xAI

Overview

AI research company founded by Elon Musk (2023). Develops the Grok model series. Leverages real-time data access to the X (Twitter) platform as a key strength. In 2026, expanding into agentic coding (Grok Build) and voice AI.

Key People

• Elon Musk — founder, CEO

Models & Products

• Grok 4.6 — Grok 4.6: released 2026-08-07 (reported; no first-party page read, and four write-ups say 2026-08-12 instead). 500K context, $2/$6 per M with the whole request re-pricing above 200K, Artificial Analysis Intelligence Index 61 at $0.84/task — all third-party, read 2026-08-16. Still no xAI model card and no parameter count: disputed at 2T announced 2026-07-18 against 1.5T on the V9 foundation in launch coverage.
• Grok 4.7 — Grok 4.7: released 2026-09-21, nine days after the third and last expired window. 2.1T parameters (against Grok 4.6's 1.5T), 500K context, $2/$6 per M at ≤200K prompts — unchanged from Grok 4.6 — doubling above 200K. Four effort levels, high the default. Vendor benchmarks: CursorBench 4.0 46.3%, EEBench 64.0%, Harvey Legal Agent Benchmark 19.6%. Still no model card, no licence statement and no parameter confirmation. (source)
• Grok 4.5 — Grok 4.5 public launch 2026-07-08/09: V9 1.5T + Cursor-trained; $2/$6 per Mtok; Opus-class (wins 2/4 benchmarks vs Opus 4.8); default model in Grok Build + all Cursor plans; not yet in EU. (source)
• Grok V9-Medium — Grok V9-Medium shipped to consumers (Grok on X + SuperGrok) 2026-06-16: 1.5T parameters (3× the prior 500B V8-small), trained on Cursor workflow data — coding-specialized, targets Claude's coding lead. Not in xAI API docs as of 2026-06-19. Musk: "Much better at coding." (source)
• Grok 5 (in training): 6T-parameter MoE. Training on Colossus 2 (Memphis, 1GW, 550K GB200/GB300). Q2 (June 30) window closed without release; July now target.
• grok-4.3 — current production model (May 2026, consolidates the grok-4 line; grok-4/4-fast/4.1-fast deprecated → redirect targets)
• Grok 4.1 Fast (xAI) — best tool-calling model, 2M context, includes Agent Tools API ($0.20/$0.50 per 1M tokens)
• Grok Build — terminal-native agentic coding CLI (2026-05-14 early beta, SuperGrok Heavy)
• Grok Automations — scheduled AI tasks (all users) + email-triggered automations (SuperGrok); launched 2026-07-16
• Grok Connectors — Gmail, Drive, Notion, GitHub, Linear, SharePoint, Outlook integration + BYO MCP (2026-05-06)
• Grok Voice API — multi-step workflow voice agent, 25+ languages; engine is Grok Voice Think Fast 2.0 since 2026-07-29 (the grok-voice-latest alias moves to it 2026-08-05), previously grok-voice-think-fast-1.0
• Grok Voice Transcribe 2.0 — Grok Voice Transcribe 2.0, released 2026-09-18: speech-to-text on the Grok Voice audio foundation model, $0.10/hour batch · $0.20/hour streaming (unchanged from 1.0), claimed "twice as accurate" as 1.0 and first among 32 streaming models on Artificial Analysis. Word timestamps, diarization, multichannel, automatic language detection. No API model id, no WER figure, no licence. Captured day +7
• Grok STT / TTS — standalone speech APIs
• Grok Imagine Image 2.0 — Grok Imagine Image 2.0, released 2026-08-07 as Quality Mode on grok.com/imagine and the mobile apps. Editing-first: region-level magic wand, segmentation, background removal, up to five reference images per generation. Claimed #2 on both Arena image tracks behind gpt-image-2. No API at launch. (source)
• Grok Imagine Quality Mode API — enhanced image generation

Recent Activity

• 2026-09-29: Grok is now one of two models behind a US federal services portal. America.gov launched as a chat interface answering questions drawn from roughly 29,000 federal websites, powered by Google Gemini and xAI Grok, led by Joe Gebbia as US Chief Design Officer (source). Named example queries: job postings, UFO files, passport, Medicare, veteran benefits, campsite availability.

