v2.1.154: Opus 4.8 por defecto y trigger ultracode para workflows dinámicos
Opus 4.8 es ahora el modelo por defecto con máximo esfuerzo. El keyword ultracode (antes llamado workflow) lanza workflows dinámicos que orquestan decenas a cientos de agentes en paralelo para tareas de gran escala. Fast mode corre a 2.5× la velocidad y es 3× más barato que antes.
v2.1.157: Plugins auto-cargados desde .claude/skills sin marketplace
Los plugins en directorios .claude/skills se cargan automáticamente sin marketplace. El nuevo comando claude plugin init <nombre> genera el andamiaje de un plugin en un solo paso.
v2.1.158: Auto mode disponible en Bedrock, Vertex y Foundry para Opus 4.7/4.8
Auto mode is now available on Amazon Bedrock, Google Vertex AI, and Azure Foundry for both Opus 4.7 and Opus 4.8. Enterprise teams using managed cloud providers can now leverage automatic model selection without switching to self-hosted setups.
v2.1.160: Confirmación antes de escribir en archivos de inicio de shell o configs de build
Claude Code now prompts before writing to shell startup files (e.g. .bashrc, .zshrc) and build-tool configuration files that grant code execution. Prevents silent, hard-to-detect environment modifications. Also fixed Windows/WSL clipboard issues and vim mode bugs.
v2.1.161: Métricas OTEL con atributos de recurso y herramientas paralelas más robustas
OTEL_RESOURCE_ATTRIBUTES values are now included as labels on metric datapoints. A failed Bash command no longer cancels other in-flight parallel tool calls. Includes fixes for login flow, telemetry edge cases, and Windows compatibility.
Archivado de sesiones y enlaces clicables con OSC 8 en el TUI
Sessions can now be archived from the TUI via /archive or from the CLI with codex archive / codex unarchive — archived sessions are protected from resume or fork until restored. TUI markdown now keeps web links clickable with OSC 8 metadata.
Codex Sites: crea y despliega webs, dashboards y apps directamente en el producto
Sites is now in preview in the Codex app — create, save, deploy, and inspect websites, dashboards, internal tools, web apps, and games hosted by OpenAI. Works alongside the chat workflow for full-stack prototyping without leaving the tool.
Computer Use en Windows y soporte de Amazon Bedrock como proveedor
Computer Use now operates on Windows desktop apps — Codex can see, click, and type in foreground windows. Additionally, Codex supports Amazon Bedrock as a model provider, using AWS-managed authentication, account controls, and consolidated billing.
Modelos OpenAI ahora disponibles en Amazon Bedrock vía Responses API
OpenAI models are now accessible in Amazon Bedrock through an OpenAI-compatible Responses API endpoint. Availability varies by AWS region. Enterprises can run OpenAI models with AWS-managed auth, VPC controls, and consolidated billing — without leaving their existing cloud infrastructure.
Facturación de sesiones Container cambia a por minuto (mínimo 5 min)
Container session billing replaced the flat 20-minute rate with per-minute billing and a 5-minute minimum. The per-minute cost is unchanged — shorter sessions will be significantly cheaper under the new model.
Depreciaciones API: objetos prompt reutilizables, plataforma Evals y Agent Builder
Deprecation notices issued for reusable prompt objects, the Evals platform, and Agent Builder. Additionally, GPT-4.5 retires from ChatGPT on June 27, 2026 — API users should migrate to GPT-5.x models before that date.
IntelliJ Gemini Code Assist 1.53.2: correcciones de bugs y mejoras menores
IntelliJ Gemini Code Assist 1.53.2 released June 3, 2026 with bug fixes and minor enhancements. This is the last scheduled maintenance release before the June 18 individual-tier end-of-service date.
⚠️ 18 jun: fin de servicio para usuarios individuales — migración a Antigravity
Gemini Code Assist IDE Extensions and Gemini CLI stop serving requests for individual, Google AI Pro, and Google AI Ultra tiers on June 18, 2026. Google is consolidating all tools into its multi-agent Antigravity platform. Standard and Enterprise tiers continue without interruption.
Gemini 3.5 Pro: GA esperada en junio con ventana de 2M tokens y Deep Think
Gemini 3.5 Pro is in limited Vertex preview after Sundar Pichai promised it “next month” at Google I/O 2026. Targets a 2M-token context window, Deep Think reasoning, and frontier multimodal — the use cases Gemini Ultra used to cover. No confirmed GA date yet.
