If you’ve been keeping up with AI news the past few days, you’re probably feeling some fatigue. Model launch → benchmark controversy → token pricing concerns → community backlash all happened within a single week.
This post organizes the major AI updates from mid-April 2026 into Claude / OpenAI / Google / Cursor / GitHub Copilot / Others sections. Each item combines the official announcement with practical commentary, so use it as a reference when checking individual updates.
📌 Today’s 3-Line Summary
① Claude Opus 4.7 launch (4/16) — same price tag, but the tokenizer change pushes real-world cost up 1.0–1.35x. Community backlash is spreading.
② Claude Design / Cursor Canvases / Gemini 3.1 Flash TTS — previews and new features dropping across visual creation, agent artifacts, and voice synthesis.
③ Opus 4.7 GA in GitHub Copilot — default model upgraded on Pro+ plan, CLI now ships auto model selection in stable.
🟪 Claude / Anthropic
Claude Opus 4.7 launch — the week’s biggest story
Anthropic released Claude Opus 4.7 on April 16. The model ID is claude-opus-4-7, and pricing matches Opus 4.6 at $5 input / $25 output per million tokens.
Benchmark numbers are clearly up.
| Metric | Opus 4.6 | Opus 4.7 |
|---|---|---|
| CursorBench | 58% | 70% |
| Vision bench | 54.5% | 98.5% |
| Image input resolution | Previous limit | Long side 2,576px (~3.75MP) |
| Reasoning option | – | New xhigh tier |
Additional capabilities shipped alongside:
- Task budgets public beta
- New
/ultrareviewcommand in Claude Code - Simultaneous launch across Claude products + API + Amazon Bedrock + Google Cloud Vertex AI + Microsoft Foundry
⚠️ Heads up — same price, different tokenizer.
For identical input text, Opus 4.7 consumes 1.0–1.35x more tokens than Opus 4.6 — Anthropic’s own docs confirm it. Pro subscribers are reporting that quotas blow out within a handful of prompts, and Anthropic has bumped subscriber rate limits in response.
Opus 4.7 backlash — a new release isn’t always an upgrade
Honestly, reactions split from day one. Here’s the key takeaway: “new model = automatically better” isn’t holding up, and the evidence has been piling up all week.
- A Reddit post titled “Opus 4.7 is a serious regression, not an upgrade” hit 2,300 upvotes.
- An X post arguing “4.7 is not better than 4.6” got 14,000 likes.
- Reports of 1.5–3x perceived cost increases when token bloat combines with long agentic runs and heavy tool use.
- Hallucination reports: misspelling “strawberry,” silently changing school names and surnames in resumes.
On the other side, Y Combinator CEO Garry Tan publicly endorsed it, and plenty of users report clear improvements in Claude Code and long-context, high-effort tasks.
My take is simple. Keep 4.6 for casual chat and short summaries, and only route code reviews, long codebases, and agentic runs through 4.7. If it feels genuinely better, switch your default — there’s no rush.
📎 Sources
→ Anthropic — Claude Opus 4.7 announcement
→ Simon Willison — system prompt diff analysis
Claude Design — Anthropic Labs preview
Anthropic Labs unveiled Claude Design on April 17 — a visual creation tool. It’s a research preview for building designs, prototypes, slides, wireframes, and marketing assets through conversation.
Available to Claude Pro/Max/Team/Enterprise subscribers at claude.ai/design.
Core capabilities:
- Learns brand systems from your codebase and design files
- Inline comments and direct editing
- Export to Canva, PDF, PPTX, HTML
- Direct handoff to Claude Code
- Supports interactive prototypes, 3D, video, shaders, and voice
Powered by Opus 4.7. Even non-designers — PMs, founders, marketers — will probably gravitate to this for “design by talking.”
📎 Sources
→ Anthropic — Claude Design announcement
🟢 OpenAI / ChatGPT
GPT-5.4-Cyber + Trusted Access for Cyber expansion
On April 14, OpenAI released GPT-5.4-Cyber, a variant tuned for cyber defense. It’s trained to be “cyber-permissive” — meaning it refuses less on security research and defensive work.
At the same time, the Trusted Access for Cyber (TAC) program expanded to thousands of vetted individuals and hundreds of teams. Eligibility is restricted to defenders of critical software.
Not something average users can grab directly, but the pattern of “defense-focused model variants operated at a company level” will only become more common.
Codex Pro $100/month plan + GPT-5.3 Instant Mini
A new $100/month Pro plan for long, intensive Codex sessions has been added. Pricing positions it against Claude Code Max at $200/month.
The ChatGPT fallback model has been swapped to GPT-5.3 Instant Mini. OpenAI cites improvements in conversational naturalness, writing, and context awareness. Free/Go tier ad placements have started rolling out in Australia, New Zealand, and Canada. The US isn’t in scope yet.
📎 Sources
→ OpenAI — Trusted Access for Cyber expansion
🟦 Google / Gemini
Gemini 3.1 Flash TTS — directing voices with prompts
On April 15, the Gemini API added gemini-3.1-flash-tts-preview. What stands out is its prompt-based directing approach to TTS.
It goes beyond plain text-to-speech — you can control accent (London/Newcastle/Exeter, etc.), emotion, breathing, and pacing through prompts. Inline tags like [yawn] and [excitedly] support multi-speaker dialog generation.
From a practical standpoint, this opens a real alternative to ElevenLabs. If you already have a Gemini API account, you can experiment without adding billing — worth a look if you automate narration, podcasts, or short-form video dubbing.
Importing context from other AIs into Gemini
Gemini’s settings page now has a switching tool. It imports personal context, preferences, and full chat history from other AI apps into Gemini.
