AI News of the Week: Sol Price Cut (Aug 23)

OpenAI cut GPT-5.6 Sol API pricing by over 20% on August 21, making its flagship model cheaper than Claude Opus 5 on both input and output tokens for the first time — a move Reuters attributed directly to growing competition from Anthropic and Chinese AI labs.

This week's AI news covers pricing shifts, a new teen-focused ChatGPT experience, a critical security patch for Microsoft Copilot, and the continued acceleration of open-weight model releases that are closing the gap with proprietary frontiers.

OpenAI Slashes GPT-5.6 Sol Pricing

OpenAI dropped GPT-5.6 Sol's API rates to $4 per million input tokens and $20 per million output tokens — down from $5 and $30 respectively. The promotional pricing is guaranteed through at least November 21, 2026. This is the first price cut to Sol since the GPT-5.6 family launched on July 9; OpenAI had previously cut only the lower tiers (Terra down 20%, Luna down 80%) on July 30.

The practical impact: Sol now undercuts Claude Opus 5's $5/$25 rate on both input and output. For output-heavy workloads, the savings are substantial — a 33% drop. The cut applies to API calls and eligible ChatGPT Work and Codex credits, while Pro, Plus, and Business subscription usage remains unchanged.

For practitioners tracking token economics, the pattern is clear: frontier model pricing is compressing fast. In six weeks, the entire GPT-5.6 family has seen price reductions. Combined with cached-input discounts (Sol's cache read is now $0.40/MTok), the effective cost of running frontier models is at a 2026 low.

ChatGPT for Teens and Replit Free Mode

OpenAI launched ChatGPT for Teens on August 18, a dedicated mode for users aged 13-17 that activates automatically based on account age signals. The teen version keeps full chatbot capability but adds meaningful guardrails: a Study Mode that guides students through problems rather than handing them answers, homework reminders that detect shortcutting, break reminders during extended sessions, and stronger content safeguards around self-harm, eating disorders, and violence.

Anti-Anthropomorphism Measures

Notably, ChatGPT is barred from using romantic language or terms of endearment with teen users and is explicitly instructed not to suggest it has feelings, consciousness, or emotions. OpenAI said these changes are designed to reduce emotional dependence on the chatbot — a responsible AI deployment consideration that other providers will likely need to address.

Replit Free Mode

Separately, Replit launched Free Mode on August 19, powered by GPT-5.6 Luna. Despite the name, it requires a Core ($20/month) or Pro ($100/month) subscription — it lets subscribers handle chats, ideation, and simple coding tasks without consuming their main AI budget. Harder tasks auto-escalate to stronger models. Replit credits the July 30 Luna price cut (80% reduction) with making the economics viable. With 50 million users and adoption across 85%+ of Fortune 500 companies, this is a significant distribution channel for OpenAI's budget tier.

AI Security: Microsoft Copilot CoSnitch Vulnerability

Microsoft patched CVE-2026-24301 on August 18 — a critical vulnerability in Copilot Personal nicknamed "CoSnitch" by Varonis Threat Labs, who discovered it. Rated CVSS 8.8, the flaw chained three weaknesses: an undocumented URL parameter that auto-executed prompts, Copilot's broad OAuth access to connected services (Gmail, Drive, Calendar), and a persistent memory-poisoning path.

A single crafted link could silently exfiltrate data from every connected account with no user confirmation beyond the initial click. This is the third Copilot vulnerability Varonis has uncovered in 2026, following Reprompt and SearchLeak — all sharing the same one-click attack pattern.

The discovery method is itself noteworthy: Varonis researchers found the undocumented autorun=1 parameter by repeatedly asking Copilot why it couldn't auto-run prompts. The AI helpfully explained its own hidden functionality — a pattern they call "meta-hacking" that highlights risks when AI systems with agentic workflows can introspect and disclose undocumented capabilities.

AI Model Leaderboard: Where Things Stand

According to the Artificial Analysis Intelligence Index (data verified August 22, 2026), the current top models are:

1. Claude Opus 5 (max effort) — Intelligence Index 63
2. Claude Fable 5 (max, with fallback) — 62
3. Grok 4.6 — 60.9
4. GPT-5.6 Sol (max) — 61
5. Kimi K3 — ~57 (strongest open-weight)

Pricing Snapshot

The fastest models are Celeris-1 (1,468 tokens/sec) and Mercury 2 (881 t/s). For the largest context windows, Llama 4 Scout and Grok 4.20 lead the field. Google's Gemini 3.5 Pro remains unshipped; the Flash line (now at 3.7) is Google's entire public release cadence.