Facturación por AI Credits activa: controles de presupuesto por usuario y nuevo tier Max
As of June 1, every Copilot interaction — completions, chat, and code review — consumes GitHub AI Credits. Code review also consumes GitHub Actions minutes. User-level budget controls are now GA for organizations and enterprises. The new Copilot Max tier with higher quotas is available for existing subscribers via plan upgrade.
App de Copilot en preview para todos: Canvases, voz, sesiones cloud y navegación agéntica
The Copilot app technical preview is now available to all Pro, Pro+, Business, and Enterprise customers on Windows, macOS, and Linux. Headline feature: Canvases — bidirectional surfaces where the agent updates and you edit, reorder, approve, or redirect on the same surface. Also ships on-device voice (STT), cloud sessions for remote agent work, cloud automations for scheduled tasks, and agentic browsing to verify changes in an integrated browser.
Lanzamiento de Claude Security: escaneo de código y sugerencias de parches de vulnerabilidades
Anthropic released Claude Security, a new product that uses frontier models to scan codebases and suggest patches for vulnerabilities. Launched alongside the Project Glasswing expansion — extending Anthropic's Mythos model to 150 more organizations in 15+ countries for automated software vulnerability detection.
Anthropic presenta confidencialmente S-1 ante la SEC para su IPO
Anthropic filed a draft S-1 registration statement with the SEC on June 1, 2026 — the formal first step toward going public. The filing follows the Claude Opus 4.8 launch and a reported $65B funding valuation. No IPO date has been announced. Anthropic joins OpenAI and xAI in signaling a 2026 liquidity event.
Línea de Tiempo — Lanzamientos de Modelos
| Fecha | Modelo | Proveedor | Tier | Tipo | Notas |
|---|---|---|---|---|---|
| 2026-06-01 | Qwen3.7-Plus | Alibaba / Qwen | A+ | Propietario | Agente multimodal GA. GUI grounding con ScreenSpot Pro 79.0. Workflows agénticos GUI+CLI híbridos. Precio: $2.50/$7.50 por 1M tokens. |
| 2026-05-28 | Claude Opus 4.8 | Anthropic | S+ | Propietario | Modelo GA más capaz. Contexto 1M tokens, 61.4 Intelligence Index (#1 global). Fast mode a 2.5× velocidad, 3× más barato que versiones anteriores. |
| 2026-05-19 | Gemini 3.5 Flash | S | Propietario | GA en Google I/O 2026. Inteligencia frontier a 4× la velocidad. 76.2% Terminal-Bench 2.1. $1.50/$9 por 1M tokens, contexto 1M. | |
| 2026-05-20 | Qwen3.7 Max | Alibaba / Qwen | A+ | Propietario | Flagship solo texto. Intelligence Index 56.6 (#1 en China). 1M tokens. 60.6% SWE-Bench Pro. Menor tasa de alucinaciones de la frontera: 22.9%. |
| 2026-04-20 | Kimi K2.6 | Moonshot AI | A+ | Abierto | 1T params (32B activos), open-weight. 58.6% SWE-Bench Pro. Swarm de agentes: 300 sub-agentes y 4.000 pasos coordinados. |
| 2026-04-24 | DeepSeek V4-Pro | DeepSeek | S | Abierto | MoE 1.6T parámetros, entrenado en 32T tokens, contexto 1M nativo. Lanzamiento más significativo desde R1. Variante V4-Flash para alto rendimiento. |
| 2026-06-30 | Gemini 3.5 Pro | S+ | Próximo | Prometido para junio 2026 en Google I/O. Objetivo: contexto 2M tokens, razonamiento Deep Think, multimodal frontier. Preview limitado en Vertex. Sin fecha GA confirmada. |
Tendencias Destacadas
La semana estuvo dominada por la expansión a la distribución en la nube: OpenAI en Amazon Bedrock vía Responses API, Auto mode de Claude Code en Bedrock/Vertex/Foundry, y Codex con soporte nativo de Bedrock como proveedor. La gobernanza se desplaza hacia los controles del proveedor de nube, y los equipos empresariales pueden ahora centralizar autenticación y facturación.
La facturación agéntica granular se está convirtiendo en el nuevo estándar del sector: GitHub Copilot completó su transición a AI Credits por interacción con controles de presupuesto por usuario, mientras OpenAI simplificó los contenedores a un modelo por minuto. El modelo de suscripción fija está siendo reemplazado por economía de consumo a medida que los agentes multiplican el volumen de llamadas.
