Week 31 Briefing: AI Pricing Updates: July 27 - August 2, 2026

On July 30, OpenAI announced the second API price cut in two weeks for the GPT-5.6 family. GPT-5.6 Luna, the cheapest model in the lineup, saw an 80% reduction: dropping from $1.00/$6.00 to $0.20/$1.20 per million tokens (input/output). GPT-5.6 Terra, the mid-tier workhorse, dropped 20% from $2.50/$15.00 to $2.00/$12.00.

OpenAI also launched Sol Fast Mode, a 2.5x faster variant of GPT-5.6 Sol at the same $5.00/$30.00 pricing. The mode targets time-sensitive applications like chatbots and real-time agents where latency matters more than peak reasoning depth.

These cuts follow the July 18 launch of the full GPT-5.6 family. In under two weeks, GPT-5.6 Luna went from $1.00 to $0.20 per 1M input tokens: a 5x reduction from launch pricing. ChatGPT and Codex subscription prices remain unchanged, but Terra and Luna usage now consumes fewer API credits, giving subscribers more effective usage for the same price.

On July 29, xAI launched Grok Voice Think Fast 2.0, a new speech-to-speech voice API. The upgrade brings improved voice quality and lower latency, but at a price: $0.08 per audio minute, up 60% from Think Fast 1.0 at $0.05/min.

This is a meaningful increase for developers building voice agents, real-time translation, and conversational AI on xAI's platform. At $0.08/min, a voice agent handling 10,000 minutes per month costs $800/month in API fees alone, up from $500.

xAI also announced that unpinned Grok Voice apps using the old alias will be auto-switched to the new model on August 5. Developers with pinned 1.0 aliases can stay on the old pricing, but only temporarily.

Two major AI transparency regulations took effect on August 2, 2026. The EU AI Act now requires AI-generated content to be clearly labeled, and high-risk AI systems must meet stricter documentation and compliance standards. The California AI Transparency Act mandates C2PA-compliant provenance metadata on AI-generated images and videos, with fines of $5,000 per day for non-compliance.

For consumers, the immediate change is more "AI-generated" labels on content. For providers, this means compliance overhead: content labeling systems, watermarking infrastructure, and risk assessments. Small providers may restrict access rather than implement the required systems. Midjourney currently has no C2PA-compliant watermarking solution and is now directly exposed to California fines.

Expect these compliance costs to flow into subscriber pricing over the coming months as AI tools build the required infrastructure into their cost base.

On July 27, Moonshot AI released Kimi K3's 2.8 trillion parameter model weights on HuggingFace alongside a technical report. The full model is now available for free self-hosting, bypassing the paused subscription sign-ups that have been in effect since July 19.

API pricing remains unchanged at $3/$15 per 1M tokens (with cache-hit at $0.30). But for developers and enterprises with their own compute infrastructure, the open weights release effectively makes K3 free to run, excluding hardware costs. This follows the growing trend of frontier models releasing weights: DeepSeek V4, Kimi K3, and Meta's models all now have open-weight paths.

New user subscriptions are still paused, so self-hosting is currently the only way for new users to access K3. Existing API users are unaffected.

On July 27, Cursor introduced Cursor Start, a new India-specific tier priced at ₹649/month (~$7), GST included and billed in INR. This represents a 65% discount from the global Pro plan at $20/month.

The move comes ahead of Cursor's acquisition by SpaceX and targets India's large and growing developer market. The pricing is among the most aggressive regional discounts in AI tools: even GitHub Copilot's India pricing is closer to $10/month.

In a related change on July 31, Cursor also removed the dollar Cost column from the Usage page for self-serve plans, switching to a tokens-only display. This change affects billing transparency for users tracking their per-request spend.