Grok 4.5 ranks second with Grok CLI on FrontierSWE and is free to sample in Grok Build, but an undisclosed usage cap, mixed agent harnesses and newer rivals leave its measured cost efficiency—not a broad capability lead—as the clearest advantage.
Grok 4.5 has given SpaceXAI a credible coding-agent result and a low-cost offer for developers. The stronger conclusion is narrower than the leaderboard headline: the Grok 4.5–Grok CLI system is competitive on demanding, long-running tasks, and its measured economics stand out, but neither the benchmark nor capped free access establishes durable adoption or model supremacy.
SpaceXAI made Grok 4.5 available through its paid API on July 8 for coding, agentic tasks and knowledge work. The release notes price it at $2 per million input tokens and $6 per million output tokens, with low, medium and high reasoning settings.
The company then said in a July 10 Grok account post that anyone with an X or Grok account could try the model in Grok Build. The free route has a token limit, but no published number; the report reproducing the announcement says the company described the allowance only as sufficient for testing. That supports a sampling offer, not a free API or a standing free quota on which a team can plan production work.
Grok Build itself is new. SpaceXAI put the coding agent into beta on May 14 and followed on May 19 with early access to grok-build-0.1, a model specifically developed for agentic coding workflows. Grok 4.5 therefore arrives through a first-party harness that had been public for less than two months.
The headline API price also needs deployment context. Independent testing says cached input costs $0.50 per million tokens but rates double for inputs longer than 200,000 tokens. The same analysis puts the context window at 500,000 tokens, down from 1 million for Grok 4.3. A separate launch account also reports the 500,000-token limit. The maximum context is therefore available, but not at the base input rate throughout.
On FrontierSWE's mean@5 table, a lower average rank is better:
| Rank | Model | Harness | Average rank | Dominance |
|---|---|---|---|---|
| 1 | Claude Fable 5 | Claude Code | 2.47 | 89% |
| 2 | Grok 4.5 | Grok CLI | 4.09 | 78% |
| 3 | Claude Opus 4.8 | Claude Code | 4.82 | 73% |
| 4 | GLM-5.2 | Claude Code | 4.85 | 72% |
| 5 | GPT-5.5 | Codex | 5.21 | 70% |
The benchmark is unusually demanding. Its 17 tasks comprise five implementation problems, three research problems and nine performance problems, including building a PostgreSQL-compatible server on SQLite and improving machine-learning optimizers. Agents may run for as long as 20 hours. Each model-and-harness combination gets five attempts per task, and partial solutions receive scores instead of a binary pass or fail.
That design makes the runner-up result meaningful, but its unit of comparison is the deployed system. Grok 4.5 runs in Grok CLI, Claude models run mainly in Claude Code, and GPT models run in Codex. The table does not isolate the model under a common agent, tool setup or token budget. It shows that one Grok configuration performed well on this task set—not that Grok 4.5 is the second-best coding model for every workload.
The archived benchmark page also shows the hazards of treating a live table as a fixed verdict. Its current standings include Grok 4.5 and newer Claude entries, while its results narrative and qualitative analysis still focus on an earlier field led by GPT-5.4 and Claude Opus 4.6. FrontierSWE says it is continuing to evaluate models, and the ranking can change as systems are added.
Another evaluation reaches a less dramatic capability result and a stronger economic one. It gives Grok 4.5 a score of 54 and fourth place on a broad Intelligence Index, behind Fable 5, GPT-5.5 and Opus 4.8. In Grok Build, Grok 4.5 scores 76 on a Coding Agent Index built from DeepSWE, Terminal-Bench v2 and SWE-Atlas QnA—level with GPT-5.5 in Codex and below Fable 5 in Claude Code.
The cost comparison is more favorable, although the analysis is internally inconsistent on Grok's exact total. Its overview gives $2.59 per task; the detailed comparison twice gives $2.49, against $5.07 for GPT-5.5 and $11.80 for Fable 5. It reports average token use of 1.9 million per task for Grok, compared with 6.2 million and 7.2 million respectively. Either Grok figure is well below the two comparison systems, but these are evaluation-specific totals, not production price guarantees.
Reliability prevents that result from becoming a general trust claim. On the evaluator's separate AA-Omniscience test, Grok 4.5's accuracy rose from 35% to 52% compared with Grok 4.3, while its hallucination rate increased from 25% to 54%. That test does not measure coding, but it shows why one coding-agent rank cannot stand in for performance across tasks.
FrontierSWE's current table compares Grok 4.5 with GPT-5.5, but OpenAI launched the GPT-5.6 family on July 9. In its announcement, OpenAI said Free and Go users could access the mid-tier Terra model in Codex. That means SpaceXAI is not alone in using free coding-agent access to attract users.
OpenAI's own launch table lists scores of 80 for GPT-5.6 Sol and 77.4 for Terra on version 1.1 of the Artificial Analysis Coding Agent Index, compared with 76.4 for GPT-5.5 and 77.2 for Fable 5. Those vendor-published results are not FrontierSWE scores, and the archived Grok analysis does not specify the same index version, so they are not a clean reranking of the FrontierSWE field. They do show how quickly a July 8 comparison against GPT-5.5 became incomplete.
Pricing is similarly segmented rather than uniformly lower. OpenAI lists Luna at $1 per million input tokens and $6 per million output tokens, Terra at $2.50 and $15, and Sol at $5 and $30. Grok 4.5 has the lower base price than Terra and Sol, while Luna has the lower input price and the same output price. Model tier, reasoning effort, token consumption and successful-task rate all affect the bill.