  What is not stated in anything read is most of what would matter: no model version for either Gemini or Grok, no account of how the two are routed or divided, no accuracy or evaluation figure, no grounding or citation mechanism beyond "scans across" the corpus, no contract value or procurement vehicle, and no statement of what happens when the two models disagree. Two agreeing search passes, no first-party read. Recorded here as a deployment surface, not as a capability claim about Grok 4.7 or any other named model — nothing read attaches a version to it. For the government-deployment thread see AI Governance

• **2026-09-18 [_, Grok Voice Transcribe 2.0, a speech-to-text model built on the audio foundation model behind Grok Voice, trained on live, noisy, multilingual audio and refined with post-training. $0.10 per hour batch, $0.20 per hour streaming — unchanged from 1.0. Claimed "twice as accurate" as 1.0, rendered by another pass as "half the errors"; claimed first for accuracy among 32 streaming models on the public Artificial Analysis leaderboard. Features: multilingual transcription (named as the largest gain) with automatic language detection and mid-recording language switching in a single pass, word timestamps, diarization, multichannel support, smarter turn detection. Why it matters: the cause is identical to the Grok 4.7 capture three days ago at +3, and it is structural rather than incidental — xAI publishes no feed, so nothing it ships reaches state/prefetch.json; the Chinese-lab rotation does not cover it; and x.ai answers EGRESS_BLOCKED, so the non-feed sweep is a search query against a lab whose releases do not reliably surface under its own name. Two misses in one week is a pattern with two instances, and it is carried to the W39 lint. Not established: no API model id; no word error rate, baseline or dataset behind either accuracy claim; the Artificial Analysis placement could not be verified — this repo's snapshots carry the intelligence board, not a speech board; no language list; no licence; no audio-length limit; no latency figure; no model card or technical report (source)

• 2026-09-21: Grok 4.7 shipped, and this wiki said it had not for two more days — Grok 4.7 was released 2026-09-21, nine days after the third expired window and thirty after the first. 2.1T parameters against Grok 4.6's 1.5T, 500K context, four effort levels with high the default, and $2/M input · $0.50/M cached · $6/M output at ≤200K prompts, doubling above 200K — the ≤200K rates identical to Grok 4.6 (3 passes). Vendor benchmarks: CursorBench 4.0 40.4% → 46.3%, EEBench 53.0% → 64.0%, Harvey Legal Agent Benchmark 15.8% → 19.6%, the last reported against GPT-5.6 Sol 2.5% and Fable 5.1 6.7%. Available in the Grok API, Cursor, Grok Build, OpenRouter, Vercel and Cloudflare. Why it matters: cadence is this lab's central competitive claim and it has now been re-established at the cost of a thirty-day slip — but the release is measured entirely against GPT-5.6 Sol and Fable 5.1, and nothing read compares it to Claude Opus 5.5, GPT-6 Sol or GPT-6 Luna, which shipped the next day. The capture is this pipeline's failure and is recorded as one: the 2026-09-22 and 2026-09-23 ingests both wrote that Grok 4.7 was unreleased, after it had shipped. xAI publishes no feed, so it never enters state/prefetch.json, and the lab-named rotation query returned nothing — the release surfaced today only through a version-named query, the third instance of that pattern since 2026-09-20. Not established: no max output figure, no licence, no model card or system card, no safety evaluation of any kind, no independent measurement, and whether the RL length-penalty defect Musk named on 2026-09-11 was fixed — a "longer RL pass targeting problems that require extended reasoning" is adjacent to that diagnosis and is not stated as a fix for it → Grok 4.7 (new), Eval Harness Configuration (source)

• 2026-09-12/13: Musk endorsed a rival's plan to slow AI down in three words, and xAI committed to nothing — responding to Amodei's We Must Pace the Frontier, Musk posted "Dario is right", one pass placing it within roughly an hour of publication (4 passes for the quote). Why it matters: this is the shortest distance between xAI and any safety-coordination proposal this page holds, and it arrived in the same week the lab's own flagship missed a third release window. It is an endorsement, not an action: of the four lab leaders who replied, only Altman adopted a step, and nothing read reports xAI committing to embedded evaluators or to anything else in the plan. Musk signed the 2026-07-28 employee statement's object nowhere — xAI does not appear in the employer table on Frontier Pacing at all — so this is his first recorded position on it. Not read first-party — x.com is not fetchable from this run → Frontier Pacing (source)

