Grok 4.5 is cheap enough to win on cost and wrong enough to lose on trust
SpaceXAI shipped Grok 4.5 on July 8, and within 24 hours the picture was clearer than the launch hype or the skeptics allowed. The model is a genuine near-frontier release that runs agentic coding tasks at roughly a fifth the cost of Claude Fable 5. It also hallucinates more than half the time.
The pricing is where the story starts. Grok 4.5 costs $2 per million input tokens and $6 per million output tokens. Claude Opus 4.8 sits at $5/$25. Fable 5, Anthropic’s top-tier model, runs $10/$50. On per-task economics measured by Artificial Analysis, Grok 4.5 in Grok Build cost $2.49 per coding task. Fable 5 in Claude Code cost $11.80. GPT-5.5 in Codex cost $5.07. For teams running thousands of agentic sessions per month, that gap is not marginal. It changes which model you default to.
The token efficiency claim is the operational detail behind that pricing. SpaceXAI reports that Grok 4.5 resolves SWE-Bench Pro tasks using an average of 15,954 output tokens. Opus 4.8 used 67,020 for the same benchmark. That is a 4.2x gap. It flows from the architecture. Grok 4.5 is a mixture-of-experts model with roughly 1.5 trillion total parameters, three times larger than its predecessor, running at 80 tokens per second. A sparse MoE routes each token to a small subset of specialized experts instead of activating the full parameter set. More parameters, more coverage, fewer tokens per task.
On the Artificial Analysis Intelligence Index, Grok 4.5 scored 54. That puts it fourth out of 168 models, behind Claude Fable 5, GPT-5.5, and Claude Opus 4.8. The 16-point jump from Grok 4.3 is the largest single-generation leap any lab has posted on that index. SpaceXAI vaulted past every Google Gemini model and every open-weight model on the board. On the Coding Agent Index, Grok 4.5 scored 76 in Grok Build, tied with GPT-5.5 in Codex and one point behind Fable 5 in Claude Code.
There are specific wins. On SWE Marathon, which tests pass-at-one resolution on extended software engineering tasks, Grok 4.5 took the top spot at 29% against Opus 4.8’s 26% and Fable 5’s 24%. On AutomationBench-AA, an independent evaluation of agentic SaaS workflow completion, it became the first model to clear 50%, posting 51.4% against Fable 5’s 48.6%. Snorkel AI ran approximately 2,000 GDPval+ tasks authored by domain professionals who produce real workplace deliverables. Grok 4.5 led at 29% mean pass rate, ahead of GPT-5.5 at 22% and Opus 4.8 at 21%. The lead was concentrated in domains requiring professional judgment: legal work (40% vs 27-28%), education (58% vs 35-42%), healthcare (35% vs 23-25%).
The picture shifts on neutral ground. On DeepSWE 1.1, which uses a standardized agent harness rather than each lab’s own tools, Grok 4.5 scored 53%. GPT-5.5 hit 67%. Fable 5 reached 70%. The gap between Grok 4.5’s performance in its own Grok Build harness and in a neutral harness is the standard vendor-harness advantage, amplified here because Grok 4.5 is the first model co-trained with Cursor since SpaceX acquired the coding platform for $60 billion in June. Cursor contributed trillions of tokens of real developer-session data: debugging traces, multi-file edits, error-recovery sequences. That signal makes the model genuinely strong on agentic coding workflows. It also means benchmarks run inside Cursor-derived tooling favor the model by construction.
Cursor disclosed that an earlier snapshot of its own codebase accidentally entered Grok 4.5’s training data, inflating its score on at least one internal benchmark. The data has been removed from future runs. Cursor stated this publicly, which is the right call, but it means anyone treating benchmark charts as gospel should apply a discount.
The hallucination number is the one that should change deployment decisions. Artificial Analysis measures hallucination through its AA-Omniscience Index, which rewards correct answers and penalizes confident incorrect ones. Grok 4.5’s accuracy rose from 35% to 52% over Grok 4.3. Its hallucination rate more than doubled, from 25% to 54%. The model knows more. When it is wrong, it is more likely to sound certain about it.
The MoE architecture and the hallucination pattern are connected. Sparse expert routing improves knowledge breadth as the model scales. Expert specialization does not automatically improve factual calibration. A router that confidently assigns a token to the wrong expert produces confident wrongness. This is the same engineering tradeoff behind the token efficiency gain. More parameters, more coverage, more confident activation, without a proportional increase in the model’s ability to recognize when it is outside its reliable knowledge.
For agentic tool-calling workflows where outputs are validated before they are acted on, Grok 4.5 is the right model at the right price. For legal research, financial analysis, medical documentation, or any workflow where the factual accuracy of the output itself is the quality metric, a 54% hallucination rate is a disqualifying defect. Web search integration at $5 per call reduces hallucination substantially when enabled, but does not eliminate it.
Grok 4.5 ships with a 500,000-token context window, down from Grok 4.3’s 1 million tokens. SpaceXAI offered no explanation. Pricing doubles for requests above 200,000 tokens. The reduction is plausibly a serving tradeoff: a model three times larger, running at competitive speed, trading away working memory. Cached input is $0.50 per million tokens, a 75% discount that benefits agentic loops reusing the same context repeatedly.
The model is not available in the EU at launch. SpaceXAI expects European access later in July.
The loudest discussion on Hacker News was not about benchmarks. It was about trust. Elon Musk controls SpaceXAI and has documented instances of editing Grok system prompts to steer outputs on political topics. The concern is whether a model from a company with active political interests can be used reliably for professional work where output independence matters. Counter-arguments in the same thread noted that Grok has often tested as more politically restrained than GPT or Gemini in practice. The evidence is mixed enough that the question stays open.
Musk’s own framing was measured. He called Grok 4.5 “roughly comparable to Opus 4.7, but much faster.” The independent numbers support that read. Grok 4.5 is not the strongest model available. It does not need to be. It matches close enough on capability while winning decisively on cost per completed task. That is the same playbook Chinese vendors used to compress margins in the API market. Close enough on benchmarks, aggressive on price, capture the volume segment where frontier accuracy is not the decision variable.