Dateline: April 02, 2026 | Next update: April 09, 2026
Over the past week, Anthropic has focused on expanding agentic capabilities, developer tooling, and model infrastructure — capping a historically fast shipping month.
Claude / Anthropic
Claude Sonnet 4.6 — latest balanced model
Anthropic's newest Sonnet model brings improved agentic search and stronger efficiency across everyday tasks.
Launched as the default balanced model with a 1M token context window (beta) and improved agentic search that uses fewer tokens.
Context: 1M tokens (beta) | Output: up to 300k tokens via Batches API (beta header: output-300k-2026-03-24) | Extended thinking supported
Best for: Everyday tasks, agentic workflows, long-document processing
Claude Opus 4.6 — flagship model
The most capable released Claude model, featuring a 1M token context window, advanced coding, and a 14.5-hour task horizon.
Message Batches API max_tokens cap raised to 300k for Opus 4.6 — enabling long-form content and large code generation in batch jobs.
Context: 1M tokens | Batch output cap: 300k tokens | Top of Finance Agent benchmark
Best for: Deep analysis, long autonomous tasks, enterprise pipelines
Claude Haiku 4.5 — fast / low-cost model
Fast, cheap model for high-volume tasks. Note: Claude Haiku 3 (claude-3-haiku-20240307) is now deprecated and retires April 19, 2026.
Haiku 3 deprecation announced — migrate to Haiku 4.5 before April 19, 2026 to avoid service interruption.
Haiku 3 retirement: April 19, 2026 | Recommended migration: claude-haiku-4-5-20251001
Best for: Quick tasks, high-volume API calls, cost-sensitive workflows
Claude Code — agentic coding tool
Claude Code has shipped multiple updates this week, with new security controls, developer ergonomics, and Windows support.
PowerShell tool for Windows (opt-in preview). Transcript search (press / in transcript mode). Credential scrubbing from subprocess environments via CLAUDE_CODE_SUBPROCESS_ENV_SCRUB=1. Agents can now declare an initialPrompt in frontmatter to auto-submit a first turn.
New env vars: CLAUDE_CODE_SUBPROCESS_ENV_SCRUB, CLAUDE_STREAM_IDLE_TIMEOUT_MS | New hooks: TaskCreated, PermissionDenied | --bare flag for scripted -p calls
Best for: Developer coding workflows, CI pipelines, terminal-based automation
Cowork — desktop automation agent
Cowork received major upgrades this week: computer use is now live, letting Claude open apps, click, and navigate your Mac screen.
Computer use rolled out as a research preview for Pro and Max — Claude can use your computer on your behalf. Persistent agent thread now available from mobile and desktop to manage Cowork tasks from anywhere. Recurring and on-demand task scheduling added.
Computer use: Mac screen control + fallback to screen API | Mobile agent thread rolling out to Max first, then Pro
Best for: Autonomous task delegation, desktop workflow automation
Claude Developer Platform — API updates
Several API infrastructure updates shipped this week, focused on longer outputs and migration deadlines.
300k output cap on Message Batches API for Opus 4.6 and Sonnet 4.6. 1M context window beta retired for Sonnet 4.5 and Sonnet 4 on April 30, 2026 — migrate to Sonnet 4.6 or Opus 4.6. Web search and code execution now generally available (no beta header required).
Beta header for 300k output: output-300k-2026-03-24 | Context-1m beta for older models expires April 30 | Models API now returns max_input_tokens, max_tokens, and capabilities object
Best for: API developers, batch processing, long-output generation
Claude for Excel & PowerPoint — add-ins
Both add-ins now share full conversation context across applications, and support skills and LLM gateway connections.
Full cross-app context sharing between Excel and PowerPoint add-ins. Skills support added. LLM gateway support for Bedrock, Vertex AI, and Microsoft Foundry.
