Python News

The patch fixed conversion of String fields containing dates to DateTime when writing to the database, integration errors with 1C and Linux hosts, unified the authorization error message against credential guessing, and optimized event handling in the monitoring center.

Engineers from two partners have completed the vendor's official certification in operating and configuring DATAREON Platform, expanding the pool of certified implementers for the platform.

The RC from May 21, 2026 became a stable spec with no material changes: initialize/Mcp-Session-Id were removed, a unified subscriptions/listen flow replaced HTTP GET and resources/subscribe, server/discover is mandatory, Roots/Sampling/Logging and the HTTP+SSE transport are deprecated (with a 12-month deprecation window). Multi Round-Trip Requests was added (resultType: input_required) to replace server-initiated requests, and JSON-RPC error code allocation policy was defined (-32020 to -32099 reserved for MCP).

Anthropic, through Dario Amodei, denied allegations of lobbying for a ban on open-weights models: the company says it never supported a ban and is promoting three measures - chip export controls against China, countering industrial-scale distillation, and mandatory pre-release safety testing for any sufficiently powerful model regardless of weight openness. Anthropic itself does not publish weights for any model in its lineup.

Opus 5 leads Frontier-Bench and GDPval-AA, but trails Mythos 5 on cybersecurity tasks. Pricing is unchanged: $5/$25 per million input/output tokens, the same as Opus 4.8. It is the default in Claude Max and the top model in Claude Pro. The weights are closed, with access only through the API (claude-opus-5) and cloud.

A workhorse update (-17% tokens versus 3.5 Flash) plus a dedicated cybersecurity model, Flash Cyber. All models have closed weights and are available via API/cloud.

Last scheduled beta; RC1 due 2026-08-04, RC2 2026-09-01, final 3.15.0 due 2026-10-01.

The 8B model scored 78.5% on RTEB and 75.5% on MMTEB Retrieval - first place overall. The 1B variant reaches 72.4% on RTEB (27% fewer errors than the previous 1B version), and NVFP4 quantization preserves 99%+ of BF16 accuracy at 2x throughput. Weights and training recipes are open and available on Hugging Face.

LightOn-rerank-LW-2B (2B, a LoRA adapter on Qwen3.5) ranks text passages and document scans with a single model and a single scale - 62.66 NDCG@10 on ViDoRe V3 versus 59.18 for Qwen3-VL-Reranker-2B and 59.40 for jina-reranker-m0. The 4B variant (64.69) beats Qwen3-VL-Reranker-8B (64.23) with half as many parameters. Weights are open on Hugging Face.

The new Pulse app (coming soon to Saleor Cloud) shows revenue, orders, returns, customers, products, fulfillment, period trends, and anomaly alerts (spikes in cancellations, unpaid shipments). It closes the gap with the ready-made BI dashboards of other platforms.

Yandex was the first in CIS to receive ISO/IEC 42001 for the base YandexGPT model back in 2025; certification now covers the full Alice AI lineup (LLM, ART, VLM) - development processes and data protection passed an independent audit.

On July 14, 2026, .NET 10.0.10, 9.0.18, and 8.0.29 were released. The scheduled update closed 17 CVEs across all three supported branches (10, 9, 8). Details for each vulnerability are linked to MSRC guidance, without exploit descriptions in the post itself.

MuleSoft published the @salesforce/mulesoft-vibes-skills package, reusable instructions for Mule development that can be ported between MuleSoft Vibes, Claude Code, Cursor, Cline, and Codex.

The agent carries a goal through to a finished result - spreadsheets, slides, documents, web apps - by breaking the task into steps and working for hours without intervention. Codex is integrated with it in a single desktop app. Weights are closed; OpenAI cloud/API only.

A three-model lineup: Sol as the flagship, $5/$30 per million tokens input/output; Terra at GPT-5.5 level for half the price, $2.5/$15; Luna as the budget option, $1/$6. Ultra mode uses subagents for complex tasks, with claimed support for defensive security (threat modeling, patching, blue team). On 2026-07-30, Luna was cut by 80% and Terra by 20%. Weights are closed; access is only through API/ChatGPT.