Open-Weight Models Continue Closing the Gap

The open-weight ecosystem had a prolific August. Key releases in the past two weeks include Qwen3.8-Max from Alibaba (2.4T parameters — the largest open-weight release to date), GLM-5.3 from Z.AI (released August 14, benchmarks comparable to Kimi K3), and Meta's Muse Spark 1.2 with Muse Code for multi-agent orchestration workflows.

Kimi K3 from Moonshot AI remains the strongest open-weight model at 2.8T parameters with a 1M-token context window. On the Artificial Analysis Intelligence Index it scores 57 against Claude Opus 5's 63 — a gap that continues to narrow. For practitioners evaluating tool selection criteria, the cost advantage of self-hosted open models is increasingly compelling, especially with DeepSeek V4-Flash available at $0.14/$0.28 per MTok.

Qwen models have reached approximately 2 billion downloads on Hugging Face in 2026, surpassing open models from Google and Meta — reflecting rapid global adoption of Chinese open-weight models regardless of the geopolitical tensions discussed below.

EU AI Act Transparency Rules Now Enforced

The EU AI Act's Article 50 transparency obligations became enforceable on August 2, 2026, and the effects are now rippling through the industry. Every chatbot, voice assistant, and agentic AI system reaching EU users must now disclose that users are interacting with AI at the start of each interaction. Deepfakes must be labeled, and AI-generated content must carry machine-readable marks.

Fines reach up to €15 million or 3% of global annual turnover. Anthropic has responded by implementing machine-readable watermarks for Claude content in the EU, with new models required to support this marking from August 2. Over 180 companies including Anthropic, OpenAI, and Google have signed the EU's voluntary Code of Practice on Transparency.

For practitioners building AI-powered products, this is no longer a future concern. If your chatbot or agent serves EU users, Article 50 compliance is a current operating requirement. The high-risk AI system deadline has been extended to December 2027, but the transparency rules are live now.

Infrastructure: Cloudflare Kitesurf and Agent Payments

Cloudflare released Kitesurf in early August — a browser built from the ground up for AI agents, running on Cloudflare Workers without Chromium. It uses 3-7x less CPU and memory than headless Chromium for common agentic tasks like screenshots and HTML extraction, and supports existing Puppeteer, Playwright, and MCP clients.

The browser is built in Rust and WebAssembly, using Firefox's Stylo CSS parser and the Boa JavaScript engine. It passes over 215,000 Web Platform Tests and is available free during beta. Cloudflare also introduced the x402 protocol enabling AI agents to pay for services autonomously, with over 20 companies already participating.

For teams building browser-based agents, Kitesurf addresses a real bottleneck: running one Chromium instance per agent is expensive at scale. The efficiency gains matter when you're orchestrating dozens or hundreds of concurrent agent sessions.

What This Means for Your AI Skills

This week's developments touch several competencies that the AI skills rubric measures. The Sol price cut and the broader pricing compression across frontier models mean that model comparison and cost optimization are no longer niche skills — they're table stakes for anyone building with AI. Understanding promotional pricing windows, long-context surcharges, and cache economics directly affects project budgets.

The CoSnitch vulnerability demonstrates why AI security literacy matters. As AI assistants gain broader access to enterprise data, the attack surface expands in ways traditional security training doesn't cover. And the EU AI Act enforcement means that responsible AI deployment knowledge is now a compliance requirement, not just a best practice.

Aisa's conversational AI reads this snapshot every week to stay current on the AI landscape, so when you take the assessment, your conversation reflects the latest developments — not a frozen training set.

Take the free AI skills assessment to see where you stand.

---

Related reading: Top 10 AI Skills Certifications in 2026 — Which credentials actually matter this year.

Related reading: AI Fluency: The New Digital Literacy — Why understanding AI models is becoming as fundamental as reading.

Related reading: How Good Are Most People at AI? — Benchmark data on where professionals actually score.

Frequently Asked Questions

What are the biggest AI developments this week?

The biggest development is OpenAI cutting GPT-5.6 Sol API pricing by over 20%, making it cheaper than Claude Opus 5 for the first time. OpenAI also launched ChatGPT for Teens with parental controls and study-focused features, and Microsoft patched the critical CoSnitch vulnerability in Copilot Personal.

Which new AI models launched this week?

No brand-new frontier models launched this specific week (August 17-23), but OpenAI repriced GPT-5.6 Sol on August 21, and Gemini 3.7 Flash (released August 13) continued its rollout. The open-weight ecosystem saw GLM-5.2 Turbo from Z.AI on August 17. The Artificial Analysis leaderboard now tracks 169 ranked models with Claude Opus 5 at the top.

How do this week's AI changes affect professionals?

Pricing compression means AI practitioners need to regularly reassess which models they're using and whether cheaper options deliver equivalent results. The EU AI Act enforcement now requires anyone deploying chatbots or AI agents in the EU to implement transparency disclosures. Take the AISA assessment to evaluate how your skills align with these evolving requirements.