Gateway support: Amazon Bedrock, Google Vertex AI, Microsoft Foundry
Best for: Enterprise knowledge workers, cross-app AI workflows
On the Horizon — Claude Mythos
Anthropic confirmed it is testing Claude Mythos, described internally as "the most capable model we've ever developed," with early access customers. A leaked draft blog post also referenced a model tier called "Capybara" — positioned above Opus — with dramatically higher benchmark scores on coding, reasoning, and cybersecurity. No public release date has been announced.
Claude Mythos confirmed in testing with early access customers. Represents a described "step change" in performance per Anthropic.
No API access yet | Anthropic flagged unprecedented cybersecurity risk profile for Mythos | Early access only
Best for: Watch this space — no action needed yet
Plans and Pricing
No pricing changes this week.
Sonnet 4.6: $3/$15 per M tokens | Opus 4.6: $5/$25 per M tokens | Batch API pricing unchanged
ChatGPT / OpenAI
Dateline: April 02, 2026 | Next update: April 09, 2026
Over the past week, OpenAI has focused on improving reliability, agent behavior, and system integration.
GPT-5.3 Standard — default model
Improved stability and tool usage.
Improved tool selection and reduced hallucination.
Context ~128k | Output ~4k–8k | No benchmark change
Best for: General use
GPT-5.3 Pro — high-reasoning model
Improved multi-step reasoning.
Increased reasoning stability.
Context ~200k (est.) | Slower but more reliable
Best for: Deep analysis
GPT-5.3 Mini — fallback model
Used for fast responses.
More aggressive fallback routing.
Context ~64k | Weaker reasoning
Best for: Quick tasks
Agent Mode
Handles multi-step workflows.
Better ambiguity handling.
Cloud execution | limited integrations
Best for: Task delegation
Deep Research
Combines browsing and reasoning.
Improved source ranking.
High latency | variable sources
Best for: Research
Memory & Projects
Persistent context across chats.
Better relevance filtering.
Global + project memory
Best for: Ongoing workflows
Advanced Voice Mode
Improved conversational flow.
Better turn-taking.
Mobile-first | no offline mode
Best for: Voice interaction
Other Features & Pricing
No major updates to Canvas, Sora, Codex, Shopping, or Image Generation. No pricing changes this week.
Gemini
Date: April 2, 2026 | Next update: April 9, 2026
Google DeepMind continues to accelerate the rollout of the Gemini 3.1 architecture. This week's focus has been on expanding "Extended Thinking" capabilities to more tiers and finalizing the transition of experimental creative tools into the unified Flow production environment.
For general readers, Gemini is becoming more specialized. Instead of one AI for everything, you now have specific versions for high-speed tasks (Flash Lite) and deep, slow-reasoning projects (Pro with Extended Thinking).
Gemini 3.1 Pro — the flagship
Gemini 3.1 Pro remains the primary model for complex reasoning and "hard" problems like debugging code or synthesizing data. It features "Extended Thinking," which allows the model to pause and process logic before providing an answer.
Extended Thinking "High Effort" mode is now live for all Gemini Advanced subscribers, significantly improving performance on mathematical proofs and complex architectural coding tasks. (April 2, 2026)
Context Window: 1M to 2M tokens | Output Window: 65,536 tokens | Benchmarks: ARC-AGI-2: 77.1% (SOTA), GPQA Diamond: 94.3% | API Pricing: $2.00 per 1M input / $12.00 per 1M output (up to 200k context); rates double above 200k
Best for: Scientific & logic research, complex coding, synthesis tasks
Gemini 3.1 Flash Lite — the efficiency specialist
The newest and most affordable model in the family, designed for high-volume, low-latency tasks like text classification and simple Q&A.
