GPT-5 (medium) vs Grok 4: The Ultimate Performance & Pricing Comparison

GPT-5 (medium) is the stronger default for cost-sensitive developers, but neither model has enough verified deployment detail for an evidence-free production decision. The data brief gives GPT-5 (medium) a 33.7 Artificial Analysis Intelligence Index, compared with 33.3 for Grok 4. Grok 4 leads the Artificial Analysis Math Index at 92.7, compared with 91.7 for GPT-5 (medium). Both models have a reported latency of 0.3 seconds, while neither has a reported median output speed. GPT-5 (medium) also costs $3.4375 per 1M blended tokens, compared with $6 for Grok 4. The price gap matters for applications that generate frequently, process long conversations, or serve many users. However, the comparison has an important evidence asymmetry. OpenAI's current model documentation does not list gpt-5-medium as a dedicated model and does not provide model-specific limits or benchmark claims at OpenAI Models. OpenAI's current pricing page also does not list gpt-5-medium at OpenAI API Pricing. No reliable, citable official or community source was available for Grok 4 in the research brief. Data provided by https://artificialanalysis.ai/

GPT-5 (medium) wins the broad comparison narrowly because its overall index is higher and its listed prices are lower, while Grok 4 retains a small mathematics advantage. The Intelligence Index difference is only 0.4 points, so it should not be treated as proof that GPT-5 (medium) will outperform Grok 4 on every development workload. Grok 4's Math Index is higher by 1 point, which may matter for workloads dominated by formal reasoning, quantitative verification, or mathematical code generation. The available evidence does not show whether that mathematics difference transfers to production tasks such as debugging, repository changes, tool use, or structured output. The same gap applies to the overall index: the score favors GPT-5 (medium), but the brief does not provide task-level error analysis or methodology details for this comparison. Developers should therefore interpret the benchmark result as directional evidence, not a substitute for a workload-specific test. The commercial conclusion is clearer. GPT-5 (medium) is listed at $1.25 per 1M input tokens and $10 per 1M output tokens, compared with $3 and $15 for Grok 4. That price structure favors GPT-5 (medium) for both prompt-heavy and response-heavy systems. The main unresolved issue is availability. The OpenAI Models page does not confirm whether gpt-5-medium remains directly callable or has a stable alias. The research brief provides no equivalent verified source for Grok 4's availability or API contract.

Grok 4 is the better mathematical benchmark choice, while GPT-5 (medium) has the slightly higher general intelligence score and neither model has a verified speed advantage. Grok 4's Math Index reaches 92.7 against 91.7 for GPT-5 (medium), but the difference is small enough that prompt design, sampling settings, tool integration, and evaluator choice could change the practical outcome. GPT-5 (medium) leads the Intelligence Index at 33.7 against 33.3, which suggests a narrow general benchmark advantage rather than a broad performance guarantee. Developers should map these results to their actual failure costs. A mathematics-focused system may prioritize exact symbolic reasoning and verification. A general coding assistant may care more about instruction following, patch reliability, repository context, and refusal behavior. The supplied evidence does not answer those questions for either model. Both models report 0.3 seconds of latency, so request responsiveness does not separate them in this dataset. Neither model has a median output-tokens-per-second value, leaving streaming throughput unverified. That missing measurement matters for long answers, code generation, and interactive interfaces where time to finish can matter more than initial latency. OpenAI describes broad capabilities for its latest models, including text and image input, text output, multilingual support, and vision, through OpenAI Models, but the page does not confirm that every capability applies to gpt-5-medium. No reliable source in the brief documents Grok 4's corresponding capabilities. The practical performance winner is therefore workload-dependent, with evidence favoring Grok 4 for mathematics and GPT-5 (medium) for the small overall-index lead.

GPT-5 (medium) is the cheaper option across blended, input, and output pricing, but lower token rates do not guarantee a lower total application cost. The blended price is $3.4375 per 1M tokens for GPT-5 (medium), compared with $6 for Grok 4. Input pricing is $1.25 versus $3, and output pricing is $10 versus $15. These differences make GPT-5 (medium) the natural starting point for high-volume workloads, especially when the application sends large prompts or produces substantial generated text. The conclusion can reverse if the cheaper model needs more retries, longer prompts, extra validation calls, or human review to reach the same task success rate. The supplied materials do not report reliability, tool-call accuracy, output quality by task, or failure rates, so they cannot establish cost per successful outcome. Developers should measure those factors before assuming token price equals operating cost. Grok 4 may still be economically rational for a narrow mathematical workload if its higher Math Index reduces correction work. The available evidence does not show whether that advantage is large enough to offset its listed token prices. GPT-5 (medium) also has an unresolved pricing-status problem. The OpenAI API Pricing page does not list gpt-5-medium, so the data brief's price should be treated as comparison data rather than independently confirmed current pricing. Grok 4 has no citable pricing source in the research brief either. Data provided by https://artificialanalysis.ai/

GPT-5 (medium) is the recommended first candidate for developers who prioritize lower listed cost and a slightly higher general benchmark score. Choose GPT-5 (medium) when the application has high token volume, needs lower input and output rates, or benefits from the broadest available evidence in the supplied materials. Before production, confirm the model identifier, availability, context window, output limit, supported parameters, and current price directly with OpenAI. The OpenAI Models page does not provide those details for gpt-5-medium, and the OpenAI API Pricing page does not list it. Choose Grok 4 when mathematics is the dominant requirement and your own evaluation confirms that its 92.7 Math Index advantage improves the target workflow. The benchmark alone is insufficient to justify paying its higher listed rates. A sensible evaluation should use representative prompts, expected outputs, tool calls, long-context cases, and adversarial examples. The research brief supplies no verified Grok 4 documentation, community testing, failure cases, or deployment details. That absence does not prove Grok 4 is weaker or unavailable. It means the procurement and engineering risk is harder to assess from the supplied evidence. The final selection should remain conditional until both models pass the same task-specific acceptance tests. GPT-5 (medium) has the better initial case because its price advantage is clear in the data brief, its overall score is slightly higher, and OpenAI provides at least general model documentation. The evidence is still insufficient for claims about production reliability, context limits, or long-form generation speed.

GPT-5 (medium) requires direct verification of its API contract before a production commitment, because the supplied official pages do not confirm model-specific limits or current listing status. The OpenAI Models page provides general information about the latest OpenAI models, but it does not identify gpt-5-medium as a dedicated entry. The OpenAI API Pricing page also does not list the model. Grok 4 requires an equivalent direct verification process because the research brief contains no reliable, citable official source for its capabilities, pricing, availability, or limitations. The benchmark data is useful for prioritizing tests, not for answering every deployment question. GPT-5 (medium) should be tested first for general-purpose and cost-sensitive workloads. Grok 4 should receive focused testing where mathematical reasoning is central. Both models need measurements for output throughput, failure recovery, structured output accuracy, and tool-use reliability before a final architecture decision.