Mercury 2.5's 80% discount ends at 07:00 UTC today — a 5x price jump
The diffusion LLM writes tokens in parallel — ~275/sec at Haiku-tier cost — and its intro price won't last the morning. Plus: China's new deepfake rules.
Copy markdown
The 5x jump lands at 07:00 UTC — hours away
Mercury 2.5's launch promo — $0.04/M input, $0.15/M output — reverts to list at 07:00 UTC (11 AM Dubai): $0.20 input, $0.75 output, cache reads $0.004→$0.02/M. Every promo figure is exactly one-fifth of list, so it's a clean 5x across the board. Requests you send before the cutoff still bill at the old rate.
Why it's quick: it writes tokens in parallel
Instead of emitting one token at a time like a standard autoregressive model, Mercury is a diffusion LLM — it drafts and refines many tokens at once. Inception claims 1,107 tok/s on standard GPUs; OpenRouter's live endpoint clocks ~275–477 tok/s, still several times a typical fast model. This is the tier where time-to-last-token, not just quality, is what your users feel.
The quality tier — with an asterisk
Inception pegs 2.5 as a 10-point intelligence jump over Mercury 2 and roughly on par with Claude Haiku 4.5, Gemini 3.5 Flash-Lite, and GPT-5.6 Luna (low). It ships a 260K-token context, 64K max output, and an OpenAI-compatible API. No independent benchmark has confirmed the parity claims yet, so treat the tier as reported, not proven.
What to point it at this weekend
Cheap, fast, and big-context favors high-volume inner loops: bulk classification and extraction, autocomplete, agent tool-routing, and streaming UIs where every dropped millisecond shows. Even at the new list price it undercuts most frontier minis — but grabbing the promo before 07:00 UTC keeps your bill at a fifth of that.
Elsewhere: China sets deepfake liability rules
China's Supreme People's Court issued its first judicial guidelines on AI disputes Monday, making AI-generated deepfakes and voice clones legally actionable and spelling out platform liability. If you ship generative audio or video into China, provenance and consent handling just moved from nice-to-have to compliance.