Grok 4.5 is SpaceXAI's first model release since the company went public and acquired Cursor, according to a contemporaneous report. SpaceXAI said the model was trained alongside Cursor, the report notes, and made it available in Grok Build, Cursor and the SpaceXAI console. As of July 8, it was not available in the European Union.
That context redistributes credit for the product result. Developers encounter a model inside an agent harness and, in Cursor, inside an editor tied to the same corporate group. The available evidence does not separate the contribution of the model, Grok CLI, Grok Build or Cursor to either performance or distribution.
The report also says Grok 4.5 used the same compute capacity that SpaceXAI leases to Anthropic and Google. It says growing internal compute demand may force a choice between allocating more capacity to SpaceXAI's models and retaining lease revenue. The free allowance does not resolve that potential tradeoff; its undisclosed cap limits what outsiders can infer about the cost of serving sustained demand.
Three pieces of evidence would determine whether Grok 4.5's low-cost benchmark showing becomes a lasting competitive advantage:
Until then, FrontierSWE supports a focused claim: Grok 4.5 with Grok CLI is a serious long-horizon coding-agent system. The independent cost-per-task result gives developers a reason to test it. Whether that efficiency persists under matched conditions—and whether free testers become paying users—remains the decision-changing evidence still missing.
Zhipu reportedly reached $1 billion in annual recurring revenue in July, roughly four times a March estimate, but the unconfirmed run rate is not annual sales and still sits far ahead of recognized cloud revenue while margins remain thin.
Anthropic’s 20-year lease gives TeraWulf a customer for 401 megawatts of future AI infrastructure, but rent starts only after delivery and the proposed utility deal passes power and grid costs to TeraWulf, leaving financing and project margin unresolved.
OpenAI has put conversations and Projects back in its redesigned ChatGPT desktop app and enabled cloud Work threads to move across devices, correcting the launch's biggest usability failures without merging local Work or Codex histories.
Twenty-nine countries signed an agreement creating WAICO as an independent intergovernmental organization, while China paired the launch with capacity-building offers that are not yet confirmed as WAICO programs.
Meta reportedly plans to put departing AWS compute executive Dave Brown to work on its data-center buildout, adding hyperscale operating experience while leaving any customer-facing cloud business conditional and undefined.
Moonshot AI has made Kimi K3 available through its apps and API, pairing a 2.8-trillion-parameter architecture with early frontier-level results, but the model's open-weight claim cannot be tested until its weights and technical report arrive.
CIA Director John Ratcliffe said US intelligence is consistent with an estimate that Russian recruits last 20 to 30 minutes on Ukraine’s battlefield, but the public trail leads to an unsourced claim about assault troops and does not establish a representative average.
China has paired a five-year AI training offer for developing countries with cooperation centers, a weather-warning rollout and a new 29-country organization. The package gives Beijing a platform for influence, but no budget, selection rules or delivery timetable has been published.
Nebius has arranged its first senior secured facility against an operating GPU deployment and one customer’s cash flows. The deal adds project-level debt to its expansion toolkit, but an unnamed customer and a larger rival financing limit what it proves about the rest of Nebius’s backlog.
Huawei publicly displayed a 16-cabinet Atlas 950 configuration rated at 1 EFLOPS in FP8, providing tangible evidence of its system-scale AI strategy while leaving price, power use and sustained workload performance undisclosed ahead of the full system's planned fourth-quarter release.
UK testing places leading open-weight models four to seven months behind selected closed-model cyber results, yet longer attack chains, U.S. benchmarks and mixed cost comparisons show why that interval is a warning signal rather than a universal capability clock.
Moonshot AI's Kimi K3 added to a global technology selloff with near-frontier performance, but unreleased weights, mixed cost comparisons and a recommended 64-accelerator deployment leave its effect on chip demand unresolved.
Two binding EU decisions require Google to give rival AI services comparable access to 11 Android features and offer eligible search competitors a restricted, anonymized dataset, but phased deadlines, certification, pricing and privacy safeguards leave the competitive effect unproven.
Databricks has signed a term sheet for a Coatue-led financing at a $188 billion valuation, while unidentified sources put the round at $3 billion. The proposed capital would deepen its push into AI governance, data agents and operational databases, but the transaction remains open and the company supplied no new operating figures.
OpenAI says a handful of GPT-5.6 Sol file-deletion reports most commonly involved Full Access without sandboxing or Auto-review. Its own evaluations show a more complicated risk picture, and the company has not yet published an incident rate or evidence that its promised safeguards stop the failure.
Netflix says generative AI workflows were used on roughly 300 titles in 2026, but its only quantified example tied to that disclosure covers 17 minutes and does not establish how much finished material, spending or labor changed across the slate.
A rogue browser extension can trigger Claude for Chrome’s fixed Gmail, Docs, Calendar and business workflows; the attack is constrained by default approvals but can run silently for users who enabled unattended action.
Gemini 3.5 Pro missed its expected June rollout. An anonymous-source account says a late-June data change fell short of Google's coding goals, but Google has confirmed only partner testing—not the reported cause, a new date, or public results and pricing.
Gold Eagle has begun collecting and prioritizing AI-discovered software vulnerabilities, but the voluntary federal clearinghouse has not disclosed results, operating rules or the resources that would turn findings into deployed fixes.
Google is bringing Gemini Omni editing and a reusable face-and-voice avatar to Vids, but the sharper distinction is account-level identity across Vids and Gemini rather than a new category of AI video; Vids already offered Veo generation and customizable avatars, while specialist rivals already sell digital presenters.