xAI Galaxy Day 3: Voice Agent API at $0.08/min, no-code Voice Agent Builder, Agent Tools API and 72-hour Grok Bot company build

The event. xAI (SpaceXAI) ran Grok Bot Galaxy, a three-day showcase at The Howard, 661 Howard Street, San Francisco, with a free global livestream, September 15–17, 2026. The premise, per the official event page and Luma registration: "Can three people build a company in three days?" Three staffers — Matt Palmer (@mattyp), Lauren Tan (@poteto), Roshan Sadanani (@roshan_s) — started with an empty GitHub org and no product, and used Grok Bot (xAI's persistent-agent product, launched Aug 11, 2026) as the primary workforce across ideation, business planning, product, engineering, sales and marketing. Each day also featured department deep dives (Engineering, Product, Founders on Day 1; Sales, SDRs, Support on Day 2; Marketing Ops, Post-Sales, Marketing on Day 3). Day 1 covered "Day 3 Livestream Starts" per the schedule; the finale (Day 3, Sept 17) wrapped the build with a livestream showcase.

The outcome (Day 3). Over the 72 hours the team researched, pivoted, spun a game studio ("Cupcake"), and shipped Thursday Arena (thursdayarena.com) — a live 3v3 browser auto-battler whose cards are real Grok Bot marketplace templates, with Clerk "Login with X", Vercel hosting and PlanetScale data (per the CodeSolutionsLLC Day 1–3 wrap-up, which explicitly flags on-stream metrics as un-audited). Thursday Arena is verified live and playable as of 2026-09-18 (practice as guest, X sign-in for rated play, leaderboard, rules page).

The products showcased. Forkast's analysis (18:06 UTC, Sept 17 — the discovery primary) frames the day's enterprise narrative around three products:

• Voice Agent API — xAI's speech-to-speech voice-agent platform, presented at its current flagship price of $0.08/min (forkast: "launching with a price point of $0.08 per minute for speech-to-speech processing").
• Voice Agent Builder — the no-code builder, "in beta since July 1, 2026" (forkast itself dates it), which lets teams configure and deploy production voice agents without engineering.
• Agent Tools API — described by forkast as enabling "the integration of custom tools, allowing agents to interact with proprietary enterprise software."

The enterprise proof points cited. (a) Haggle Bot, an internal procurement Bot launched in the Bot Marketplace Sept 3, 2026, with an official case study (Sept 4) claiming "more than $100,000 in direct savings" for xAI's own spend; (b) Grok Bot for Enterprise (Sept 3, 2026) adding access, network and audit controls, Action Recording and OpenTelemetry Export; (c) a "VM-per-Agent architecture" the company says backs these deployments, which forkast links to a "$50 million investment in AIR Security" — UNVERIFIED (see section 1; TechCrunch shows AIR independently raised $50M).

Timeline anchor. Forkast's piece is dated 2026-09-17 18:06 UTC (11:06 a.m. PT, mid-Day-3). The same text syndicated to CryptoRank and Yahoo Tech the same day.

• Before: The three products existed but lived on separate pages and release notes: Voice Agent API (Dec 2025)/Think Fast 2.0 at $0.08/min (Jul 29, 2026); Voice Agent Builder beta (Jul 1, 2026); Agent Tools API (Nov 19, 2025); Grok Bot in consumer beta (Aug 11) with enterprise controls (Sep 3). There was no consolidated "autonomous enterprise" narrative: no demo of a full company run by agents, and no procurement-ROI showcase beyond the Sep 4 case study.
• Change (event, Sept 17): Galaxy Day 3 (i) concluded the first live, unscripted 72-hour company build built on Grok Bot, shipping a real product (Thursday Arena); (ii) consolidated and fronted the enterprise agent stack — Voice Agent API, Voice Agent Builder, Agent Tools API — as one "full business automation" offer; (iii) promoted Haggle Bot's $100K+ savings claim as the ROI proof point; (iv) positioned enterprise governance (controls, audit, Action Recording, OpenTelemetry) and an agent-isolation architecture as the enterprise-ready wrapper.
• After: xAI now has a demo-able, livestreamed reference case ("three people + Grok Bot built a company in 72 hours"), a coherent enterprise-agent product story (voice + no-code + tools + governance), and a visible pricing anchor for voice ($0.08/min flagship). For buyers, what actually changed operationally is modest — the components shipped weeks/months earlier — but the enterprise sales narrative is materially stronger: there is now a public record of an end-to-end agent-run business, and a repeated ROI figure.

