|
The Grind
The most unusual consensus in AI right now: Sam Altman, Elon Musk, and Dario Amodei all publicly agree that frontier AI development needs to slow down. Amodei published an essay titled "We Must Pace the Frontier" outlining a three-part plan, and both Altman and Musk posted their agreement on X within days. For solo builders, the immediate read is counterintuitive — a pacing agreement among lab CEOs is not a product announcement, but it's a signal about what the labs think the ceiling looks like and how long the current access environment holds. If the people building the models are publicly debating whether to ease off, the current window of cheap-and-capable is worth treating as finite.
The data on why that window matters is concrete. Across 230-plus models and 18 providers, DeepSeek V4.1 Flash runs at $0.15 per million input tokens and Llama 3.1 8B sits at $0.02 — the floor references for cost-sensitive workloads. Meanwhile the frontier clusters at $10/$50 (input/output). That 500× spread is the actual playing field. Google's April 2026 removal of Pro models from its free tier already narrowed the free lane, pushing API users onto Flash and Flash-Lite under tighter rate limits. Models with identical Intelligence Index scores can still differ by 57% in per-task pricing — so the headline rate continues to describe only one dimension of real spend.
OpenRouter's July discount experiment on GPT-5.6 Terra and Luna models crystallizes the dynamics. Token usage exploded 13.8× during the campaign — a clean demonstration of Jevons Paradox. Critically, nearly one-third of users retained some usage after the discounts ended, and daily volumes stayed elevated. That retention figure is the one to file: when price drops, new use cases get discovered, and some of them stick. Which is exactly why a pacing agreement among the people setting those prices deserves a line on your radar.
Also worth noting: GLM-5.3 from Zhipu AI launched on OpenRouter with a 1M-token context window and 131K max output at $0.075/M input tokens through September 9, rising to $0.15/M after — open-weight, publicly routable, and priced well below the workhorse tier. The model was previously OpenRouter's stealth "Ox Alpha" before the identity was confirmed.
More this week:
|
Solo founders tap AI tools for $1K–$3K monthly
Entrepreneurs are building side hustles using off-the-shelf AI agents—video creation, voiceover work, resume optimization—with Claude and Grokbot running ~$200/month, equivalent to 1990s corporate department capabilities. Real examples include SendRoq for LinkedIn automation and AI lesson planning.
Read on →
|
|
Musk backs Amodei's AI development slowdown plan
Elon Musk publicly endorsed Dario Amodei's 'We Must Pace the Frontier' essay proposing a three-part industry slowdown on frontier AI development, signaling potential industry-wide shifts.
Read on →
|
|
DeepSeek Flash undercuts rivals at $0.15 per million tokens
API pricing across 230+ models shows 57% spreads on identical performance tiers. DeepSeek V4.1 Flash hits $0.15/million input tokens while Llama 3.1 8B reaches $0.02; Google cut Pro models from Gemini's free tier on April 1, leaving only Flash and Flash-Lite with daily request limits.
Read on →
|
|
OpenAI launches Defense Factory for continuous cyber response
OpenAI's new continuous defense system aims to close the 'Defender's window'—the gap between open-weight model capabilities and active defenses—by automating vulnerability discovery, validation, and fixes faster than attackers leveraging public models.
Read on →
|
|
Astra generates full games in Excel and RPG engines
Ethan Mollick built a Space Invaders clone in VBA and an Ultima-style RPG called 'The Ninth Shore' using Astra (GPT-5.6), revealing the model's strength at interlocking system generation alongside persistent LLM weaknesses in narrative and voice.
Read on →
|
|
Anthropic enables centralized auth for enterprise MCP connectors
Enterprise-managed authentication for MCP connectors on Claude went GA in August, letting IT control all connector access through a single identity provider instead of forcing employees through per-connection OAuth prompts.
Read on →
|
|
OpenRouter opens public API routing across 50+ models
OpenRouter released public API access to route calls across 50+ text, image, and video models with OpenAI-compatible auth, automatic failover, and :nitro and :floor variants to tap service tiers for throughput and priority access.
Read on →
|
|
