SpaceXAI’s Grok 4.5 Goes Public at $2 Per Million Input Tokens

Elon Musk confirmed on Wednesday that SpaceXAI’s Grok 4.5 goes public Thursday, July 9, in the first flagship model release from SpaceXAI since SpaceX absorbed xAI in February. SpaceXAI priced the model at $2 per million input tokens and $6 per million output tokens, well below Anthropic’s Claude Opus 4.8 and OpenAI’s premium tiers. The release lands alongside two rival launches that turn Thursday into the most crowded model-ship day of the year.

Anthropic’s Claude Fable 5 returns to global users this week after an 18-day export-control blackout, with access restored following a June 30 decision from the US Commerce Department. OpenAI is staging its GPT-5.6 family for a narrow set of trusted partners under a new federal review regime. Three frontier labs are answering the same question three different ways on the same day. The price gap between the leaders is the widest it has been in this cycle.

Grok 4.5 Goes Public at Half the Price

Musk announced the July 9 public release “based on strong positive feedback from customers in our beta test program.” He added that the model is “an Opus-class model, but faster, more token-efficient and lower cost.” Thursday’s release is the first flagship model to ship under the SpaceXAI name since the February merger.

SpaceXAI’s official launch page lists pricing at $2 per million input tokens and $6 per million output tokens, with cached input at $0.50 per million. Anthropic’s Claude Opus 4.8 runs $5 and $25 on the same line items. OpenAI’s cheapest GPT-5.6 tier, Luna, sits at $1 and $6 per million tokens, placing Grok 4.5 between Anthropic’s premium and OpenAI’s budget. Artificial Analysis measured Grok 4.5 at $0.49 per completed task and called it “nearly 90% cheaper than the models ahead of it” on its leaderboard.

Musk said the model has been in private beta at SpaceX and Tesla since June 28. Grok 4.5 is live today in Grok Build, in Cursor across all paid plans, and via the SpaceXAI console, but the European Union gets it in mid-July rather than at launch.

• $2.00 per 1M input tokens
• $6.00 per 1M output tokens
• $0.50 per 1M cached input tokens
• $0.49 cost per completed task (Artificial Analysis)
• No. 4 on Artificial Analysis GDPval-AA v2, Elo 1543

How Grok 4.5 Was Built

Grok 4.5 sits on the V9 foundation model, which Musk has framed at 1.5 trillion parameters. The technical post on V9 and Colossus training described the architecture as “roughly three times the size of the v8-small architecture behind Grok 4.3” and as “trained on Blackwell GPUs at the Colossus cluster in Memphis.” SpaceXAI’s official launch page does not publish a parameter count, and Artificial Analysis notes the firm has not disclosed model size. The 1.5T figure remains Musk’s self-reported framing.

The Cursor deal closed weeks after the SpaceX IPO. SpaceX agreed in April to buy the AI coding startup at $60 billion, then closed the all-stock acquisition in June, with Cursor posting on Wednesday: “We’ve partnered with SpaceXAI to train Grok 4.5. It’s our most powerful model yet and the first we’ve built for more than software engineering.” The launch lands against an unresolved xAI safety lawsuit filed before SpaceX’s record IPO.

What ‘Opus-Class’ Actually Hides

The independent numbers are more sober than Musk’s framing. Artificial Analysis placed Grok 4.5 fourth on its agentic knowledge-work index with an Elo of 1543, between Opus 4.8 at 1600 and GLM-5.2 at 1513. The jump over Grok 4.3 is real, but the absolute position is not top of the leaderboard.

SpaceXAI’s own published benchmark chart shows mixed results. Grok 4.5 beats Opus 4.8 on SpaceXAI’s DeepSWE 1.0 provider-harness run with 62.0% to 55.75%, but the neutral-harness DeepSWE 1.1 from DataCurve flips the result with Opus 4.8 at 59% to Grok 4.5 at 53%. On Terminal Bench 2.1, Fable 5 posts 84.3%, GPT-5.5 83.4%, and Grok 4.5 83.3%, with Opus 4.8 at 78.9%.

The disruption case lives in dollar-and-token math rather than the leaderboard. Artificial Analysis reports Grok 4.5 uses around 14,000 output tokens per Intelligence Index task, more than 60% fewer than Opus 4.8. SpaceXAI’s own SWE Bench Pro example claims 15,954 output tokens per task versus 67,020 for Opus 4.8, a 4.2x gap. The launch page frames this as “roughly 2x the token efficiency of comparable leading models,” and the gap holds whether or not the model tops any single benchmark.

Musk added a second comparison. “Our internal assessment is that Grok 4.5 is roughly comparable to Opus 4.7, but much faster,” he wrote. The bar he set is Opus 4.7 rather than the current top-of-line Opus 4.8, and the gap matters to any buyer choosing on capability alone.

OpenAI Lands GPT-5.6 on Government Terms

OpenAI previewed GPT-5.6 Sol, Terra, and Luna on June 26 and is making the models available initially to a narrow set of approximately 20 organizations. A general release is planned for “the coming weeks.” The staggered rollout directly follows Executive Order 14409, the June 2 directive asking federal agencies to assess frontier AI capabilities before wide release. OpenAI says it previewed its plans and the models’ capabilities with the government ahead of the launch.

The three tiers split the workload by use case and price. Sol is the flagship at $5 per million input tokens and $30 per million output tokens, the same headline price as GPT-5.5 but with new reasoning tools, while Terra is the mid-tier at $2.50 and $15 and Luna is the budget tier at $1 and $6.