• 2026-09-11/12: Grok 4.7 misses a third window, and xAI names its own RL reward design as the cause — Musk said on 2026-09-02 that "Grok 4.7 comes out in 10 days", giving 2026-09-12; on 2026-09-11 he said it "needs a few more days to cook", and the window closed with no new date offered (2 passes each). This is the third window this model has been given and the third to expire — ~2026-08-22 from the 2026-07-25 "~4 weeks", then two more. The stated reason is specific enough to be checkable: the model "still stops too early on some difficult tasks" and "does not check its own work rigorously enough", because xAI "may have penalized response length too aggressively during reinforcement learning", causing it to abandon problems it is otherwise capable of solving. Reported at 2.1T parameters, in final training, trained in part on SpaceX engineering data (1 pass); no model card, benchmark package, API pricing or rate limits published. Why it matters: cadence is xAI's central competitive claim — Grok 4.5 on 2026-07-08 and Grok 4.6 on 2026-08-07 both landed inside their windows — and this is the first month it has not held. The more useful half is the diagnosis: a lab publicly attributing a capability regression to its own reward shaping, and the failure mode it describes (a length penalty producing premature termination on hard tasks) is exactly the exploratory-behaviour suppression Agentic Reinforcement Learning and Negative Self-Distillation: Learning to Reason by Avoiding Flaws record from the research side this week. The 2.1T figure is new here and does not resolve anything: Grok 4.6 carries 2T announced against 1.5T in launch coverage under ## Conflicting Reports, and nothing read this run settles it. Not read first-party — x.com is not fetchable from this run and every line is a search extract with a pass count (source) (datastudios) (OrcaRouter)

• 2026-08-28: The Cursor acquisition has its first cost, and it is a supplier walking away — OpenAI published Our decision on Cursor following its acquisition by SpaceX, ending Cursor's direct access to OpenAI models on 2026-11-12, and named as its reason that it cannot be confident SpaceX will honour its terms of service — citing the Twitter acquisition and xAI having admitted violating those terms. Why it matters: this page has recorded three separate Grok models as Cursor-trained — Grok V9-Medium, Grok 4.5 and the V9 foundation behind them — so Cursor is not an unrelated SpaceX property but the source of the workflow data xAI's coding line was built on. Owning it now costs the product a supplier. Cursor CEO Michael Truell put OpenAI's share at about 5% of traffic, which is the counter-argument and the only quantity anyone has published. What is not established here: nothing read states the acquisition's completion date first-hand — coverage says 2026-08-14, this wiki holds the 2026-06-16 announcement of the same $60 billion all-stock deal with a Q3 close expected, and the completion is recorded on the capture rather than promoted to this page (source) (acquisition)

• 2026-08-23: Fastest business-adoption growth of any lab in the index, from the smallest base — the Ramp AI Index for August 2026 puts xAI at 4.0% of US businesses paying for subscriptions or tokens in July, +0.94pp month-over-month and its fastest growth since July 2025, against Anthropic at 43.5% (+1.1pp) and OpenAI at 39.7% (+0.23pp) (source). Why it matters: in percentage-point terms xAI added nearly as many business customers as Anthropic did this month while holding a tenth of the base, so it is growing several times faster in relative terms — and it did so in the month Grok 4.6 shipped without a model card, a price or a context window published. The index measures payers, not spend, and 4% is small enough that a percentage point is a modest absolute number; recorded as a rate, not a position. → Model Routing (Ramp)

• 2026-08-16: Grok 4.6's specification arrived nine days after the model did — from documentation sites, not from xAI — three rows that had read unknown since launch now have values: 500,000-token context, $2/M input · $6/M output, and an API. Above 200K tokens the entire request re-prices at $4.00/M input · $1.00/M cached input · $12.00/M output — a cliff, not a tier, so the marginal cost of crossing it applies to every token before it. Knowledge cutoff 2026-02-01; text and image in, text out; low/medium/high/xhigh effort levels. Artificial Analysis, absent Grok 4.6 on 08-09, now measures it at 61 on the Intelligence Index at $0.84 per task — level with GPT-5.6 Sol (max) and Claude Opus 5 (high), against $1.23 and $2.34 for the Anthropic rows. Why it matters: xAI's competitive claim has been cadence, and the price is now the sharper one — frontier-band parity at roughly a third of Opus 5 (max)'s measured cost per task. What has not changed is that no xAI model card, pricing page or benchmark table has been read from this environment, so all of the above is third-party; there is still no parameter count; and four of the sources read date the release to 2026-08-12 against the 2026-08-07 this wiki has published since 08-08. Both dates are third-party and the conflict is recorded, not resolved. → Grok 4.6 (source) (AA snapshot) (appwrite) (mem0)