An operations-focused release: Fastly CDN integration with targeted cache invalidation, dark mode and theming in Studio, and an SBOM for compliance visibility. The last release with feature backports is 2025.4; after that, new features will ship only in newer versions.

Akeneo replaced the conversational Ask Ziggy with Agentic Ziggy, a layer of coordinated specialized agents that carry out data modeling, schema mapping, and quality checks from natural-language commands, rather than only providing suggestions.

The patch release fixes three vulnerabilities through a Django upgrade (CVE-2026-48588 - private data leakage from the response cache with Set-Cookie, CVE-2026-53877 - heap buffer over-read in GDALRaster, CVE-2026-53878 - header injection via DomainNameValidator).

432B parameters (28B active, MoE) versus 702B in the previous Ultra - a hybrid of standard attention and linear attention (GatedDeltaNet): 4x less KV-cache per token, 2.14x more context in the same memory. Sber's first model to publish not only weights but also key training stages; claimed to outperform DeepSeek V4 Flash and V3.2 on coding.

Fable 5 and Mythos 5 (launched in June 2026) were suspended by the U.S. Department of Commerce after a jailbreak was discovered. Access was restored on 2026-07-01 with a new classifier blocking more than 99% of exploit repeats; flagged requests are routed to Opus 4.8. Anthropic, together with Amazon, Microsoft, and Google, proposed an industry jailbreak severity scale, CJS-0 through CJS-4 (Glasswing project).

Major release: the configuration/monitoring/development center UI has been redesigned; the process handling service and message store have been moved to core cluster services; the internal DatareonDB architecture has been reworked (indexes, compression, maintenance); the 1C connector now logs delete events in information registers and exchange resilience has been improved for unstable networks; a security audit has been completed.

Sonnet 5 replaced Sonnet 4.6 as the default for Free/Pro on 2026-07-01 and is close to Opus 4.8 on agentic tasks such as planning, tools, and browser/terminal use. Launch pricing is $2/$10 per million input/output tokens until 2026-08-31, then $3/$15. The weights are closed, with API/cloud access only.

Workflow- and node-level spans are exported via OTEL to observability systems; most features are also available in the free self-hosted community edition, while custom telemetry tags are enterprise only.

Microsoft confirmed that the .NET 8 LTS release (36 months of support from November 2023) and the .NET 9 STS release (24 months from November 2024) will both lose security patches and technical support on November 10, 2026. Migration to .NET 10 LTS, supported through November 2028, is recommended.

Four Tier 1 SDKs have received beta releases for the new spec: Python mcp v2.0.0b1, TypeScript v2 (split packages), Go v1.7.0-pre.1, and C# v2.0.0-preview.1. They implement a stateless architecture, Multi Round-Trip Requests, routing headers Mcp-Method/Mcp-Name, and standard JSON-RPC error codes. Migration to v2 is optional, and existing servers continue to work.

Existing MuleSoft integrations exposed as MCP servers directly from API Catalog — Agentforce discovers/uses them with no custom wiring, governed through Omni Gateway.

A patch release after 4.3.0. The main fix is a RocksDB native memory leak in Kafka Streams; other critical bugs have been closed.

Partner Programming Store implemented an integration solution on DATAREON Platform for a customer in the oil industry; the scope and system coverage were not disclosed in the announcement.

The model gained native GUI/browser control for agents; closed weights, available via API/cloud.

The HOMEPAGE_WIDGETS mount point lets apps render as widgets in a responsive grid on the main screen, not just full-screen tabs. App Bridge actions RefreshEntity and WidgetResize were added (auto-height iframe instead of a fixed 200px).

Anthropic launched Claude Tag (beta, Enterprise/Team plans), where Claude joins a Slack channel as a regular participant and anyone can assign a task with an @ mention, without a separate bot for the channel. The work model is an agent in the team’s shared channel, not a developer’s personal CLI tool.

Python 3.15.0 beta 3 is out - the feature freeze has passed, and stabilization is underway before release.

Auth bypass in python.org release management API let a request with admin username + arbitrary API key process as admin, able to alter release/file metadata on python.org/downloads. Existed since 2014; reported 2026-02-23, patched within 48h; PSRT audit found no evidence of exploitation.