Las interfaces bidireccionales agente-humano emergen como paradigma dominante: Canvases de Copilot, workflows ultracode de Claude Code y Qwen3.7-Plus con GUI grounding apuntan todos hacia el mismo patrón — agente y humano trabajando sobre la misma superficie de trabajo, con el humano dirigiendo y el agente ejecutando de forma visible y revisable.
v2.1.154: Opus 4.8 default and ultracode trigger for dynamic workflows
Opus 4.8 is now the default model at high effort. The ultracode keyword (renamed from workflow) launches dynamic workflows that orchestrate tens to hundreds of background agents for large-scale tasks. Fast mode runs at 2.5× speed and is 3× cheaper than previous versions.
v2.1.157: Plugins auto-loaded from .claude/skills without a marketplace
Plugins placed in .claude/skills directories are now auto-loaded without requiring a marketplace. The new claude plugin init <name> command scaffolds a new plugin in a single step.
v2.1.158: Auto mode now available on Bedrock, Vertex, and Foundry for Opus 4.7/4.8
Auto mode is now available on Amazon Bedrock, Google Vertex AI, and Azure Foundry for both Opus 4.7 and Opus 4.8. Enterprise teams using managed cloud providers can now leverage automatic model selection without switching to self-hosted setups.
v2.1.160: Prompt before writing to shell startup files or build-tool configs
Claude Code now prompts before writing to shell startup files (e.g. .bashrc, .zshrc) and build-tool configuration files that grant code execution. Prevents silent, hard-to-detect environment modifications. Also fixed Windows/WSL clipboard issues and vim mode bugs.
v2.1.161: OTEL metrics with resource attributes and more robust parallel tool execution
OTEL_RESOURCE_ATTRIBUTES values are now included as labels on metric datapoints. A failed Bash command no longer cancels other in-flight parallel tool calls. Includes fixes for login flow, telemetry edge cases, and Windows compatibility.
Session archiving and OSC 8 clickable links in TUI
Sessions can now be archived from the TUI via /archive or from the CLI with codex archive / codex unarchive — archived sessions are protected from resume or fork until restored. TUI markdown now keeps web links clickable with OSC 8 metadata.
Codex Sites: create and deploy websites, dashboards, and apps in-product
Sites is now in preview in the Codex app — create, save, deploy, and inspect websites, dashboards, internal tools, web apps, and games hosted by OpenAI. Works alongside the chat workflow for full-stack prototyping without leaving the tool.
Computer Use on Windows and Amazon Bedrock as model provider
Computer Use now operates on Windows desktop apps — Codex can see, click, and type in foreground windows. Additionally, Codex supports Amazon Bedrock as a model provider, using AWS-managed authentication, account controls, and consolidated billing.
OpenAI models now available on Amazon Bedrock via Responses API
OpenAI models are now accessible in Amazon Bedrock through an OpenAI-compatible Responses API endpoint. Availability varies by AWS region. Enterprises can run OpenAI models with AWS-managed auth, VPC controls, and consolidated billing — without leaving their existing cloud infrastructure.
Container session billing switches to per-minute model with 5-minute minimum
Container session billing replaced the flat 20-minute rate with per-minute billing and a 5-minute minimum. The per-minute cost is unchanged — shorter sessions will be significantly cheaper under the new model.
API deprecations: reusable prompt objects, Evals platform, and Agent Builder
Deprecation notices issued for reusable prompt objects, the Evals platform, and Agent Builder. Additionally, GPT-4.5 retires from ChatGPT on June 27, 2026 — API users should migrate to GPT-5.x models before that date.
IntelliJ Gemini Code Assist 1.53.2: bug fixes and minor product enhancements
IntelliJ Gemini Code Assist 1.53.2 released June 3, 2026 with bug fixes and minor enhancements. This is the last scheduled maintenance release before the June 18 individual-tier end-of-service date.
⚠️ Jun 18: end of service for individual users — migrate to Antigravity
Gemini Code Assist IDE Extensions and Gemini CLI stop serving requests for individual, Google AI Pro, and Google AI Ultra tiers on June 18, 2026. Google is consolidating all tools into its multi-agent Antigravity platform. Standard and Enterprise tiers continue without interruption.
Gemini 3.5 Pro: GA expected in June with 2M-token context and Deep Think reasoning
Gemini 3.5 Pro is in limited Vertex preview after Sundar Pichai promised it “next month” at Google I/O 2026. Targets a 2M-token context window, Deep Think reasoning, and frontier multimodal — the use cases Gemini Ultra used to cover. No confirmed GA date yet.