• 2026-08-07: Grok Imagine Image 2.0 shipped the same day as Grok 4.6 — and the image line went app-first while the video line went API-first — xAI released Grok Imagine Image 2.0 on 2026-08-07, generally available as Quality Mode on grok.com/imagine and in the iOS and Android apps. The release is an editing tool rather than a generator refresh: a magic wand that changes only the region pointed at, segmentation for precise area selection, background removal with transparency, multi-reference editing accepting up to five input images in one generation, and smart resize across nine aspect ratios. xAI claims #2 in the world on both the Arena text-to-image and image-edit leaderboards as of 2026-08-07, behind OpenAI's gpt-image-2 in both — a claim relayed by reporting, since no first-party page and no leaderboard was readable this run, and this repo's own sources/evals/ snapshots carry no image-edit column to check it against. No API at launch, which is reported rather than inferred. Why it matters: xAI shipped a frontier text model and a top-two-claimed image model on the same day, but in opposite directions — Grok Imagine Video 1.5 (Preview) launched API-first at $0.08/second and this one launched app-first with no API, which reads as consumer capture rather than developer platform for the image line. → Grok Imagine Image 2.0 (source) (SpaceXAI) (Unite.AI)

• 2026-08-07: Grok 4.6 shipped on schedule — and the parameter count in the launch coverage is not the one Musk announced — Grok 4.6 is reported released on 2026-08-07, a day inside the ~August 8 target Musk gave on July 25 and consistent with his August 4 "likely next week". The cadence xAI has been claiming since June therefore holds for a second month: 4.5 on July 8, 4.6 on August 7, 4.7 targeted ~August 22. What does not hold is the specification: Musk announced 4.6 on 2026-07-18 as a 2T model, "33% larger" than Grok 4.5; launch coverage describes 1.5T on the same V9 foundation, with gains from improved SFT and RL instead of scale. Both are recorded on Grok 4.6 under ## Conflicting Reports rather than resolved. No first-party xAI launch page, model card or pricing was readable from this run — the egress proxy blocked every host attempted — so the release itself is carried on third-party reporting, and Context window, Pricing and Availability all remain unknown a day after launch. Why it matters: a monthly frontier cadence is xAI's central competitive claim, and it is being met. But the two releases this wiki can compare are a scale story that became a post-training story somewhere between announcement and ship — and with no model card behind either, there is nothing to check it against. → Grok 4.6 (source) (kie.ai) (AIToolsReview)

• 2026-07-29: Grok Voice Think Fast 2.0 released — and today it becomes the default for every grok-voice-latest caller — SpaceXAI released its next-generation speech-to-speech model on July 29, positioned on intelligence, transcription accuracy, conversational behaviour and tool use. The design claim is that it reasons in parallel with speech, holding latency down while handling harder queries: ~0.70 s to first audio response, against 1.25 s for the previous Think Fast model. Reported figures on Artificial Analysis' speech-to-speech benchmark, overall: 82.9%, up from 75.7% for Think Fast 1.0, ahead of GPT-Realtime-2.1 at 79.1% and Gemini 3.1 Flash at 69.5% — recorded as reported, since this repo holds no Artificial Analysis snapshot carrying a speech-to-speech column. Pricing $0.08/minute speech-to-speech plus $0.004 per text input. On 2026-08-05 the grok-voice-latest alias moves automatically from grok-voice-think-fast-1.0 to 2.0, with xAI stating the upgrade should improve almost all use cases without prompt edits. No parameter count, architecture or context window published. Why it matters: xAI's voice stack has until now been measured on its own tau-voice Bench (Think Fast 1.0 at a self-administered 67.3%); a third-party benchmark placing it ahead of OpenAI's realtime model is a different kind of claim, and the automatic alias flip means every existing integration inherits it without opting in. → Grok Voice Think Fast 2.0 (source) (SpaceXAI) (TestingCatalog)