A unified product page conversion framework built on MDM/PIM data (core inherited from Riversand): enriched content, reviews, real-time social signals. It targets both human shoppers and AI discovery agents (ChatGPT, marketplace assistants). The claimed figures (+39% add-to-cart, +108% conversion when reading reviews) are vendor-reported and not independently verified.

Zero-touch OAuth via Identity Assertion JWT Grant (ID-JAG): IdP-issued token exchanged for MCP server access token, no per-server consent screen. Adopted at launch by Anthropic, Microsoft, Okta; 7 MCP providers support it (Asana, Atlassian, Canva, Figma, Granola, Linear, Supabase).

The platform received an interface redesign, natural-language access within the tools already in use, and a separate hub for programmatic access, enabling management by autonomous agents alongside people.

Compatibility has been confirmed for the sales management suite with the built-in CRM v3.1 features and the CRM module for 1C:ERP.Holding Management.

Routine bugfix release: ~179 fixes since 3.14.5, ~240 fixes since 3.13.13; build/docs only.

A fix for the 1C subsystem (version 3.0.6.1) has been released to restore stability after platform patch release 3.2.0.5.

1C launched a training program for the mandatory transition to EPD from 01.09.2026; support is being implemented in sales and ERP configurations. A regulatory requirement in CIS.

The code of new and republished marketplace modules (2,000+ solutions) is now additionally checked by an AI analyzer, which classifies findings by CWE/OWASP Top 10 and CVSS; vulnerable code is blocked from publication until the author fixes it.

On May 27-28, 2026, Saby (formerly the SBIS brand) held an EDI industry forum with case studies on adapting businesses to the mandatory transition to electronic waybills from September 1, 2026.

25 KIPs, 600+ commits since 4.2.0. New: broker cordoning before removal from the cluster, partition size metrics, share group tuning, tiered storage improvements. Phase 1 of deprecating the consumer's classic rebalance protocol has begun.

RC was locked on 2026-05-21, and the spec final is 2026-07-28 (a 10-week validation window). The initialize/initialized handshake and the Mcp-Session-Id header were removed, so the server runs behind a standard round-robin load balancer without sticky sessions. Tasks were moved out of core into an optional extension. The new MCP Apps extension provides server-side UI in a sandboxed iframe with JSON-RPC. A formal deprecation policy was introduced (Active → Deprecated → Removed, minimum 12 months); Roots, Sampling, and Logging are marked for removal.

The first model in the "frontier intelligence with action" line, outperforming Gemini 3.1 Pro on coding and agent benchmarks at lower cost. Closed weights.

The 17M model beats 33M ms-marco-MiniLM-L12-v2 by +0.051 NDCG@10 (MTEB) with half as many parameters; the 150M model is the strongest mid-tier reranker up to 600M, outperforming the 596M Qwen3-Reranker-0.6B; the 1B model matches the quality of the 1.54B teacher model (a 0.0001 NDCG@10 difference) with 2.4x faster inference. Apache 2.0, open weights.

Validated GS1 data from Syndigo product MDM flows directly into Blue Yonder planning and execution workflows (space planning, network sync), removing manual data cleansing during onboarding. Riversand is not mentioned explicitly, but it is the same multi-domain MDM engine under the hood of Syndigo.

New Bitrix24 mode: the user describes the task in text, and an AI agent based on BitrixGPT 5.5 studies the documentation, writes the code, and assembles the app without a developer. It works only within the Bitrix24 cloud ecosystem, and model access is closed (API/subscription, weights are not published) - closed-weights tag.

500+ fixes; support for PHP 8.5 (PHP 8.2 has been deprecated); Symfony upgraded to 7.4 LTS; support for OpenSearch 3.x and Valkey 8.1/9 added. Laminas MVC replaced with native PHP, Zend_Cache with Symfony Cache, TinyMCE with HugeRTE, and OAuth moved to native PHP functions. Local payment methods added (BLIK, Pay Upon Invoice, ELO), Apple Pay expanded to Chrome/Firefox, Google Pay card vaulting added, and the clearCart mutation introduced in GraphQL.

