Inference Just Got 5x Cheaper - So the Fight Moved to the Agent Layer

The price/perf cliff

OpenAI's GPT 6.1 Sol is dominating HN's AI feed this week: "near-Astra intelligence for a fifth of the price." That's not an incremental discount - it's a price/perf cliff that resets API budgets. If it's real, near-frontier inference is commoditized again.

Anthropic is on the same path from the other side. Claude Haiku 5.5 landed Oct 7 and runs roughly 75% cheaper than Haiku 4.5. And FT reports that Anthropic's best model is struggling to attract users as cheaper tools thrive - the price war is moving customers, not just benchmarks.

It all lands on top of the GPT-6 rollout for all 1.2B weekly ChatGPT users, live since Oct 7 with a new interactive interface - OpenAI's biggest consumer rollout yet, per OpenAI. FT put annualized revenue near $50B at end of September, below the ~$70B rumored earlier.

Speed is collapsing on the same curve: Cerebras runs a GPT-5.6 Sol Ultrafast tier at up to 750 tok/s with no quality loss - a 7x latency cut that rewrites what interactive agent UX can be (HN).

The fight moved to the agent layer

When intelligence gets cheap, value moves up the stack. GitHub trending this week is almost entirely agent behavior: Anthropic's official plugin directory, DeepSeek's agent plugin harness, ECC for taming agent skills and memory, and viral persona skills like ponytail that make your agent "think like the laziest senior dev."

The #1 repo is the clearest signal: morluto/rea - an agent that reverse-engineers apps down to native binaries - gained ~233k stars in a week (308k total). The velocity smells partly viral, but agent-driven reverse engineering is a frontier no coding assistant touches.

Product Hunt ran the same theme: MoDev puts an AI dev environment on your phone, Clipline cuts viral short videos inside Telegram, and TrackNotch tracks LLM spend in the Mac notch. And Epic Games confirmed a company-wide AI-first pivot - the delayed Disney collaboration will be built "no more handwritten code."

The local lane keeps getting fatter

On Hugging Face, NVIDIA shipped Qwen3.8-Flash-Next-FP4-Local - a 96B MoE in FP4, roughly 50GB of weights, explicitly packaged for fat workstations. LiquidAI's LFM2.5-24B-A2B (2B active params, official GGUF) targets fast local token generation, openbmb's MiniCPM-V-4.7-35B-A3B is a 35B MoE vision model with just 3B active, and DeepSeek's V4.1-Flash flagship stays cloud-only. The week's sweet spot: small-active MoE for everyday work, FP4 96B for headroom, 700B+ for the cloud.

Signal to watch

Nvidia is reportedly in talks to acquire Hugging Face for $13B - not closed. As inference prices collapse, the open ecosystem's model hub is the most valuable asset left: whoever owns distribution owns the next platform.