NEWS · MODELS · #431
SpaceXAI launches Grok 4.5, a model optimized for coding and agentic tasks
SpaceXAI says it has released Grok 4.5, its newest model trained alongside Cursor and optimized for coding, agentic workflows, and knowledge work. The company reports training across tens of thousands of NVIDIA GB300 GPUs, claims roughly 2× token efficiency versus leading models, 80 TPS serving speed, and priced the model at $2 per million input tokens and $6 per million output tokens; Grok 4.5 is available in Grok Build, Cursor, and via the SpaceXAI console with limited free usage offered.
KEY POINTS
- SpaceXAI says it has released Grok 4.5, its newest model trained alongside Cursor and optimized for coding, agentic workflows, and knowledge work.
- The company reports training across tens of thousands of NVIDIA GB300 GPUs, claims roughly 2× token efficiency versus leading models, 80 TPS serving speed, and priced the model at $2 per million input tokens and $6 per million output tokens; Grok 4.5 is available in Grok Build, Cursor, and via the SpaceXAI console with limited free usage offered.
- A model that claims substantial gains in coding/agentic performance, serving speed, and cost efficiency could shift developer tooling, competitive pricing, and benchmark standings among advanced LLM providers.
DECISION BRIEF
CONFIDENCE
WHAT CHANGED
SpaceXAI publicly released Grok 4.5, a model the company describes as optimized for coding, agentic workflows, and knowledge work; it reports training Grok 4.5 alongside Cursor on tens of thousands of NVIDIA GB300 GPUs, serving speed of 80 TPS, token-efficiency claims of roughly 2× versus leading models, and a published usage price of $2 per million input tokens and $6 per million output tokens. Grok 4.5 is available in Grok Build, Cursor, the SpaceXAI console, and across web, X, iOS, and Android, with Microsoft 365 add-ins (Word, Excel, PowerPoint, Outlook) noted. SpaceXAI also describes agentic products (Grok Bot) and a subsequent Grok 4.6 that builds on 4.5.
WHY NOW
SpaceXAI has made public claims about a new default, broadly available model (Grok 4.5) with specific performance, cost, and distribution attributes (serving speed, token-efficiency claim, price, multi-surface availability) and is simultaneously promoting agentic features and follow-on models (Grok Bot, Grok 4.6). These published product, performance, and pricing claims create immediate observable points for competitors, integrators, and users to validate or respond to.
WHO IS AFFECTED
Directly affected groups named in the sources: SpaceXAI users of Grok and Cursor (desktop, iOS, Android, web, X), Grok Build users, Microsoft 365 add-in users (Word, Excel, PowerPoint, Outlook), subscribers eligible for Grok Bot beta tiers (SuperGrok, Cursor Pro/Teams tiers), and developers or teams using agentic workflows as described by SpaceXAI.
CONFIRMED
- SpaceXAI announced Grok 4.5 as a model built for coding, agentic tasks, and knowledge work [497]. - SpaceXAI states Grok 4.5 was trained alongside Cursor on datasets spanning coding, science, engineering, and math and that training ran across tens of thousands of NVIDIA GB300 GPUs [497]. - SpaceXAI reports Grok 4.5 serving speed of 80 TPS and claims roughly 2× token efficiency versus leading models for the same tasks [497]. - SpaceXAI published pricing for Grok 4.5 at $2 per million input tokens and $6 per million output tokens and says limited free usage is offered; Grok 4.5 is available in Grok Build, Cursor, and via the SpaceXAI console [summary/event and 497]. - SpaceXAI states Grok 4.5 is live on grok.com, X, iOS, and Android and that Grok integrates with Word, Excel, PowerPoint, and Outlook via add-ins [507]. - SpaceXAI announced Grok Bot (an agentic teammate product), described its beta availability for specified subscription tiers, and that Bots run in the cloud and can operate across apps and websites [530]. - SpaceXAI published a Grok 4.6 announcement stating 4.6 builds on 4.5 and describing additional training and benchmark claims for 4.6 [531].
UNCERTAIN
Missing or unresolved items based on supplied excerpts: - Independent verification: efficiency ("~2× token efficiency"), 80 TPS serving at scale, and training scale ("tens of thousands of NVIDIA GB300 GPUs") are company claims in the excerpts; no third‑party or independently-verified benchmark data is present here [497, 531]. - Comparative context: the excerpts reference comparisons to "leading models" and competitor figures drawn from system cards or leaderboards but do not include the underlying competitor data or independent head-to-head results for Grok 4.5 in these excerpts [497, 531]. - Real-world agent reliability and safety: claims about Grok Bot finishing end-to-end jobs, signing into apps, and handling multi-step work are presented as product descriptions and internal examples; the excerpts do not include external evaluations, failure rates, or security/privacy assessments [530]. - Pricing and billing details: the quoted per-token prices and "limited free usage" are asserted, but excerpts do not show billing terms, limits, or enterprise contract details [497]. - Scope of availability: Grok Bot beta availability lists specific subscription tiers and an enterprise waitlist, but the excerpts do not provide rollout timelines or geographic/enterprise eligibility details [530].