Flash Lite has exited "Preview" and is now General Availability (GA) with a 99.9% uptime SLA for Enterprise users. (March 30, 2026)
Context Window: 1M tokens (32K via API Preview) | API Pricing: $0.25 per 1M input / $1.50 per 1M output
Best for: Enterprise production, high-volume API calls
NotebookLM
NotebookLM is a personalized AI research assistant that uses your specific documents as its only knowledge source. Audio Overviews let AI hosts discuss your documents in a podcast format. Cinematic Video Overviews transform notes into 4K narrated video presentations (Ultra Only).
Added "Source Anchoring" for Video Overviews — click a timestamp in the video to jump directly to the relevant page in the PDF. (March 28, 2026)
Source Limits: up to 50 sources per notebook; 500k words per source
Whisk and Flow transition
Google is consolidating its creative tools. Whisk (experimental image/animation) is being phased out in favor of Flow (professional unified creative studio).
Final reminder issued: all users must migrate assets from Whisk to Flow before the April 30, 2026 deadline or they will be permanently deleted. (March 31, 2026)
Lyria 3 — music generation
Lyria 3 creates high-fidelity, 3-minute music tracks with vocals and lyrics from text prompts.
Updated to support multi-language vocal performances including Serbian, Spanish, and Japanese. (March 25, 2026)
Plans and Pricing
| Plan | Price | Key Includes | Not Included |
|---|---|---|---|
| Basic (Free) | $0.00 | Gemini 3.1 Flash, 32k Context | Deep Research, Jules, 4K Video |
| AI Plus | $7.99/mo | 200GB Storage, 200 Credits | 3.1 Pro Model, High Jules limits |
| AI Pro | $19.99/mo | 2TB Storage, Gemini 3.1 Pro | Veo 3.1 Standard (4K) |
| AI Ultra | $249.99/mo | 30TB Storage, Gemini 3.1 Ultra | — |
This document is kept current as of April 2, 2026. For the most up-to-date figures, please request a re-run of this guide monthly.
Microsoft Copilot
Dateline: April 02, 2026 | Next update: April 09, 2026
Over the past week, Microsoft has focused on expanding education tools, improving research accuracy, and refining Copilot Chat access.
Copilot for Education — six AI-driven teacher tools
Lesson planning and classroom activities now benefit from AI support, including alignment with standards in 35 countries, fill-in-the-blank exercises, and matching activities with leaderboards.
Six AI-driven features for teachers shipped.
Standards alignment engine (35 countries) | "Add Supporting Examples" module | Fill-in-the-blank generator with analytics | Matching activity builder with gamification
Best for: Teachers and schools
Copilot Researcher — upgraded reports
Reports now draw from hundreds of sources with improved accuracy via multi-agent validation.
Multi-agent reliability checks added.
Multi-agent validation pipeline | Enhanced source reliability scoring
Best for: Analysts, enterprises
Copilot Cowork — shared drafting (soft launch)
Teams can co-create research and documents using Copilot with Anthropic integration.
Shared AI-driven drafting launched.
Multi-user drafting environment | Anthropic collaboration layer
Best for: Team research and drafting
Copilot Chat — video recap
Meeting summaries now include narrated highlight reels with relevant clips.
Video recap feature added to Copilot Chat.
Clipchamp integration | Audio recap expanded to more languages
Best for: Business teams
Copilot Notebooks — expanded export formats
Reports can now be exported as PowerPoint, PDF, infographic, or audio.
Expanded export formats.
Multi-format export engine | Enhanced layout options
Best for: Report generation
Dragon Copilot — pricing update
Dragon Copilot is now more affordable for partners worldwide.
Lower per-user license cost — reduced pricing applies globally.
Best for: CSP adoption
Copilot Chat — licensing changes (effective April 15, 2026)
Copilot Chat remains in Outlook but will be removed from Word, Excel, PowerPoint, and OneNote unless users have a paid Copilot license.
Refined access rules — Outlook retains full Copilot Chat access; other Office apps require a paid Copilot license.
Outlook grounding (inbox + calendar) | Restricted access in other apps
Best for: Licensed users
Plans and Pricing
No other changes this week. Plans unchanged | API token billing.