Inception Labs Releases Mercury 2.5, a Diffusion LLM Built for Speed

Inception Labs Releases Mercury 2.5, a Diffusion LLM Built for Speed

Inception Labs shipped Mercury 2.5 on September 8, claiming a 40% intelligence jump over Mercury 2 while running at 1,107 tokens/sec on NVIDIA GPUs [4]. It's positioned as the largest diffusion language model ever trained, with a 260K context window, tunable reasoning, parallel tool calls, and clean schema-aligned JSON output [5]. Pricing lands at $0.20/$0.75 per million tokens, with a launch discount cutting that to $0.04/$0.15.

This isn't trying to out-think GPT-6 Astra — it's built for the unglamorous, high-volume layer underneath agents: fast, cheap, structured output that doesn't choke on tool-calling. Diffusion-based LLMs have been a research curiosity for a while; this is the clearest signal yet that they're becoming a production category, going head-to-head with speed-optimized models like GPT-5.6 Luna and Claude's Haiku tier.

The buzz on X isn't about raw intelligence — it's about latency and cost per agent-step, which is the metric that actually matters once you're running thousands of orchestrated calls a day instead of one chat response.

Clay Raises $115M Series D at $7.1B Valuation

Clay closed a $115M Series D on September 9, led by Wellington Management with Sequoia, a16z Perennial, DST Global, and CapitalG piling in [6]. Valuation more than doubled to $7.1B. The customer list is the tell: 17,000+ companies including Anthropic, OpenAI, Google, and Stripe, plus 80% of the Forbes AI 50 [7].

Clay's pitch is "self-learning growth agents" for go-to-market — and its new Sequencer feature leans hard into agentic workflows replacing traditional sales-ops headcount. ARR is on pace to hit $200M this quarter, with a target of $240M by year-end and a plan to double again next year. That's not hype-cycle growth, that's a company selling picks and shovels to every AI-native company building revenue engines right now.

The fact that AI labs themselves are Clay customers is the real signal: even the companies building the foundation models need orchestration layers on top to actually run a business. Nobody's exempt from the judgment layer.

Forus Secures $150M Series C at $3B Valuation for AI Medicine Network

Forus (formerly Tandem) raised $150M on September 8, led by Bain Capital Ventures, tripling its valuation to $3B in roughly four months [8]. Total funding now exceeds $300M. The platform uses per-prescription AI agents to automate insurance approvals, patient assistance, and supply logistics between biopharma, doctors, patients, payers, and pharmacies — free for providers and patients [9].

Coverage is already at all 50 US states and 85% of ZIP codes, with partnerships across 9 of the top 15 global pharma companies. That's not a pilot — that's a working national infrastructure layer, and the valuation math (3x in four months) reflects investors betting that healthcare admin is one of the biggest agentic-AI wedges available.

This is a good reminder that "AGI era" headlines aside, some of the most consequential AI deployment right now is unglamorous: paperwork, approvals, logistics. The stuff nobody tweets about until it's the thing that got their prescription approved in an hour instead of three weeks.