3.14's incremental GC reverted to 3.13's generational collector after production reports of excessive memory strain; ~154 other bugfixes.

3.15 hits feature freeze at beta 1; final release scheduled 2026-10-01 per PEP 790.

Node 26 includes the Temporal API (Temporal.PlainDate, Temporal.ZonedDateTime) without flags as a replacement for the outdated mutable Date; V8 is updated to 14.6 (Chromium 146), and Undici to 8.0. Current release, entering LTS in October 2026.

Connectors have been added for Business, Finance, Marketing, CX, ITSM, Security, DevOps, IT, Productivity, and Media (QuickBooks GraphQL, Shippo, BoldSign, Adobe Experience Platform, and others).

The node gives the n8n agent an identity in a Microsoft tenant: an email for @mentions in Teams, access to SharePoint sites, and the ability to act as a team member in Microsoft 365 apps. Early preview, available only through the Frontier program.

GPT-5.5 Instant replaced GPT-5.2/5.1 as ChatGPT's default model. Weights closed, access only via OpenAI cloud.

Native workflow automation inside Elastic Security replaces a separate SOAR tool, GA for Enterprise. Agent Builder now includes Skills (on-demand instruction packages - 5 for SecOps, 5 for SRE), Attachments, Connectors, and Plugins. PromQL has been added as a source command in ES|QL (tech preview).

IdP role mapping is now configured inside n8n instead of in an external IdP config; instance bootstrap via environment variables has been added for automated deployment, along with expression-based grouping for complex structures.

The Talend 7.3 line reached EOL on 2024-11-30; paid extended support (Severity-1 production blockers only, with no fixes for third-party components such as ActiveMQ or Jetty) ends in December 2026 and cannot be extended. Talend 8 or Qlik Talend Cloud is the only supported migration path.

Odoo has received Plateforme Agréée status from the French tax authorities, so receiving and sending electronic invoices and VAT e-reporting go directly through the product. From September 2026, receiving e-invoices will be mandatory for all companies in France, and issuing them as well for medium and large businesses.

Akeneo embedded the context-aware AI assistant Ask Ziggy into the interface, with guidance for configuration and workflows, and added Custom Channels for template-based retailer onboarding without manual engineering for each channel.

Second consecutive Magic Quadrant inclusion (after DXP in 2025); in MDM, it is a Niche Player, not a Leader. Eighteen vendors were evaluated.

The model ships with automatic filters that block forbidden high-risk cyber requests; it is available natively on Claude Platform and through Bedrock, Vertex AI, and Microsoft Foundry. Self-hosting is unavailable, and gateway routing must keep sessions in the provider's cloud.

DetailPage analyzes digital shelf signals and AI search visibility (Amazon, Walmart, ChatGPT), and attribute/content recommendations are fed back into the managed Syndigo PIM/MDM workflow for syndication and ROI reporting.

Business users (marketers, data stewards) assign tasks to agents in natural language instead of configuration; ready-made agents - Product Creation, Onboarding, Enrichment, Syndication - operate under human oversight. The AI model/provider is not disclosed (closed managed service, not downloadable weights).

Studio, with a React, TypeScript, and Ant Design interface, is production-ready after more than two years of development and is available in 2026.1 and 2025.4. 2026.1 is the first release without the legacy ExtJS admin UI, and all platform bundles now share a single version number.

The SDK is built into the platform core: agents receive identity, permissions, and an audit trail on par with people (the Agentic PXM concept). LLM weights are not published: the SDK orchestrates external models and is not itself a model; the announcement does not name specific providers or models.

A single validated source of truth above PIM/MDM and content platforms: quality rules, workflow, permissions, and audit trail apply equally to people and agents. The article does not disclose the licensing openness of Data Spine itself, only a mention of the platform's open-core Community Edition.

An agent inside Alice AI searches products across billions of offers online and completes the purchase directly in chat: average savings of 25% (about 6,000 RUB), with search taking 5-8 minutes. Available to 30.5 million monthly chat users; no specific agent model was named.