• 2026-07-25: Grok 4.7 first confirmed — Musk gives explicit timelines for both 4.6 and 4.7 — Elon Musk publicly confirmed both Grok 4.6 and Grok 4.7 release timelines on July 25, 2026: Grok 4.6 is "~2 weeks" out (~August 8); Grok 4.7 is "~4 weeks" out (~August 22). This is the first public confirmation of Grok 4.7 as a distinct upcoming model. The July 25 statement also sharpens the previously ambiguous "~early August" window for Grok 4.6 to a specific ~Aug 8 target. The monthly cadence (Grok 4.5 Jul 8 → 4.6 ~Aug 8 → 4.7 ~Aug 22) is now publicly confirmed through at least August 2026. No specs, benchmarks, parameters, or pricing for Grok 4.7 have been disclosed. Why it matters: confirms SpaceXAI intends to ship two more frontier updates before end of August — both faster and more frequent than any other frontier lab's current confirmed release cadence. → Grok 4.6 (source) (NextBigFuture)

• 2026-07-24: Grok Google Workspace add-on launched — free integration in Sheets, Slides, and Docs — SpaceXAI launched a free Grok add-on for Google Workspace (install: Google Workspace Marketplace). Grok appears as a sidebar panel in each application. Capabilities by app: Sheets — read selected ranges, explain data with cell citations, write formulas, update values and rebuild financial scenarios; Slides — generate full presentations from an outline, add theme-matching slides, create charts from web or X data; Docs — transform rough notes into formatted documents, draft and refine content. Cross-platform context: one week earlier (July 17–18), SpaceXAI launched Microsoft Office add-ins (Excel, Word, PowerPoint). The two launches are a deliberate cross-platform productivity strategy — Grok as the AI layer spanning both major office suites. Why it matters: Grok's prior integration advantages were tied to X data and voice; the dual office-suite rollout extends Grok into knowledge-worker workflows where ChatGPT and Copilot are entrenched. Free pricing removes the adoption barrier. (source) (SpaceXAI) (Marketplace)

• 2026-07-18: Grok 4.6 training announced — 2T parameters, targeting ~early August launch (updated Jul 27: Musk posted Jul 24 "~2 weeks" → implies ~Aug 7) — Elon Musk announced on X (July 18) that xAI's next model, expected to be named Grok 4.6, will complete its initial training run the week of July 20, 2026. Key claims: 2T parameters (vs. Grok 4.5's 1.5T — 33% larger), "better than [Grok 4.5] in every way." Speed target: keeping token efficiency close to Grok 4.5. No benchmarks, pricing, context window, or launch date published. Musk framed it as a direct response to Moonshot AI's Kimi K3 (2.8T MoE, announced July 16). Updated (Jul 27): Musk posted on X (Jul 24) that Grok 4.6 is "~2 weeks" out — implying ~Aug 7 launch for 4.6 and ~Aug 21 for 4.7 if the monthly cadence holds. → Grok 4.6 → Grok 4.6 (source) (Dataconomy) (WION)

• 2026-07-16: Grok Automations launched — scheduled AI tasks and email triggers — xAI launched Grok Automations, a feature that lets users describe a recurring job once, set a schedule or trigger, and have Grok execute it automatically with a run history. Schedule types: once, daily, weekdays, weekly, monthly, yearly in the user's timezone. Email triggers (SuperGrok only): watches inbox for matching messages (filter by sender, recipient, or subject); when triggered, passes incoming email as context and Grok responds to the actual message. Execution model: each automation run is a fresh conversation (same instructions, current data); connectors and skills can be attached; results saved to run history with notification options (email, app, both, or none). Available across grok.com, iOS, and Android; email triggers require SuperGrok. Why it matters: extends Grok from a conversational interface to an always-on background agent. Email-triggered automations position Grok as an inbox agent competing with ChatGPT Work, Anthropic Cowork, and enterprise automation tools (n8n, Zapier). Combines with xAI's existing Voice Agent Builder and Grok Connectors to form a coherent ambient agent stack. (source) (xAI)