AI Credits billing live: per-user budget controls and new Copilot Max tier
As of June 1, every Copilot interaction — completions, chat, and code review — consumes GitHub AI Credits. Code review also consumes GitHub Actions minutes. User-level budget controls are now GA for organizations and enterprises. The new Copilot Max tier with higher quotas is available for existing subscribers via plan upgrade.
Copilot app preview expanded to all plans: Canvases, voice, cloud sessions, agentic browsing
The Copilot app technical preview is now available to all Pro, Pro+, Business, and Enterprise customers on Windows, macOS, and Linux. Headline feature: Canvases — bidirectional surfaces where the agent updates and you edit, reorder, approve, or redirect on the same surface. Also ships on-device voice (STT), cloud sessions for remote agent work, cloud automations for scheduled tasks, and agentic browsing to verify changes in an integrated browser.
Claude Security launch: codebase scanning and vulnerability patch suggestions
Anthropic released Claude Security, a new product that uses frontier models to scan codebases and suggest patches for vulnerabilities. Launched alongside the Project Glasswing expansion — extending Anthropic's Mythos model to 150 more organizations in 15+ countries for automated software vulnerability detection.
Anthropic confidentially submits draft S-1 to the SEC for IPO
Anthropic filed a draft S-1 registration statement with the SEC on June 1, 2026 — the formal first step toward going public. The filing follows the Claude Opus 4.8 launch and a reported $65B funding valuation. No IPO date has been announced. Anthropic joins OpenAI and xAI in signaling a 2026 liquidity event.
Model Launches Timeline
| Date | Model | Provider | Tier | Type | Notes |
|---|---|---|---|---|---|
| 2026-06-01 | Qwen3.7-Plus | Alibaba / Qwen | A+ | Proprietary | GA multimodal GUI agent. ScreenSpot Pro 79.0, hybrid GUI+CLI agentic workflows. Pricing: $2.50/$7.50 per 1M tokens — half the cost of Opus 4.7. |
| 2026-05-28 | Claude Opus 4.8 | Anthropic | S+ | Proprietary | Most capable GA model. 1M-token context, 61.4 Intelligence Index (ranked #1 globally). Fast mode at 2.5× speed, 3× cheaper than previous models. |
| 2026-05-19 | Gemini 3.5 Flash | S | Proprietary | GA at Google I/O 2026. Frontier intelligence at 4× the speed. 76.2% Terminal-Bench 2.1. $1.50/$9 per 1M tokens, 1M context. | |
| 2026-05-20 | Qwen3.7 Max | Alibaba / Qwen | A+ | Proprietary | Text-only flagship. Intelligence Index 56.6 (#1 Chinese model). 1M tokens. 60.6% SWE-Bench Pro. Lowest hallucination rate in frontier: 22.9%. |
| 2026-04-20 | Kimi K2.6 | Moonshot AI | A+ | Open | 1T params (32B active), open-weight. 58.6% SWE-Bench Pro. Agent swarm scaling to 300 sub-agents and 4,000 coordinated steps. |
| 2026-04-24 | DeepSeek V4-Pro | DeepSeek | S | Open | 1.6T-parameter MoE, trained on 32T tokens, native 1M-token context. Most significant release since R1. V4-Flash variant for high-throughput workloads. |
| 2026-06-30 | Gemini 3.5 Pro | S+ | Upcoming | Promised for June 2026 at Google I/O. Targets 2M-token context, Deep Think reasoning, frontier multimodal. Currently in limited Vertex preview. No confirmed GA date. |
Notable Trends
The week was dominated by cloud distribution expansion: OpenAI on Amazon Bedrock via Responses API, Claude Code Auto mode on Bedrock/Vertex/Foundry, and Codex with native Bedrock provider support. Governance is shifting toward cloud-provider controls, allowing enterprise teams to centralize authentication and billing under existing AWS/Azure/GCP agreements.
Granular agentic billing is becoming the new industry standard: GitHub Copilot completed its move to per-interaction AI Credits with per-user budget controls, while OpenAI simplified container billing to per-minute. The flat subscription model is giving way to consumption economics as agents multiply the volume of interactions.