A clear strategy to lower switching costs from competitors. Whether ChatGPT memory actually transfers cleanly is something we’ll have to verify in testing.
📎 Sources
→ Simon Willison — Gemini 3.1 Flash TTS review
🟧 Cursor
Cursor Canvases — agent outputs as editable documents
On April 15, Canvases shipped in Cursor 3.1+. Agents can now produce interactive “canvases” as conversational responses.
First-party visuals supported include dashboards, tables, diagrams, and charts. They live as persistent artifacts in the Agents Window side panel — document-style outputs that sit alongside the terminal and source control.
Especially useful for work where the artifact shouldn’t scroll away in chat history: reports, plans, architecture diagrams.
Bugbot self-learning + MCP integration
Features that shipped on April 8 stabilized this week. The core piece is learned rules — Bugbot learns from PR feedback and reactions — and reviews can now connect to MCP servers.
Cursor’s published resolution rate is 78%, with improvements to batch autofix and autofix trigger conditions. Personal and team settings have been split apart cleanly. If your team is evaluating Bugbot for PR reviews, now’s the moment.
📎 Sources
→ Cursor Changelog — 04/08 (Bugbot)
🟦 GitHub Copilot
Claude Opus 4.7 GA on Copilot Pro+
In the Copilot Pro+ model picker, Opus 4.7 has replaced Opus 4.5/4.6. GitHub cites meaningful gains on multi-step and agentic tasks.
Caveat: Pro+ subscription required — regular Pro and Free users don’t get it. Also, through April 30, 2026 there’s a promo period applying a 7.5x premium request multiplier to Opus 4.7 calls. If you run heavy sessions, watch your monthly quota burn rate.
Copilot CLI auto model selection GA
From April 17, auto model selection in Copilot CLI is generally available across all plans. Copilot picks the most efficient model per task (Claude, Codex, etc.).
When resuming sessions, short session ID prefixes (7+ characters) are now accepted — much less typing to get back into a session.
US/EU data residency + FedRAMP compliance
Updates that remove enterprise adoption blockers landed at the same time. Inference and related data stay within the specified region, and FedRAMP compliance for government and regulated industries is supported.
Currently US and EU only — other regions aren’t in scope yet.
📎 Sources
→ GitHub — CLI auto model selection
→ GitHub — Data residency + FedRAMP
⚫ Others — xAI Grok 4.3 Beta, Qwen3.6 open source
xAI Grok 4.3 Beta
Grok 4.3 Beta rolled out to iOS, Android, and web simultaneously on April 17. Native generation of PDFs, PPTs, and spreadsheets is supported, along with new video input.
Catch: SuperGrok Heavy ($300/month) only. On regular SuperGrok, the option shows up but stays locked. The price-of-entry is steep, so most developers will probably want to sit this one out for now.
Qwen3.6-35B-A3B — the open-source model that beat Opus on a laptop
On April 16, Alibaba released a quantized build of Qwen3.6-35B-A3B. It’s 20.9GB and runs locally on a MacBook Pro M5 via LM Studio.
On Simon Willison’s “pelican on a bicycle” SVG benchmark, it produced a better result than Opus 4.7. A symbolic moment for local, open-source models outperforming a frontier model on a specific creative task — the case for evaluating local options on privacy- and cost-sensitive workloads just got stronger.
📎 Sources
→ Simon Willison — Qwen3.6 beats Opus 4.7
💬 My take
Honestly, Opus 4.7’s launch felt underwhelming in my own usage. The benchmark numbers definitely went up, but in identical conversations the token consumption is noticeably higher — Pro quota empties out fast. The criticism that “same price tag but 1.35x the tokens” is a real price hike is fair.
That said, for coding and agentic workflows I think switching to 4.7 is the right move. Long-codebase comprehension and refactoring suggestion quality clearly stepped up a tier. For everyday writing and short summaries, 4.6 is lighter — using both situationally is the realistic move for now.
The most interesting thread in this whole cycle is actually Qwen3.6 beating Opus on a specific benchmark. It’s a signal that the era of “pick whichever model is most expensive” is fading. The skill of choosing the right model for the job is going to matter more and more.
❓ FAQ
Q. Should I upgrade to Claude Opus 4.7 right now?
A. Not necessarily. Opus 4.6 is more efficient for everyday chat and short summaries. Run 4.7 on code reviews, long-context tasks, and agentic workflows first, confirm the actual difference, then decide whether to switch your default.
Q. Why does Opus 4.7 cost more than Opus 4.6? The price sheet looks identical.
A. The price sheet is the same, but the tokenizer changed. Identical input text now consumes 1.0–1.35x more tokens. Stack that on top of agentic loop length and tool usage, and the perceived cost can hit 1.5–3x. It’s documented in the official changelog.
Q. Can I use Opus 4.7 on GitHub Copilot Pro (regular/free)?
A. No — it’s Pro+ only. Regular Pro and Free users stay on the previous models. That said, Copilot CLI’s auto model selection is available across all plans.
🔗 All links
🔗 Official links for everything covered today
🟪 Claude / Anthropic
→ Claude Opus 4.7 announcement
→ Opus 4.6 → 4.7 system prompt diff
🟢 OpenAI / ChatGPT
→ Trusted Access for Cyber expansion
🟦 Google / Gemini
🟧 Cursor
🟦 GitHub Copilot
⚫ Others
This post is a curated AI update roundup summarizing and reorganizing publicly announced and verified sources as of April 20, 2026. For final pricing, features, and regional availability, please confirm on each company’s official page.
This post is part of an English mirror of a Korean dev experiments blog. Cross-posted for the global developer audience.