All three GPT-5.6 models sit at OpenAI’s “High” risk level for cyber and biological/chemical capability under the new system card, even Luna. Sol reached 91.91% on TerminalBench 2.1 in ultra thinking mode, the highest score OpenAI has published, with max mode at 88.76% and GPT-5.5 at 83.4%, while Anthropic’s Mythos 5 sat at 88% on the same test. The smaller-tier coding gain is large enough that even Luna is competitive with the prior generation’s flagship. OpenAI’s broader push into AI security, including its Daybreak cybersecurity initiative, is accelerating in parallel.

Anthropic Pulls Fable 5 Back From the Brink

Anthropic disabled access to Claude Fable 5 and Claude Mythos 5 in mid-June to comply with an export control directive from the US Commerce Department. The June 12 directive cited “national security authorities” and instructed Anthropic to suspend all access “by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees.” Fable 5 had been public for only three days at that point.

Anthropic co-founder Tom Brown led the negotiations with the Trump administration, replacing CEO Dario Amodei in that role. Amodei had drawn the administration’s ire for his vocal support of Kamala Harris in 2024 and his outspoken AI safety positions. Commerce Secretary Howard Lutnick cleared Mythos 5 for a selected group of companies and federal agencies on June 26. Lutnick told Anthropic in a letter that “appropriate safeguards” were in place for trusted partner access to Fable 5.

Commerce’s lift of the Fable 5 export controls came on June 30, with global access to Claude.AI and Claude Code restarting this week. Fable 5 is included in up to 50% of weekly usage limits for Pro, Max, Team, and selected enterprise plans through July 7, with AWS, Google Cloud, and Microsoft Foundry access to follow. The release lands alongside Anthropic’s separate launch of Claude Sonnet 5 on June 30, which closes the gap with Opus on coding at half the price.

Government scrutiny now hangs over all three labs. OpenAI committed to a “small group of trusted partners” preview in response to the Trump administration’s June 2 executive order, dedicating 700,000 A100e GPU hours to automated red-teaming of GPT-5.6. SpaceXAI trained Grok 4.5 on the same Colossus capacity it leases to Anthropic and Google, a structural overlap that will draw closer review as the federal framework takes shape. Each of the three labs now operates inside a different government-relations loop on the same week.

Anthropic’s Glasswing cybersecurity program continues to expand access to Mythos 5 for domestic and international partners. Per Ramp spending data cited by CNBC, Cursor’s share of the AI coding category fell from 41% in June 2025 to about 26% by May 2026, while Anthropic came to control roughly half the market. The 18-day export standoff cost Anthropic weeks of momentum in a category where Claude now sets the price benchmark.

• June 9, 2026 – Anthropic releases Claude Fable 5 publicly
• June 12 – US Commerce Department issues the export control directive
• Mid-June – Anthropic disables global access to Fable 5 and Mythos 5
• June 26 – Commerce Secretary Lutnick clears Mythos 5 for select partners
• June 30 – Commerce Department lifts all export controls
• Early July – Global access to Fable 5 resumes on Claude.AI and Claude Code, with token limits through July 7

Who Now Owns the Coding-AI Stack

SpaceXAI’s launch crystallizes what Musk’s dealmaking was building toward. The February xAI-SpaceX merger valued the combined company at $1.25 trillion, with xAI at $250 billion. The June IPO raised $75 billion at a fixed $135 share price, valuing SpaceX at roughly $1.77 trillion. Musk holds 82.4% of voting power through Class B shares. SpaceX now controls nearly the entire stack: Colossus for training compute, the Cursor distribution channel, the Grok model line, and captive demand from Tesla and SpaceX engineering.

Anthropic and OpenAI cannot match that integration. Both reach developers through third-party tools, and Musk now owns the channel that collected Grok 4.5’s training data. Investor Gavin Baker captured the calculus on social media the morning of the launch: “Pareto dominant for coding by the numbers. We will see on the all-important vibes.”

Musk’s framing is deliberate. Arriving late and battered by Grok’s mid-2025 safety failures and a full xAI co-founder turnover, he has chosen cost-per-task as the competitive axis. The release lands as Anthropic recovers from its export-control blackout and OpenAI waits out the federal review clock. Three labs, three different competitive theories, one crowded day, with the federal framework set to rewrite the rules again before any of them consolidates a lead.

Frequently Asked Questions

When does Grok 4.5 go public?

SpaceXAI released Grok 4.5 publicly on Thursday, July 9, 2026, after Musk confirmed the rollout on social media the day before. The model had been in private beta at SpaceX and Tesla since June 28. EU users will see availability in mid-July rather than at launch.

How much does Grok 4.5 cost?

The official API price is $2.00 per million input tokens, $0.50 per million cached input tokens, and $6.00 per million output tokens. Requests above 200,000 tokens use separate higher-context rates, and the model is free for a limited time inside Grok Build and Cursor.

How fast is Grok 4.5?

SpaceXAI serves Grok 4.5 at 80 tokens per second on the fast-model tier. The model supports text and image inputs with text output, function calling, web search, X search, and code execution. SpaceXAI has not published a context-window figure on the launch page, and Artificial Analysis notes the firm has not disclosed model size.

How does Grok 4.5 compare to Claude Opus 4.8 and GPT-5.6?

On Artificial Analysis’s GDPval-AA v2 agentic index, Grok 4.5 landed at an Elo of 1543, with Opus 4.8 ahead at 1600 and GLM-5.2 behind at 1513. Grok 4.5 is faster and significantly cheaper per task than Opus 4.8 on the metrics measured, but the model does not lead on raw capability.

Is this the first SpaceXAI flagship since the xAI merger?

This is the first model release from SpaceXAI since the combined xAI-SpaceX structure came into being in February 2026, and pricing was disclosed in the same window as SpaceX’s record-setting IPO.