Bidirectional agent-human interfaces are emerging as the dominant paradigm: Copilot Canvases, Claude Code ultracode workflows, and Qwen3.7-Plus GUI grounding all converge on the same pattern — agent and human on the same work surface, with the human steering and the agent executing visibly and reviewably.
v2.1.154: Opus 4.8 par défaut et déclencheur ultracode pour les workflows dynamiques
Opus 4.8 est le modèle par défaut à effort élevé. Le mot-clé ultracode (renommé depuis workflow) lance des workflows dynamiques orchestrant des dizaines à centaines d'agents en arrière-plan. Fast mode tourne à 2.5× la vitesse, 3× moins cher.
v2.1.157: Plugins chargés automatiquement depuis .claude/skills sans marketplace
Les plugins dans .claude/skills sont chargés automatiquement sans marketplace. La nouvelle commande claude plugin init <nom> génère l'échafaudage d'un plugin en une seule étape.
v2.1.158: Mode Auto disponible sur Bedrock, Vertex et Foundry pour Opus 4.7/4.8
Auto mode is now available on Amazon Bedrock, Google Vertex AI, and Azure Foundry for both Opus 4.7 and Opus 4.8. Enterprise teams using managed cloud providers can now leverage automatic model selection without switching to self-hosted setups.
v2.1.160: Confirmation avant d'écrire dans les fichiers de démarrage shell ou configs de build
Claude Code now prompts before writing to shell startup files (e.g. .bashrc, .zshrc) and build-tool configuration files that grant code execution. Prevents silent, hard-to-detect environment modifications. Also fixed Windows/WSL clipboard issues and vim mode bugs.
v2.1.161: Métriques OTEL avec attributs de ressource et exécution parallèle plus robuste
OTEL_RESOURCE_ATTRIBUTES values are now included as labels on metric datapoints. A failed Bash command no longer cancels other in-flight parallel tool calls. Includes fixes for login flow, telemetry edge cases, and Windows compatibility.
Archivage des sessions et liens OSC 8 cliquables dans le TUI
Sessions can now be archived from the TUI via /archive or from the CLI with codex archive / codex unarchive — archived sessions are protected from resume or fork until restored. TUI markdown now keeps web links clickable with OSC 8 metadata.
Codex Sites: créez et déployez des sites web, dashboards et apps directement
Sites is now in preview in the Codex app — create, save, deploy, and inspect websites, dashboards, internal tools, web apps, and games hosted by OpenAI. Works alongside the chat workflow for full-stack prototyping without leaving the tool.
Computer Use sur Windows et Amazon Bedrock comme fournisseur de modèle
Computer Use now operates on Windows desktop apps — Codex can see, click, and type in foreground windows. Additionally, Codex supports Amazon Bedrock as a model provider, using AWS-managed authentication, account controls, and consolidated billing.
Modèles OpenAI désormais disponibles sur Amazon Bedrock via l'API Responses
OpenAI models are now accessible in Amazon Bedrock through an OpenAI-compatible Responses API endpoint. Availability varies by AWS region. Enterprises can run OpenAI models with AWS-managed auth, VPC controls, and consolidated billing — without leaving their existing cloud infrastructure.
Facturation des sessions Container passe au modèle à la minute (minimum 5 min)
Container session billing replaced the flat 20-minute rate with per-minute billing and a 5-minute minimum. The per-minute cost is unchanged — shorter sessions will be significantly cheaper under the new model.
Dépréciations API: objets prompt réutilisables, plateforme Evals et Agent Builder
Deprecation notices issued for reusable prompt objects, the Evals platform, and Agent Builder. Additionally, GPT-4.5 retires from ChatGPT on June 27, 2026 — API users should migrate to GPT-5.x models before that date.
IntelliJ Gemini Code Assist 1.53.2: corrections de bugs et améliorations mineures
IntelliJ Gemini Code Assist 1.53.2 released June 3, 2026 with bug fixes and minor enhancements. This is the last scheduled maintenance release before the June 18 individual-tier end-of-service date.
⚠️ 18 juin: fin de service pour les utilisateurs individuels — migration vers Antigravity
Gemini Code Assist IDE Extensions and Gemini CLI stop serving requests for individual, Google AI Pro, and Google AI Ultra tiers on June 18, 2026. Google is consolidating all tools into its multi-agent Antigravity platform. Standard and Enterprise tiers continue without interruption.
Gemini 3.5 Pro: GA attendue en juin avec contexte 2M tokens et raisonnement Deep Think
Gemini 3.5 Pro is in limited Vertex preview after Sundar Pichai promised it “next month” at Google I/O 2026. Targets a 2M-token context window, Deep Think reasoning, and frontier multimodal — the use cases Gemini Ultra used to cover. No confirmed GA date yet.
