AI Landscape Snapshot — Week 28

GPT-5.6 Goes Generally Available in Three Tiers

OpenAI made GPT-5.6 generally available on July 9 across the API, Codex, and ChatGPT. The family ships as three models: Sol ($5/$30 per 1M tokens), Terra ($2.50/$15), and Luna ($1/$6). All three share a 1.05M token context window and 128K max output.

Sol sets new state-of-the-art on Terminal-Bench 2.1 at 88.8% (91.9% in Ultra mode with parallel agents) and leads Agents' Last Exam at 53.6, beating Claude Fable 5 by 13.1 points according to OpenAI's reported benchmarks. Terra delivers GPT-5.5-class quality at half the price. Luna is the cheapest option but has a sharp cliff on long-context recall — MRCR drops to 41.3% versus Sol's 91.5%.

The practical guidance: default to Terra for production work, escalate to Sol when correctness matters more than cost, and avoid Luna for anything requiring long-context reasoning. GPT-5.4 retires July 23.

Grok 4.5 Launches as a Coding-First Model

xAI released Grok 4.5 on July 8, its first model built specifically for coding and agentic work. Built on the V9 architecture and trained with real Cursor developer session data (SpaceX acquired Cursor for ~$60B in June), the model targets developers at $2/$6 per 1M tokens — significantly undercutting Opus 4.8 and GPT-5.5.

Grok 4.5 ranks 4th on the Artificial Analysis Intelligence Index but has a reduced 500K context window (down from Grok 4.3's 1M). It is not available in the EU at launch, with mid-July targeted for EU access.

Anthropic: Revenue Leadership and Model Lineup

Fortune confirmed that Anthropic has overtaken OpenAI on revenue, reporting a $47B annualized run rate versus OpenAI's $25-33B. OpenAI still leads on users (1.1B MAU for ChatGPT).

The Anthropic model lineup as of this week: Claude Fable 5 ($10/$50, 1M context) sits at the top as the flagship. Claude Sonnet 5, released June 30, is now the default on Free and Pro plans at an introductory $2/$10 through August 31. Opus 4.8 ($5/$25) remains the mid-tier flagship. As of July 8, Fable 5 access requires usage credits on all subscription tiers.

Mistral Ships Leanstral 1.5 for Formal Verification

Mistral released Leanstral 1.5 on July 2 under the Apache 2.0 license. The 119B-parameter MoE model (6.5B active per token, 256K context) is purpose-built for Lean 4 formal verification. It saturates miniF2F at 100%, solves 587/672 PutnamBench problems, and found 5 previously unknown bugs across 57 open-source repositories. Free during beta through September 30.

At roughly $4 per PutnamBench problem, it is dramatically cheaper than proprietary alternatives. This is a meaningful step toward making formal software verification accessible beyond specialist teams.

Cloudflare Redefines AI Crawler Access

Cloudflare announced that starting September 15, training and agent crawlers will be blocked by default on ad-supported pages. Search crawlers remain allowed. The policy introduces a three-category system — Search, Agent, and Training — replacing the binary block/allow approach.

Cloudflare handles roughly 20% of web traffic, making this operationally significant. The company also launched a Pay Per Use marketplace where publishers can charge AI companies when their content appears in AI-generated answers. Teams building RAG pipelines or training datasets from web sources should plan for this shift.

Security: First Autonomous AI Ransomware Attack

Sysdig documented JADEPUFFER (July 4-6), the first end-to-end autonomous AI ransomware attack. AI agents handled target identification, intrusion, payload deployment, and self-debugging — fixing errors in 31 seconds — without human input at each tactical step. A human still set the initial objective, but the execution skill bottleneck has been eliminated.

Framework Updates

The Vercel AI SDK shipped v7.0.19 on July 9 with fingerprintTools and detectToolDrift — the first framework-native MCP tool-definition drift detection ("rug pull" prevention). AI SDK 7 (June 25) had already introduced HMAC-signed tool approvals and WorkflowAgent for durable workflows.

DeepSeek's legacy deepseek-chat and deepseek-reasoner aliases stop responding on July 24 — migration to V4 model IDs is mandatory.

What This Means for Practitioners

Model routing is now table stakes. GPT-5.6's three-tier system and Anthropic's Fable/Sonnet/Opus/Haiku lineup both reward intelligent request routing. Teams that default to the most expensive model are overpaying by 3-10x on routine work. Take the AISA assessment to benchmark where your skills stand.

MCP security requires active management. Tool-definition drift is a real attack vector. If you're connecting to remote MCP servers, adopt fingerprinting and drift detection now.

Web data access is tightening. Cloudflare's September 15 deadline means any pipeline relying on broad web crawling needs a plan. Review the AI skills rubric for guidance on building robust data pipelines.

Cost optimization is a skill, not a feature. With Grok at $2/$6, GPT-5.6 Luna at $1/$6, and Claude Sonnet 5 at $2/$10, the floor for capable AI has dropped substantially. The differentiating skill is knowing when to spend more. Explore developer roles to see how model selection fits into the broader AI competency framework.