Anthropic Builds an In-House Custom Silicon Team to Run Claude Cheaper
Anthropic has confirmed it’s assembling an in-house team to design its own AI chips, aiming to run Claude faster and more cost-efficiently at the scale its customers now demand. The pitch is co-design: silicon and models built in tandem, with a stated target of roughly 50% cuts in per-token inference costs.
Recent job listings for the “custom silicon team” offer $320K–$485K and want engineers fluent across the hardware/software stack. Technical leadership is anchored by Clive Chan, who joined in June from OpenAI’s chip effort and, before that, Tesla’s Dojo program. Anthropic stresses this is additive — it keeps its multi-chip approach, including massive TPU capacity from Google and Broadcom.
Sonnet 5’s Introductory Pricing Ends August 31 — Budget Now
Anthropic has confirmed that Claude Sonnet 5’s promotional pricing of $2 / $10 per million input/output tokens ends Aug 31, 2026. On Sept 1 it reverts to standard $3 / $15 — a flat 50% increase on both dimensions.
The sticker isn’t the whole story. Sonnet 5 ships a new tokenizer that can turn the same text into up to 35% more billable tokens, so real bills can climb well past the headline 50%. If you’re running Sonnet 5 in production, the next three weeks are the window to audit prompt sizes, cache aggressively, and model out September before the meter changes.
Claude Code Adds Self-Hosted Environments and Zip-Based Plugin Installs
Claude Code’s latest changelog brings self-hosted environments: claude self-hosted-runner turns your own machines or containers into a place web, mobile, and desktop sessions can run — on Team and Enterprise plans. It’s a real answer for teams that can’t send workloads to managed cloud sandboxes for compliance reasons.
Also new: archive plugin source installs plugins from a zip over HTTPS with no git or npm and optional SHA-256 pinning, plus the removal of the subagent spawn cap and a stack of fixes — including stream idle timeouts on custom ANTHROPIC_BASE_URL gateways and connectors falsely flagged as needing re-auth (now a /login hint).
Public Sharing Links, Editor Roles, and Artifacts From Claude Tag Sessions
This week also brought collaboration plumbing: public sharing links for conversations and editor roles that allow shared editing on Team and Enterprise plans — more than one person can now iterate on the same thread with real permissions instead of copy-paste.
On top of that, artifacts can now be created directly from Claude Tag (Claude-in-Slack) sessions, so a conversation that starts in Slack can spin out a live document or app without leaving the flow. Small features individually; together they nudge Claude further from a single-player tool toward a team surface.
Tino Cuéllar Joins Anthropic as Chief Global Affairs Officer
Mariano-Florentino (Tino) Cuéllar joined Anthropic as Chief Global Affairs Officer on Aug 4. A former California Supreme Court justice and current president of the Carnegie Endowment for International Peace, he brings heavy policy and governance weight to a company steering an IPO through a thicket of AI regulation.
The hire fits the pattern in this week’s product releases too — DLP inference hooks, org-wide controls, self-hosted runners. Anthropic is staffing and shipping for the part of the market where governance, not just capability, decides the deal.
The Enterprise Base Keeps Compounding: 1,000+ Customers at $1M+
The backdrop to the hardware and pricing moves is a revenue curve that’s hard to overstate: Anthropic’s annualized run-rate has climbed toward roughly $47B, and the company says more than 1,000 business customers now spend at least $1M a year, with eight of the Fortune 10 on board.
Claude Code alone reportedly draws more than half its revenue from enterprise, with names like Netflix, Spotify, KPMG, and Salesforce cited. Distribution deals keep stacking too — a multi-year $200M Snowflake agreement puts Claude models inside another major data platform. This is the demand that a cheaper, self-designed inference stack is meant to serve.
Chips Plus Pricing: This Is a Margin Story Now
Read today’s two big items together — a custom silicon team targeting 50% cheaper inference, and Sonnet 5’s price rising 50% on Sept 1 — and the strategy is unmistakable. Anthropic is squeezing cost out of the stack it controls while letting introductory subsidies roll off. Both levers point the same way: toward margin, ahead of a public offering.
That’s the tell of a company transitioning from land-grab to durable business. When you’re at a ~$47B run-rate with a confidential IPO on file, the question stops being “can we win users” and becomes “what does each token cost us to serve.” Designing your own chips is the most capital-intensive way to answer that — and the fact that Anthropic is doing it anyway says a lot about how permanent it thinks this demand is. For buyers, the near-term takeaway is simpler: lock in intro pricing while you still can, and watch whether those promised inference savings ever reach the invoice.