• Grok Bot (the agent platform). Per xAI's Aug 11 introduction: Bots are persistent agents that "share a computer of their own in the cloud," sign into apps/tools/websites (including platforms with no clean API or MCP), keep working 24/7, and are messaged like colleagues. Routines, teach-a-task skills, approvals, guardrails, and vaults constrain behavior. In the Galaxy build, the team ran an ops fleet (specialist bots — research, inbox, slides, playtest, engineering — behind a front-door/"chief of staff" bot), which is the pattern guests repeated: one bot per job, a manager bot in front, routines for busywork, human gates for pricing/legal/customer-facing promises.
• Voice Agent API (Speech-to-Speech). Real-time full-duplex voice over WebSocket (wss://api.x.ai/v1/realtime?model=grok-voice-latest), a native speech-to-speech model (no text intermediate), server-side VAD, barge-in/turn-taking, tool calling (web search, X search, collections, MCP, custom functions), ephemeral tokens for client apps, phones via Builder telephony (SIP/provisioned numbers). Pricing (docs-verified 2026-09-18): $0.08/min audio ($4.80/hr) + $0.004 per text input message, 10 concurrent sessions/team, 120-min max session, us-east-1. Marketing pages still say "Starting at $0.05/min" for the deprecated Think Fast 1.0 row removed from the rate card.
• Voice Agent Builder (no-code). Browser-based builder on the same Grok Voice stack: pick/clone a voice (80+ preset; clone from ~1–2 min audio), define call handling, connect knowledge bases and tools (incl. MCP; Google Calendar, Notion), deploy via xAI phone number or SIP, monitor calls. Beta since Jul 1, 2026 at $0.05/min + $0.01/min telephony at launch; 25+ languages; sub-second first-audio claim.
• Agent Tools API (server-side tool suite). Hosted tools the model invokes on xAI infrastructure — web_search ($5/1k calls), x_search ($5/1k calls; switching to per-post billing Sep 21, 2026 at 12:00 p.m. PT), code_execution ($5/1k), attachment_search ($10/1k), collections_search/RAG ($2.50/1k), image/video understanding (token-based), and Remote MCP Tools (token-based) for custom tool servers — i.e., the "custom tools/proprietary software" integration forkast refers to. Launched Nov 19, 2025 with Grok 4.1 Fast; the tool suite now serves all Grok models (4.6 flagship).
• VM-per-Agent claim (unverified). Forkast: "These deployments connect to the company's VM-per-Agent architecture, which treats each agent as a distinct entity with its own dedicated computing environment." This conflicts with xAI's own Grok Bot security docs (Sep 3), which describe hosted computers reachable through shared static egress IPs across customers and Bots sharing a cloud computer. The architecture claim is company claim/UNVERIFIED and internally inconsistent with published docs.

• It converts "agents run the back office" from a demo clip into a visible, time-boxed experiment. Three staffers + an agent fleet shipping a live, playable product (Thursday Arena) in 72 hours is a first-of-kind public artifact for AI-agency claims — and because it was livestreamed, its failure modes (pivots, guardrail decisions, human taste calls) are on the record, which is rare. Caveat: it is a vendor showcasing its own tool (KuCoin's flash explicitly notes this is "a vendor showing off its own tool rather than an independent trial").
• The voice price point is a real market signal — but the framing in the discovery primary is inverted. The docs-verified rate is $0.08/min for grok-voice-think-fast-2.0, which is a 60% increase over Think Fast 1.0's $0.05 (Jul 29, 2026), not a discount launch. OpenAI put GPT-Live-1 into the developer API at $0.05/min on Sep 10, 2026 — exactly the price xAI vacated. Forkast frames xAI as "undercutting Deepgram and ElevenLabs"; the actual competitive picture is xAI at $0.08 vs OpenAI at $0.05, with ElevenLabs Agents at ~$0.08-equivalent. Analysts (OrcaRouter, eesel) read this as xAI pricing a capability raise, not a price war.
• The $100K+ savings figure is the strongest enterprise proof point and the weakest evidence item. It is xAI's number about xAI's own spend (~125 vendors; $14,220 from 43 idle seats; $85,662/yr in unused SKUs; a $14,629→$6,143 order example), published as a case study with the full system prompt — useful as a repeatable pattern, misleading if quoted as a customer-ROI benchmark (independent commentary: aireiter, dev.to/Toolkit).
• The agent-security angle is the quiet strategic read. Forkast's garbled "$50M AIR Security investment" points at a real category: as agents get credentials and network access (VM-per-Agent or not), the supply chain of skills/plugins/MCPs needs vetting. AIR raised $50M (Sep 1), Zenity $125M Series C (Aug), Noma $100M Series B — the market is pricing this risk independently of xAI.