Meta Superintelligence Labs introduced Muse Spark, a multimodal reasoning model powering Meta AI (WhatsApp, Instagram, Messenger, Ray-Ban Meta). The weights are closed: access is limited to a private API preview for select partners, with no public release. This is Meta's first flagship release since Llama 4 Maverick (April 2025) that does not open the weights, marking the effective end of Llama as an open flagship.

A stable unified SDK for .NET and Python for orchestrating single-agent and multi-agent AI workflows (sequential, concurrent, handoff, group chat) with a stable API and an LTS commitment. Built-in connectors include Microsoft Foundry, Azure OpenAI, OpenAI, Anthropic Claude, Amazon Bedrock, Google Gemini, and Ollama; MCP support for calling external tools is already stable, and support for the A2A 1.0 protocol is listed as coming soon.

Ultra (702B MoE) and Lightning (10B, 1.8B active parameters) were moved from a dense architecture to MoE, and the post-training stage was redesigned (DPO in FP8: higher quality than bf16 with half the memory). The weights are open under MIT on HuggingFace and GitVerse; Ultra outperforms non-reasoning Qwen3-235B-A22B and DeepSeek-V3-0324 in math and general reasoning.

First agentic PXM platform according to Syndigo, natively built on enterprise multi-domain MDM (Riversand historical core). Agents reason and act across domains (product, supplier, location, compliance, performance) as unified "digital twin", sequentially or in parallel. Only "machine learning + generative AI" disclosed, specific model/vendor unnamed — closed service.

Through the Yandex Commerce Protocol (YCP), stores on 1C-Bitrix can show product cards and accept orders directly in Yandex Search and Alice, without going to the website; order and sales management remains with the seller.

The Talend ESB runtime 8.0.1 patch (Camel 4) fixes the camel-leveldb CVE and upgrades Spring Security to 6.3.9_2; JobServer has been updated to 8.0.2.20260224. There are no separate Talend ESB releases anymore, so fixes now ship through the shared cumulative Talend Studio/Runtime track.

With Node.js 27 (October 2026), the release model changes: Alpha 6 months -> Current 6 months -> LTS 30 months. All releases become LTS, and the odd/even split is removed. The reason is minimal adoption of odd-numbered releases over 10 years and the volunteer team being overloaded by supporting four branches at once.

Scheduled security and quality patch for the 2.4.8.x LTS branch (Adobe moved to a monthly patch schedule in January 2026). No new features, only security fixes and bug fixes for the current stable release.

The ModelContextProtocol SDK for .NET has reached v1.0 with full support for the MCP 2025-11-25 specification: Protected Resource Metadata for server authorization, icons for tools/resources/prompts, incremental scope consent based on least privilege, URL-mode elicitation for sensitive data, tool calls inside sampling requests, and experimental Tasks for long-running operations.

Fixes CVE-2024-6923, CVE-2026-24515, CVE-2026-25210, CVE-2025-59375: email header injection via unsafe folding/delimiters, XML memory-amplification DoS (libexpat bumped to 2.7.4), SSL use-after-free on SSL_new() failure, plus use-after-free/OOB in concurrent list ops (3.11.15/3.10.20).

4,876 passages from the Judicial College of Victoria Criminal Charge Book plus 100 expert questions were added to the Massive Legal Embedding Benchmark (MLEB, October 2025) - evaluating not only retrieval, but end-to-end answer generation.

Paper is a production-ready reference storefront, positioned as a foundation that both developers and AI agents can extend. It closes a long-standing Saleor gap: headless commerce without an official default frontend.

Yandex AI Studio gave businesses a builder for reasoning agents based on DeepSeek-V3.2 (third-party, not YandexGPT) for multi-step tasks such as tender analysis, incident review, and code. Caching/tool tokens were added (4x cheaper), along with secure connectivity and no request logging in the enterprise environment.

35 issues fixed since 3.9.1, including security patches. KIP-1252 removes the behavior mismatch of AlterConfigPolicy between ZooKeeper and KRaft modes (an optional compatibility flag).