Facturation AI Credits active: contrôles de budget par utilisateur et nouveau tier Max
As of June 1, every Copilot interaction — completions, chat, and code review — consumes GitHub AI Credits. Code review also consumes GitHub Actions minutes. User-level budget controls are now GA for organizations and enterprises. The new Copilot Max tier with higher quotas is available for existing subscribers via plan upgrade.
App Copilot en preview pour tous: Canvases, voix, sessions cloud et navigation agentique
The Copilot app technical preview is now available to all Pro, Pro+, Business, and Enterprise customers on Windows, macOS, and Linux. Headline feature: Canvases — bidirectional surfaces where the agent updates and you edit, reorder, approve, or redirect on the same surface. Also ships on-device voice (STT), cloud sessions for remote agent work, cloud automations for scheduled tasks, and agentic browsing to verify changes in an integrated browser.
Lancement de Claude Security: analyse de code et suggestions de correctifs de vulnérabilités
Anthropic released Claude Security, a new product that uses frontier models to scan codebases and suggest patches for vulnerabilities. Launched alongside the Project Glasswing expansion — extending Anthropic's Mythos model to 150 more organizations in 15+ countries for automated software vulnerability detection.
Anthropic soumet confidentiellement son projet S-1 à la SEC pour son IPO
Anthropic filed a draft S-1 registration statement with the SEC on June 1, 2026 — the formal first step toward going public. The filing follows the Claude Opus 4.8 launch and a reported $65B funding valuation. No IPO date has been announced. Anthropic joins OpenAI and xAI in signaling a 2026 liquidity event.
Chronologie des Lancements de Modèles
| Date | Modèle | Fournisseur | Tier | Type | Notes |
|---|---|---|---|---|---|
| 2026-06-01 | Qwen3.7-Plus | Alibaba / Qwen | A+ | Propriétaire | Agent GUI multimodal GA. ScreenSpot Pro 79.0, workflows agentiques GUI+CLI hybrides. Prix: $2.50/$7.50 / 1M tokens. |
| 2026-05-28 | Claude Opus 4.8 | Anthropic | S+ | Propriétaire | Modèle GA le plus capable. Contexte 1M tokens, 61.4 Intelligence Index (#1 mondial). Fast mode à 2.5× la vitesse, 3× moins cher. |
| 2026-05-19 | Gemini 3.5 Flash | S | Propriétaire | GA à Google I/O 2026. Intelligence frontier à 4× la vitesse. 76.2% Terminal-Bench 2.1. $1.50/$9 / 1M tokens. | |
| 2026-05-20 | Qwen3.7 Max | Alibaba / Qwen | A+ | Propriétaire | Flagship texte uniquement. Intelligence Index 56.6 (#1 modèle chinois). 1M tokens. 60.6% SWE-Bench Pro. Taux d'hallucination le plus bas: 22.9%. |
| 2026-04-20 | Kimi K2.6 | Moonshot AI | A+ | Ouvert | 1T params (32B actifs), open-weight. 58.6% SWE-Bench Pro. Essaim d'agents: 300 sous-agents et 4 000 étapes coordonnées. |
| 2026-04-24 | DeepSeek V4-Pro | DeepSeek | S | Ouvert | MoE 1.6T paramètres, entraîné sur 32T tokens, contexte 1M natif. Sortie la plus significative depuis R1. Variante V4-Flash haute performance. |
| 2026-06-30 | Gemini 3.5 Pro | S+ | À venir | Promis pour juin 2026 à Google I/O. Cible: contexte 2M tokens, raisonnement Deep Think, multimodal frontier. Preview limité Vertex. Pas de date GA confirmée. |
Tendances Notables
La semaine a été dominée par l'expansion de la distribution cloud: OpenAI sur Amazon Bedrock, mode Auto de Claude Code sur Bedrock/Vertex/Foundry, et Codex avec support natif Bedrock. La gouvernance se déplace vers les contrôles des fournisseurs cloud, permettant aux équipes entreprise de centraliser auth et facturation.
La facturation agentique granulaire devient le nouveau standard: GitHub Copilot a finalisé sa transition vers des AI Credits par interaction avec contrôles de budget par utilisateur, pendant qu'OpenAI simplifie la facturation conteneur au modèle à la minute. L'abonnement fixe cède la place à l'économie de consommation.
Les interfaces bidirectionnelles agent-humain s'imposent comme paradigme dominant: Canvases de Copilot, workflows ultracode de Claude Code et GUI grounding de Qwen3.7-Plus convergent tous vers le même patron — agent et humain sur la même surface de travail, l'humain dirigeant et l'agent exécutant de façon visible.