Anthropic and OpenAI Unleash a Product Wave: Opus 4.7, GPT-5.5 Instant, Claude Dreams, and a New Supercomputer Protocol

Anthropic and OpenAI Unleash a Product Wave: Opus 4.7, GPT-5.5 Instant, Claude Dreams, and a New Supercomputer Protocol

The AI industry saw one of its busiest 24-hour windows in recent memory. Between May 5 and May 6, Anthropic delivered a production-grade model upgrade, a surreal new "dreaming" capability, and 10 specialized finance agents — while OpenAI shipped a default-model refresh, co-authored a cross-industry supercomputer networking standard, and watched its former CTO testify in the Musk v. Altman trial.

Here's what happened and why it matters for developers, enterprises, and anyone building with AI.

Anthropic Claude Opus 4.7: The Coding Workhorse Gets Sharper

Anthropic released Claude Opus 4.7 as a generally available production model — and the reviews from early testers read like a laundry list of "this thing actually works."

Opus 4.7 is not Anthropic's most capable model overall (that title belongs to Mythos Preview, a cybersecurity-focused model held back for safety reasons). But Opus 4.7 is optimised for the kinds of tasks enterprises actually run: complex software engineering, multi-step agent workflows, long-context reasoning, and high-resolution image understanding.

Key benchmarks:

Testers across the board reported that Opus 4.7 catches its own logical mistakes, resists infinite loops (a notorious agent failure mode), and handles ambiguous instructions better than any previous Claude model. Cursor called it "a meaningful jump" at 70% on CursorBench versus 58% for Opus 4.6. Replit noted it achieves "the same quality at lower cost."

Pricing: Unchanged at $5/M input tokens and $25/M output tokens — available via API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry.

The catch: Opus 4.7 has deliberately reduced cybersecurity capabilities compared to Mythos Preview. Anthropic is treating it as a test bed for real-world safeties before considering a broader Mythos release. Security professionals can apply for the Cyber Verification Program to unlock limited cybersecurity use.

Claude Dreams: AI Agents That Sleep On It

In a move that sounds like science fiction, Anthropic launched Claude Dreams — a research preview feature that lets Claude review past agent sessions and reorganize its memory stores.

Dreaming works as an asynchronous job: Claude reads an existing memory store (plus up to 100 past session transcripts), then produces a cleaned-up version that merges duplicates, replaces stale entries, and surfaces new insights from patterns it finds across sessions.

For teams running autonomous agents — performing code review, data analysis, or document workflows — this is a practical leap. Instead of memory stores accumulating contradictions and dead weight, Dreams lets agents "sleep" and wake up with a curated, deduplicated memory. It mirrors how humans consolidate learning during rest.

10 AI Agents for Wall Street

Anthropic also rolled out 10 purpose-built AI agents for financial services — handling pitchbook drafting, financial statement review, credit memo preparation, and compliance escalation.

This is a clear signal: enterprise agents are moving past the "demo" phase. Finance is an ideal first market — document-heavy, regulated, and expensive enough to justify dedicated agent tooling. If these agents meet compliance standards, expect rapid expansion into legal, insurance, and healthcare.

OpenAI GPT-5.5 Instant: Default Model Refresh with Hallucination Reductions

OpenAI quietly replaced GPT-5.3 Instant with GPT-5.5 Instant as the default ChatGPT model on May 5.

This isn't a new model generation — it's the "Instant" variant of GPT-5.5 (which shipped last month), optimised for low latency and production reliability. The headline improvement: reduced hallucinations in high-stakes domains — law, medicine, and finance — where incorrect answers are costly.

Performance gains (GPT-5.5 over 5.3):

• AIME 2025 math: 81.2 (from 65.4)
• MMMU-Pro multimodal: 76 (from 69.2)

GPT-5.5 Instant also adds a search tool for referencing past conversations, files, and Gmail — making responses more context-aware. ChatGPT now displays memory sources, letting users delete or correct outdated information directly.

Rollout: Plus and Pro users on web first, mobile and free/enterprise tiers in the coming weeks. API access via chat-latest.

OpenAI + AMD + Broadcom + Intel + Microsoft + NVIDIA: A New Supercomputer Protocol

OpenAI partnered with AMD, Broadcom, Intel, Microsoft, and NVIDIA to publish MRC (Multipath Reliable Connection) — a supercomputer networking protocol released through the Open Compute Project.

Why this matters: as training clusters grow to Stargate scale, network congestion and link failures become the dominant bottleneck. MRC uses adaptive packet spraying and static source routing to virtually eliminate core congestion and survive link failures without crashing training jobs. One late data transfer can idle thousands of GPUs — MRC is designed to prevent exactly that.

This is a rare moment of cross-vendor cooperation in the AI chip stack, and it suggests that infrastructure bottlenecks — not just model architecture — are the next frontier for AI performance.

In Other News

• Samsung crossed $1 trillion market cap on AI memory demand — only the second Asian company after TSMC.
• xAI rebranded to SpaceXAI following SpaceX's acquisition of the company.
• Google DeepMind took a minority stake in Fenris Creations, the studio behind EVE Online.
• Hut 8 signed a ~$10 billion AI data center lease in Texas — a sign of where infrastructure money is flowing.
• China reportedly weighs investing in DeepSeek at a $50 billion valuation.

Why This All Matters

Two themes dominated this week:

1. AI agents are becoming real infrastructure. Claude Dreams, Opus 4.7's autonomous coding, and OpenAI's MRC protocol all point in the same direction: AI is moving from single-turn chat to persistent, autonomous, multi-session work. The tools to manage agent memory, reliability, and compute scaling are being built right now.
   
2. Safety is being productized. Anthropic's approach — release a less capable model with guardrails, test real-world safeguards, then unlock capabilities through verified programs — is a framework worth watching. It's a middle path between "move fast and break things" and total regulatory paralysis.