• 2026-07-16: Grok Build CLI open-sourced under Apache 2.0 — 844K lines of Rust — xAI released the full Grok Build CLI codebase under the Apache 2.0 license ~72 hours after the privacy breach reached Hacker News front page (July 14). Repository: github.com/xai-org/grok-build; 844,530 lines of Rust; covers the agent loop, all tool implementations (file edit, shell exec, web search, doc retrieval), terminal UI (TUI), and extension system (Agent Client Protocol / ACP). Concurrent announcements: usage limits removed for all Grok users; data storage defaults to off since July 12. Why it matters: first agentic coding CLI by a frontier lab (other than Claude Code) to be fully open-sourced under a permissive commercial license. The open-source release enables independent security audit of the agent loop and upload behavior — a direct response to the trust erosion from the breach. Apache 2.0 allows unrestricted commercial use; forks and derivatives are possible immediately. → Grok Build (source) (xAI) (github.com/xai-org/grok-build)

• 2026-07-12–14: Grok Build CLI v0.2.93 — full Git repository upload privacy breach disclosed — Security researcher cereblab published wire-level analysis proving that Grok Build CLI silently uploaded entire local Git repositories to a Google Cloud Storage bucket (grok-code-session-traces) managed by xAI — including full commit history, untracked files, and unredacted .env secrets (API keys, database passwords, SSH keys). One test: 5.1 GiB uploaded from an 11.2 GiB repo when only 192 KiB was needed — ~27,800× excess. xAI had marketed: "nothing from your codebase transmitted to xAI servers during a session." The wire data directly contradicts this. The /privacy toggle was ineffective — it only controlled post-upload data retention, not the upload itself. Story reached Hacker News front page July 14. xAI response: server-side flag disable_codebase_upload: true deployed (stopped new uploads); new /privacy command added. Elon Musk posted July 13: "all user data that was uploaded to SpaceXAI before now will be completely and utterly deleted." Deletion not independently verified as of July 14. Security advisory: any developer who ran Grok Build CLI before July 13, 2026 on a repo containing secrets in tracked files or Git history should treat those credentials as potentially compromised. → Grok Build (source) (The Hacker News) (Cybernews)

• 2026-07-06: 21 new multilingual Grok Voices launched — voice roster expands from 5 to 26 — SpaceXAI released 21 new flagship voices for Grok Voice on July 6, joining the original 5 (Ara, Eve, Leo, Rex, Sal — all enhanced with improved pacing and emphasis). Every voice is natively multilingual, supporting all of Grok Voice's 25+ languages (English, Hindi, Spanish, French, German, Japanese, Korean, Thai, and more). New voices include: Carina, Zagan, Helix, Orion, Luna, Iris, Altair, Zenith, Perseus, Helios, Lux, Kepler, Rigel, Cosmo, Celeste, Ursa, Sirius, Lumen, Castor, Naksh, Atlas. Additional feature: voice cloning from a reference audio clip up to 120 seconds. Available across the Realtime Voice Agent API, Text to Speech API, and Grok Voice Agent Builder. Use-case optimization: support, characters, commentary, advertising, education. Speech tags for delivery control also added. Why it matters: the 5→26 voice expansion, native multilingualism, and cloning capability positions Grok Voice as the most feature-complete voice API in the frontier-lab market — directly competing with ElevenLabs and OpenAI's TTS/Realtime stack. → (source) (xAI)

• 2026-07-08/09: Grok 4.5 public launch — Musk announced July 8; went live July 9. Available in Grok Build (now default), Cursor (all plans), and SpaceXAI console. Not yet in EU. Pricing: $2/$6 per Mtok. Published benchmarks vs Opus 4.8: DeepSWE 1.0 62% vs 55.75% (wins), DeepSWE 1.1 53% vs 59% (loses), Terminal Bench 2.1 83.3% vs 78.9% (wins), SWE-Bench Pro 64.7% vs 69.2% (loses). Token efficiency: 4.2× fewer output tokens per task than Opus 4.8. → Grok 4.5 (source)