Anthropic added automatic codebase vulnerability scanning in Claude Code on the web, with ready-made targeted patches for human review (limited research preview, Enterprise/Team). In a pilot with the Government of Alberta, 466 million lines of code were scanned in 20 hours.

More than a 2x reasoning gain over Gemini 3 Pro (ARC-AGI-2: 77.1%), powering Gemini 3 Deep Think. Closed weights, preview in Vertex AI, Gemini Enterprise, AI Studio, and Gemini CLI.

38 KIPs. Share groups (native queues on top of topics, KIP-932) are production-ready; the new Kafka Streams rebalance protocol (KIP-848) is production-ready too. DLQ support has been added in Streams exception handlers (KIP-1034), along with Java 25 support.

Positioning from the agent entry point to the storefront, transactional core, and devtools as a single open stack for agentic commerce - keeping control of the commerce engine with the merchant, not a third-party agent channel.

On MRCR v2 (8-needle, 1M tokens), Opus 4.6 scored 76% vs 18.5% for Sonnet 4.5 — better at retrieving and finding facts in large document sets without explicit RAG index. Model via API/cloud only, weights not published.

Qlik has announced a mandatory move of the Talend Studio desktop IDE to Java 21 starting with version 2026-06; jobs and services themselves remain on Java 17 (Big Data Jobs on Java 8), so the requirement applies only to the development environment. Migration is available through Feature Manager, a new JDK 21 installer, or manual setup (Studio 2025-09+ required).

dense_vector with bfloat16 elements (2 bytes instead of 4) cuts vector index size in half with minimal recall loss. Elastic Agent Builder is now GA: native chat with Elasticsearch data and a builder for custom AI agents without an external RAG stack.

MuleSoft Agent Fabric received a curated catalog of third-party MCP servers from the Official MCP Registry with governance; GA for adding MCP servers by URL in January 2026.

Akeneo launched Digital Showroom for sales teams, integrated with Stripe Agentic Commerce Suite for product discovery and checkout by AI agents through a single connector, and positions the MCP Server as a layer for controlling AI model access to product data.

The final 8.5.1 release of the 8.5 line is out, with a new interface (light/dark theme, beta status within the release) and backward compatibility with 8.3. UNF 3.0.13, Retail 3.0.13, and Cashier Workstation 1.1.1 were updated at the same time.

Task runners are enabled by default: Code nodes run in an isolated environment, and access to env variables from Code nodes is blocked. Legacy options have been removed (in-memory binary data), new SQLite pooling delivers up to a 10x boost, and a Publish/Save flow has been added for safe production updates. Breaking changes apply, with a migration report before upgrading; 1.x is supported for another 3 months.

GPT-5.2 (Instant/Thinking/Pro) improved long-context, vision and tool calling; Thinking achieved 80% on SWE-bench Verified and 98.7% on Tau2-bench Telecom. Separate GPT-5.2-Codex — specialized for agentic coding. Weights closed, access only via API/ChatGPT.

The program on the Standoff Bug Bounty platform is open to independent researchers to find vulnerabilities in the cloud version of Bitrix24 for a reward, expanding the earlier closed security-audit partnerships.

Protocol anniversary: experimental Tasks (polling for long-running operations), the Extensions framework for optional modules outside the core spec, OAuth Client ID Metadata Documents for simplified client registration, URL-mode elicitation, sampling with tool-calling support (server-side agent loops). The registry has grown to ~2,000 servers (+407% since September), with 2,900+ contributors in Discord.

The flagship completed the 4.5 line after Sonnet 4.5 (Sep. 2025) and Haiku 4.5 (Oct. 2025): 200K-token context, output up to 64K, and an effort parameter that balances quality and cost (medium effort delivers Sonnet 4.5-level SWE-bench performance with 76% less output-token usage). Weights are closed - access only through Claude.ai/Code/API and the AWS Bedrock, GCP Vertex, and Azure Foundry clouds.

PHP 8.5 is out - the pipe operator |>, built-in URI extension (RFC 3986/WHATWG), clone() with syntax, and the #[\NoDiscard] attribute. Active support until 2027-12-31, security fixes until 2029-12-31.