• For xAI: a reusable, honest demo format ("build a company in 3 days") that generates artifacts, templates and case studies; an enterprise story that now includes voice ($0.08/min), no-code builder, tools, governance and a stated ROI case; and a marketplace (Bot Marketplace) whose templates became the very content of the shipped game — self-referential "dogfooding" at product scale.
• For developers and founders: a worked, copied-able playbook — one-bot-one-job, manager-bot front door, teach-a-task skills, routines, human guardrails (the CodeSolutionsLLC wrap-up's "seven durable lessons"); voice agents at a published per-minute rate with a no-code on-ramp for non-engineers; and server-side tool pricing that is per-invocation and predictable ($5/1k class, MCP token-based).
• For enterprises: a governance surface (audit logs, Action Recording, OpenTelemetry export) plus free 2-week enterprise trials — enough to run a bounded pilot; procurement automation demonstrated with explicit human-in-the-loop controls (Haggle Bot "never spends, signs, or sends without you").
• For the public record: a falsifiable claim pair — "$100K+ savings in procurement" and "a company built in 72 hours" — that later artifacts (customer case studies, audited figures, Thursday Arena's continued operation) will confirm or refute.

1. Did xAI intend "launched on Day 3" (forkast) or "showcased on Day 3"? No official Sept-17 announcement post exists — what, if anything, did the Day-3 livestream actually announce as new?
2. Is the VM-per-Agent architecture real, and does it supersede the documented shared-computer/egress model? When will it be documented?
3. Did xAI invest in AIR (air.security), or was the "AIR Security" line a conflation of AIR's independent $50M raise? No official statement found.
4. What is the actual composition of the "$100K+ savings" beyond the disclosed $14,220 + $85,662? Are any savings customer-replicated yet?
5. When do Grok Bot Enterprise per-seat prices, quotas and overage rates get published (free-trial pricing disclosed only)?
6. Will the Voice Agent Builder's published price move from the $0.05-era marketing lanes to the $0.08 rate card?
7. Does the X Search per-post billing change (Sep 21) apply inside voice agents' x_search tool, and how much does it move agent TCO?
8. What happened to Thursday Arena after the event — ongoing operation, monetization experiments, or archive?

• Immediate (this week): Sep 21, 2026 — X Search moves to per-post/per-profile billing ($5/1k posts, $10/1k profiles), affecting all agent tooling that uses x_search; post-Galaxy enterprise-trial conversions begin; watch for xAI clarifying the "AIR Security" and "VM-per-Agent" claims (or staying silent).
• 1–3 weeks: reference to Thursday Arena in xAI marketing/marketplace; any official Day-3 recap post that would retroactively define what was "announced"; Grok Bot Enterprise pricing disclosure; first customer procurement-agent case studies.
• 4–8 weeks: competitive voice-agent price responses to the OpenAI $0.05 GPT-Live-1 launch; possible Think Fast 2.x iteration (xAI's cadence) resetting the $0.08 anchor; agent-security vendor land grabs (AIR, Zenity, Noma) targeting Grok Bot fleets.
• Structural: whether "build-a-company-in-72-hours" becomes a recurring xAI format (each Galaxy/iteration → new artifacts/templates); whether the enterprise trio (voice + builder + tools) is packaged into a single enterprise SKU; whether the claimed VM-per-Agent isolation shows up in documentation or stays a pitch line.

The news is not three products shipping — they all shipped weeks or months earlier. The news is that xAI spent three public days demonstrating, live, that a small team plus a fleet of persistent agents can take a company from blank slate to a running product (Thursday Arena), and that it used the finale to consolidate its enterprise-agent stack — voice at a documented $0.08/min (a 60% raise from $0.05 in July, not the "undercutting" the discovery primary implies), a no-code builder, a server-side tools suite with published per-call prices, and governance controls. Report the event as CONFIRMED; report the ROI ($100K+ savings — internal case study), the "functioning company" framing, and the $50M "AIR Security investment" (unverified; AIR raised $50M independently) as COMPANY CLAIM/UNVERIFIED. The durable story is the format and the price signal: agent-run business-building is now demo-able rather than hypothetical, and voice-agent pricing is a contested market where xAI is no longer the low-cost leader. The discipline this story demands is the same as every agent-economy story: separate the livestream from the ledger, the case study from the benchmark, and the price card from the press line.

Key numbers for the report: $0.08/min + $0.004/msg (grok-voice-think-fast-2.0, docs-verified 2026-09-18); $4.80/audio-hour; $0.05/min (deprecated 1.0; OpenAI GPT-Live-1 at $0.05 since Sep 10); tools $5/1k (web/x/code), $10/1k (attachments), $2.50/1k (collections), MCP token-based; X Search per-post billing Sep 21; Haggle Bot >$100K claim ($14,220 idle seats + $85,662/yr unused SKUs; $14,629→$6,143 example); ~125 vendors mapped; Voice Agent Builder beta Jul 1, 2026; Agent Tools API since Nov 19, 2025; event Sept 15–17, 2026 at The Howard, SF; 72-hour build by Palmer, Tan, Sadanani → thursdayarena.com (live); AIR $50M raise (Sequoia $10M, Greenoaks $40M, Sep 1) — no xAI investment found.