• 2026-07-06: xAI rebrands to SpaceXAI — On July 6, 2026 xAI adopted the SpaceXAI name and logo, nesting the xAI identity visually inside the SpaceX brand. X handle changed to @SpaceXAI. Product names unchanged (Grok, Grok Build, Voice Agent Builder). Primary surface using the new brand is the SpaceXAI console (developer API dashboard). The rebrand formalizes xAI's deepening integration with SpaceX (Colossus data centers, Grok in Tesla vehicles, joint employee programs). Entity slug remains xai — slug changes break wikilinks. ⚠️ Note: already referenced as "SpaceXAI console" in the July 8/9 Grok 4.5 entry above. (source) (@SpaceXAI)

• 2026-07-06: Grok 4.5 web UI feature flag detected — consumer launch imminent — TestingCatalog discovered on July 6 that the Grok web client contains a feature flag with the account upgrade copy "Unlock the full power of Chat with Grok 4.5" — the consumer launch scaffolding is in place. The flag is dormant as of July 8, meaning only the flip remains before broader rollout. Based on the Grok 4.3 precedent, initial access is expected for paid SuperGrok Heavy ($300/month) subscribers. No official benchmarks, pricing details, or context window specs have been published. → Grok 4.5 (source)

• 2026-07-01: Voice Agent Builder — no-code voice agent platform in beta — xAI launched Voice Agent Builder, a no-code platform to build production voice agents on Grok Voice in under 2 minutes. Bundles telephony, knowledge retrieval, tool-calling, guardrails, MCPs, observability, and call recording in one speech-to-speech interface — replacing the fragmented STT+LLM+TTS three-vendor stack. Integrations: Gmail, Google Calendar, Outlook, Linear, Notion, OneDrive, Google Drive, custom APIs; human transfer support. Pricing: $0.05/min agent audio + $0.01/min telephony — undercutting ElevenLabs and Vapi. 80+ voices, 25+ languages, brand voice cloning from ~2 min of audio. Powered by Grok Voice Think Fast 1.0 (67.3% on internal tau-voice Bench; self-administered). Why it matters: MCP support baked in from day one makes this the first no-code voice agent tool with an open integration standard — any MCP server can be connected without writing glue code. (source) (xAI)

• 2026-06-30: Grok 5 misses Q2 (June 30) deadline — still in training on Colossus 2 as of June 30. Q2 window now closed. July is the next stated target. Monthly release cadence (Grok 4.5 → Grok 5) remains xAI's stated strategy. (felloai.com)

• 2026-06-28: Grok 4.5 enters private beta at Tesla and SpaceX — Elon Musk announced on X that Grok 4.5 — the enterprise deployment of xAI's V9-class model (1.5T parameters, Cursor-trained) — entered private beta at Tesla and SpaceX. Musk stated "early evals show performance close to, perhaps exceeding Opus" (i.e., Claude Opus 4.8). The 1.5T V9 foundation underlies both the consumer Grok V9-Medium (shipped 2026-06-16) and this enterprise-tuned variant. Musk also announced a monthly model release cadence for the remainder of 2026. Additionally, Musk stated: "My estimate of the probability of Grok 5 achieving AGI is now at 10% and rising" — a speculative CEO claim for the in-training 6T model, not Grok 4.5 itself. → Grok 4.5 (source)

• 2026-06-22: Grok Build /goal — long-running autonomous execution mode — xAI shipped /goal, a new autonomous execution mode inside Grok Build. The mode lets developers hand off a complex implementation task and step back entirely while the agent plans, executes, and self-verifies from start to finish. Implementation: two-model pipeline (Composer 2.5 for planning + Grok Build 0.1 for execution) with three-form verification (code review, webpage inspection, script execution). Users can still inject additional instructions mid-run; pause/resume/clear controls are available. Available now to SuperGrok ($30/month) or X Premium+ subscribers. Why it matters: the plan→execute→verify loop makes Grok Build competitive with Anthropic Managed Agents and OpenAI Codex Enterprise for long-running autonomous coding tasks, now on a consumer subscription model. → Grok Build (source) (xAI)

• 2026-06-03: Grok Imagine Video 1.5 Preview released — image-to-video model launched in preview on the xAI API. Took #1 on the Artificial Analysis Video Arena image-to-video leaderboard. 720p, 15 seconds, native audio sync. Built on the Aurora engine (autoregressive MoE, interleaving text/image/video/audio). $0.08/sec. Simultaneously announced a Vapi partnership — Grok adopted as the default engine for all 12 of Vapi's core voices (targeting 2.5M+ voice agents). → Grok Imagine Video 1.5 (Preview) (source)

