the $29/mo tier represents significant value for budget-conscious SEO professionals.