The Gemini 3 flagship model was rolled out on day one to Search, the Gemini app, AI Studio, Vertex AI, Gemini CLI, and the Antigravity IDE. Weights are closed; access is only through Google API/cloud.

Main feature — extension members: extension of existing extension method syntax to properties, operators, and static members. Added field-backed properties via contextual keyword field, implicit conversions between array/Span/ReadOnlySpan, null-conditional assignment, partial events and constructors, custom compound-assignment operators.

Akeneo added a native MCP Server ("Bring Your Own LLM") for connecting external models to catalog data while preserving governance, extended Data Architect Agent to reference entities, asset families, and variants, and added AI extraction of attributes from video, PDF, and images in DAM.

On 11 November 2025, Microsoft released .NET 10 as an LTS release with three-year support until 10 November 2028. Key changes: JIT compiler improvements (inlining, method devirtualization), AVX10.2 support for modern Intel processors, extended post-quantum cryptography (PQC) with Windows CNG support, Aspire 13 for cloud-native development.

The release added 50 new connectors and embedded generative AI into the low-code integration layer (MuleSoft for Flow).

Node.js 24.11.0 entered Active LTS on October 28, 2025, with support through April 2028. AsyncLocalStorage now uses AsyncContextFrame by default, URLPattern is globally available, and npm 11 is included.

Yandex relaunched the Alice chat on the new Alice AI LLM model - a MoE architecture with hundreds of billions of parameters, trained with online RL and built-in multi-stage web search. The YandexGPT brand for the chat flagship has effectively been replaced by Alice AI; weights are closed.

Current release: the --allow-net flag gives explicit control over network access in the permission model; Web Storage (localStorage/sessionStorage) is enabled without a flag; V8 is updated to 14.1 with faster JSON.stringify and built-in Uint8Array base64/hex conversion.

At DevDay 2025, OpenAI released AgentKit — Agent Builder (visual canvas workflows, beta), ChatKit (chat widget, GA), Evals for Agents, and Connector Registry for tool integration. On 3 June 2026, OpenAI announced sunset of Agent Builder and Evals — both leaving platform by 30.11.2026.

RTEB (beta) combines open and private datasets across 20 languages and domains (law, medicine, code, finance) on NDCG@10 metric to prevent model overfitting to public benchmarks—prior MTEB inflated scores due to public dataset memorization. Used as primary benchmark for 2026 embedding model evaluation (Nemotron 3 Embed and others).

Native VS Code extension (beta) with inline diffs, a checkpoint system that auto-saves the code state before changes (/rewind), subagents for parallel work (backend and frontend at the same time), and background tasks for long-running processes like the dev server. The default model is Sonnet 4.5.

A family of safety classifiers (0.6B/4B/8B) for real-time moderation of LLM traffic: Gen (instruction-following) and Stream (token-level). Apache 2.0, open weights.

AI features are natively available in every app: "Ask AI search" turns a natural-language question into Odoo filters/groupings, AI agents perform actions on request (quotes, follow-up), and fields are filled automatically by AI. ChatGPT has been added as a provider: the model is cloud-based, there are no weights, and access is only through the OpenAI API.

registry.modelcontextprotocol.io is now open as the central community-driven MCP server registry, API freeze v0.1, with no data-retention guarantees before GA.

Share groups (queues on top of topics) have been moved from Early Access (4.0) to Preview, bringing them closer to production but not ready yet. The Streams Rebalance Protocol (KIP-1071) has been added in early access status.

Yandex B2B Tech opened access to YandexGPT 5.1 Pro through Yandex Cloud AI Studio: the share of good answers rose from 60% to 71%, hallucinations fell by half to 16%, and the price was cut threefold to 40 kopeks per 1,000 tokens. The model is available only via API, with closed weights.

An image editing model based on Qwen-Image (native text rendering); open weights, self-hosting possible.

OpenAI released two reasoning models (117B and 21B parameters, MoE, Apache 2.0 license)—first open weights from the company since 2019. gpt-oss-120b runs on single 80GB GPU, benchmarks near o4-mini, exceeds it on AIME 2024/2025 and HealthBench. Weights published: self-host in your own infrastructure without OpenAI cloud dependency.