• 2026-05-25 (captured 2026-06-01): Grok V9-Medium training declared complete — announced by Elon Musk on X. 1.5T parameters (3× V8's 500B). Fine-tuned on Cursor developer workflow data. RL starting imminently. Release target: ~2 weeks out (mid-June 2026). Grok 5 (6T MoE): training separately on Colossus 2 (Memphis 1GW supercluster). → (source)

• 2026-05-28: Grok Build 0.1 API public beta — the grok-build-0.1 model became available in public beta via the xAI API. Speed: 100+ tokens/sec. Pricing: $1/M input, $2/M output. Previously limited to the SuperGrok Heavy ($300/month) CLI or the Kilo Code IDE, but the API release enables direct use in third-party integrations and enterprise workflows. → Grok Build (source)

• 2026-05-27: Grok Build → Kilo Code integration — SuperGrok and X Premium+ subscribers can use grok-build-0.1 directly in Kilo Code (open-source VS Code agentic coding extension, 500+ models supported). No separate API key required (OAuth authentication). Expands the distribution channel from the terminal-only Grok Build CLI to the entire IDE ecosystem. Includes JetBrains IDE and terminal support. → Grok Build (source)

• 2026-05-22 (ingest): Grok 4.1 Fast + Agent Tools API — best tool-calling model (2M context, −50% hallucination vs Grok 4 Fast). Agent Tools API: web browsing, X post search, code execution, document retrieval — all run on xAI infrastructure (no need for developers to manage API keys/rate limits/sandboxes/pipelines). $0.20/$0.50 per 1M tokens. → Grok 4.1 Fast (xAI) (source)

• 2026-05-18 (ingest): Colossus deal captured

• 2026-05-15: Grok + Nous Research Hermes agent — SuperGrok subscribers can use their Grok account and subscription inside the Nous Research open-source self-improving Hermes agent

• 2026-05-06: Anthropic Colossus 1 compute partnership — exclusive lease of the Memphis data center (220,000+ GPUs, 300 MW). ~$5B/year scale. xAI has completed migration of its training workloads to Colossus 2. Musk: "No one set off my evil detector". SpaceX IPO context: monetizing idle assets as xAI is being absorbed into SpaceX. → Anthropic (source)

• 2026-05-17 (ingest): Grok Connectors + BYO MCP write-up

• 2026-05-15: legacy Grok API model retirement — grok-4, grok-4-fast, grok-4-1-fast, grok-code-fast-1 retired → auto-redirect to grok-4.3

• 2026-05-14: Grok Build early beta released — agentic coding CLI for SuperGrok Heavy subscribers (source)

• 2026-05-06: Grok Connectors + BYO MCP — deep integration with 6 platforms (Google Workspace, SharePoint, Outlook, Notion, GitHub, Linear) + support for connecting custom MCP servers (source)

• 2026-05: Grok Voice API + STT/TTS APIs announced — building out a voice agent ecosystem

• X platform real-time data integration — maintains a differentiator in real-time access to news/markets

Strategic Position

• Frontier competitors: Anthropic, OpenAI, Google DeepMind
• Differentiation: exclusive real-time access to X platform data, ties into the Musk ecosystem (Tesla, SpaceX). The Aurora engine is evolving into a unified multimodal stack across text, image, video, and audio
• 2026: coding agents (Grok Build) + voice AI — competing directly with OpenAI Codex / Claude Code
• Compute: Colossus 1 → leased to Anthropic (300 MW, $5B/yr). Own training concentrated on Colossus 2.
• SpaceXAI rebrand (2026-07-06): xAI adopted the SpaceXAI name and handle (@SpaceXAI on X), formally nesting the AI unit inside the SpaceX brand identity. A restructuring is underway to absorb xAI into SpaceX ahead of an IPO. The Colossus 1 lease should be read in this context.

Related

• Grok Imagine Video 1.5 (Preview) — image-to-video, Video Arena #1
• Grok Build
• Grok 4.1 Fast (xAI) — best tool-calling model, Agent Tools API
• Agents (LLM Agents) (coding agent category)

Conflicting Reports

None