SpaceXAI launched Grok 4.5 on July 8, 2026, and independent benchmarkers spent the following 24 hours stress-testing every claim. The verdict is more nuanced than either the launch hype or the skeptics suggest: a genuine near-frontier model at a price point that changes the economics of agentic coding work \u2014 but built on the same architectural tradeoff that explains both why it runs so cheaply and why it hallucinates more confidently than its predecessor.<\/p>\n
The pricing is the starting point. At $2 per million input tokens and $6 per million output tokens, Grok 4.5 undercuts Anthropic’s Claude Opus 4.8 by more than 60% on headline rates. Per completed agentic task, the gap widens further: independent benchmarking firm Artificial Analysis<\/a> measured Grok 4.5 running inside Grok Build at $2.49 per task against $5.07 for GPT-5.5 in OpenAI’s Codex and $11.80 for Fable 5 in Claude Code. For teams running thousands of agentic sessions monthly, that arithmetic compresses costs in ways that matter far more than the per-token comparison.<\/p>\n The model landed at fourth place on Artificial Analysis’s Intelligence Index \u2014 a composite of nine real-world evaluations including knowledge work, banking, and agentic coding \u2014 behind Claude Fable 5, GPT-5.5, and Claude Opus 4.8. That ranking represents a 16-point jump from its predecessor Grok 4.3, the largest single-generation leap any lab has posted on that index, vaulting SpaceXAI past Google’s Gemini line and every open-weight model on the board. On the firm’s Coding Agent Index<\/a>, Grok 4.5 running in Grok Build scored 76 \u2014 tied with GPT-5.5 in Codex and just one point below Fable 5 in Claude Code.<\/p>\n Elon Musk’s own framing was precise: “roughly comparable to Opus 4.7, but much faster.” The independent numbers support that read. Grok 4.5 trails Opus 4.8 on most raw benchmarks \u2014 it is not the strongest model available \u2014 but it matches or approaches that tier on speed and cost in ways that make those trail margins less relevant for high-volume, cost-sensitive workloads.<\/p>\n Why Grok 4.5 Runs Cheaper Than Its Benchmarks Suggest<\/p>\n The price and speed advantages are not temporary subsidies. They are a direct consequence of how Grok 4.5 is built.<\/p>\n Grok 4.5 uses a mixture-of-experts (MoE) architecture on SpaceXAI’s V9 foundation. Rather than activating the entire parameter set for every token \u2014 as a dense model does \u2014 a sparse MoE model routes each token to only a small subset of specialized “expert” subnetworks, selected by a learned gating function. This is why a model with roughly 1.5 trillion total parameters can run at 80 tokens per second and resolve coding tasks using an average of 15,954 output tokens, against Opus 4.8’s average of 67,020 tokens for the same benchmark, according to the SpaceXAI Grok 4.5 launch page<\/a>. The 4.2x token efficiency advantage flows from the architecture, not from underinvestment in capability.<\/p>\n The training methodology added a second dimension. SpaceXAI trained Grok 4.5 across tens of thousands of NVIDIA GB300 GPUs using a highly asynchronous reinforcement learning pipeline covering hundreds of thousands of tasks centered on multi-step software engineering. Agentic rollouts ran for many hours while learning continued in parallel, a setup designed to teach the model how to complete real engineering sequences, not just single-turn prompts. Cursor, acquired by SpaceX for $60 billion in June 2026, contributed trillions of tokens of actual developer-session data: real debugging traces, multi-file edits, and error-recovery sequences from production engineering workflows. That is the specific signal that makes Grok 4.5 competitive on agentic coding despite not holding the top raw-capability position.<\/p>\n The context window, however, shrank. Grok 4.5 ships with 500,000 tokens \u2014 down from Grok 4.3’s 1 million-token window. SpaceXAI offered no official explanation. A plausible inference is that serving a model three times the parameter size of its predecessor at competitive speeds requires trading away working memory. Artificial Analysis flagged the reduction without receiving a reason; pricing also doubles for requests above 200,000 tokens, which is relevant for teams using long-context inputs regularly.<\/p>\n What the Benchmarks Show \u2014 and What They Don’t<\/p>\n Grok 4.5’s benchmark profile is genuinely strong in some areas and modest in others. On SWE Marathon, which measures pass-at-one resolution rates on software engineering tasks, it took the top spot at 29% against Opus 4.8’s 26% and Fable 5’s 24%. On Terminal Bench 2.1, which tests complex command-line tasks, it scored 83.3% \u2014 a fraction of a point behind GPT-5.5 at 83.4% and Fable 5 at 84.3%. On AutomationBench-AA, an independent evaluation of agentic SaaS workflow completion, it was the first model to clear the 50% mark, posting 51.4% clean task completions against Fable 5’s 48.6%.<\/p>\n The picture is different on evaluations that use a neutral harness. On DeepSWE 1.1 \u2014 run by Datacurve with a standardized agent rather than each lab’s own tools \u2014 Grok 4.5 scored 53%, behind GPT-5.5 at 67% and Fable 5 at 70%. On SWE-Bench Pro, it reached 64.7%, trailing Opus 4.8’s 69.2% and Fable 5’s 80.4%. SpaceXAI’s own launch page featured the benchmarks where Grok 4.5 does best and omitted its scores on reasoning evaluations such as GPQA and AIME. That is standard practice for AI launches, but worth noting when evaluating the full picture.<\/p>\n The hallucination finding is the number that matters most for any use case involving factual claims. Artificial Analysis measures hallucination via its AA-Omniscience Index, which rewards correct answers and penalizes confident incorrect ones. Grok 4.5’s knowledge accuracy improved substantially \u2014 from 35% to 52% over its predecessor \u2014 but its hallucination rate more than doubled, rising from 25% to 54%. The independent assessment framed this explicitly as “a known pattern \u2014 larger models know more but state it with more unwarranted confidence.” For legal, financial, research, or client-facing work, that combination means the model is more likely to be right, but when it is wrong, it is now more likely to sound certain about it. Web search integration \u2014 available as a billed server-side tool at $5 per call \u2014 reduces hallucination rates substantially when enabled, but does not eliminate them.<\/p>\n The MoE architecture and the hallucination pattern are connected. Sparse expert routing improves knowledge breadth as the model scales, but expert specialization does not automatically improve factual calibration. A router that confidently assigns a token to an expert can be confident about the wrong expert. This is the same engineering tradeoff that produced the token efficiency gain: more parameters, more coverage, more confident activation \u2014 without a proportional increase in the model’s ability to recognize when it is outside its reliable knowledge.<\/p>\n Pricing Against the Competitive Field<\/p>\n The competitive context clarifies the disruption case. Opus 4.8 is priced at $5 per million input tokens and $25 per million output tokens. Fable 5, Anthropic’s highest-tier model, runs $10 input and $50 output. OpenAI’s GPT-5.6, which began its public rollout today, is positioned between those tiers.<\/p>\n Musk’s strategy follows a pattern that Chinese vendors have used effectively: achieve close-enough performance on the benchmarks that matter most for the use case, then win decisively on per-task economics. For high-volume agentic workloads \u2014 where a model may generate millions of tokens per day \u2014 the gap between $2.49 and $11.80 per task is not marginal. It is transformative, and it creates pressure on incumbent labs to either match the price or justify the premium.<\/p>\n