WHAT TO WATCH
Concrete observable signals to monitor (and where to look in the supplied excerpts): - Independent benchmarks or third‑party head-to-head tests comparing Grok 4.5 token efficiency and task performance to other models (not in these excerpts; related claims appear in [497, 531]). - Independent measurements of serving throughput (80 TPS claim) under representative loads (claim in [497]). - Reports or published audits about the scale and infrastructure of the stated training run ("tens of thousands of NVIDIA GB300 GPUs") and data curation methods (claims in [497]). - User and developer reports about Grok 4.5 behavior in coding and multi-step agentic tasks, and real-world examples of Grok Bot completing end-to-end jobs or failing to do so (product descriptions and examples in [497, 530]). - Official SpaceXAI follow-ups clarifying pricing, billing terms, free‑usage limits, and enterprise availability for Grok 4.5 and Grok Bot (pricing and availability notes in [497, 530]). - Any third‑party security, privacy, or reliability evaluations of Grok Bot’s ability to sign into and act within external apps and services (capabilities described in [530]). - Further SpaceXAI technical documentation or system cards giving benchmark data and methodology, and any responses from competitor developers to the comparative claims (comparative language appears in [497, 531]).
WHY IT MATTERS
A model that claims substantial gains in coding/agentic performance, serving speed, and cost efficiency could shift developer tooling, competitive pricing, and benchmark standings among advanced LLM providers.
EVIDENCE MAP
4Editorial claims linked to specific sources, with support, contradiction and context shown separately.
SpaceXAI released Grok 4.5 and describes it as its smartest model built for coding, agentic tasks, and knowledge work; it was trained alongside Cursor.
SUPPORTEDChecked 2026-09-25 · 2 supporting
507 excerpt: 'Grok 4.5, our most intelligent model yet, is now on grok.com, X, iOS, and Android' and 'Grok 4.5 is our most capable model yet.'
SUPPORTS xAIMEDIA · 2026-09-25497 excerpt: 'Grok 4.5... SpaceXAI's smartest model built to excel at coding, agentic tasks, and knowledge work' and 'It’s our strongest model ever and was trained alongside Cursor.'
The announcement states Grok 4.5 was trained across tens of thousands of NVIDIA GB300 GPUs.
SUPPORTEDChecked 2026-09-25 · 1 supporting
SpaceXAI claims Grok 4.5 achieves roughly 2× token efficiency versus leading models.
SUPPORTEDChecked 2026-09-25 · 1 supporting
The announcement states Grok 4.5 is served at 80 TPS (transactions per second).
SUPPORTEDChecked 2026-09-25 · 1 supporting
SOURCES & TIMELINE
1Grok 4.6 builds on Grok 4.5 with a particular focus on long-running agents and more ambitious interactive and visual work. Today we are releasing Grok 4.6 . Grok 4.6 builds on Grok 4.5 with a particular focus on long-running agents and more ambitious interactive and visual work. It stays with complex tasks across many steps, whether researching a topic, analyzing information, working across a codebase, or turning an …
Grok Bot is your team of always-on agents. They have their own computer, work inside tools and apps like you do, and keep working 24/7. Today we’re launching Grok Bot, AI teammates you can give real work to. Bots have their own computer. They sign into the tools you already use and work across apps, inboxes, and more. They finish jobs end to end, and only come back when something needs your approval. You can message …
Grok 4.5, our most intelligent model yet, is now on grok.com, X, iOS, and Android. Grok 4.5 now powers Grok wherever you use it , live today on grok.com , on X, and in the Grok apps for iOS and Android. Grok 4.5 is our most capable model yet. In everyday use, it follows what you're actually asking, keeps track of longer conversations, and gives clearer, more reliable answers. Ask it to explain something unfamiliar, w…
Grok 4.5 is SpaceXAI's smartest model built for coding, agentic tasks, and knowledge work. Today, we're launching Grok 4.5 , SpaceXAI's smartest model built to excel at coding, agentic tasks, and knowledge work. It's our strongest model ever and was trained alongside Cursor . Grok 4.5 was trained on datasets spanning knowledge in coding, science, engineering, and math. With both intelligent and efficient reasoning, G…