Claude Sonnet 4.6 (Non-reasoning, Low Effort) vs GPT-5 (high): The Ultimate Performance & Pricing Comparison

GPT-5 (high) is the stronger default for developers because it combines a slightly higher Intelligence Index with published coding and math evidence at a lower blended token price.

The comparison is less settled than a simple leaderboard suggests. Claude Sonnet 4.6 (Non-reasoning, Low Effort) has measured output speed, while GPT-5 (high) has no corresponding speed value in the supplied data. Their latency is tied at 0.3 seconds, so interactive responsiveness cannot be decided from latency alone.

The model labels also describe different kinds of configuration. OpenAI documents “high” as a reasoning_effort setting on gpt-5, not as a separate API model. Anthropic’s official material does not clearly establish whether “Non-reasoning, Low Effort” is an independent API model or a benchmark configuration. See GPT-5 for developers and Anthropic’s Claude models overview.

Data provided by https://artificialanalysis.ai/ supplies the comparative performance, latency, speed, and pricing snapshot used here.

GPT-5 (high) offers the more defensible general-purpose choice because its evidence covers intelligence, coding, and math, while Claude Sonnet 4.6 has a narrower documented comparison.

The evidence does not prove that GPT-5 will win every coding task. Its coding score is available, but Claude has no matching value in the supplied data. That is an evidence asymmetry, not a demonstrated failure by Claude.

Anthropic still lists Claude Sonnet 4.6 in its official pricing system, while OpenAI’s model documentation lists gpt-5 as callable but describes the fixed snapshot as deprecated and recommends GPT-5.6. These statuses create different migration concerns. Read Claude pricing alongside the GPT-5 model documentation before committing to a long-lived integration.

GPT-5 (high) has the stronger measured capability profile, but Claude Sonnet 4.6 may feel faster during long streamed responses.

GPT-5 leads the supplied Intelligence Index at 34.7 versus 34.3 for Claude. The difference is small enough that it should not determine a purchase decision by itself. A narrow overall gap can hide large task-specific differences, especially because GPT-5 has supplied coding and math scores while Claude has no corresponding values.

The clearest practical signal favors GPT-5 for work where developers need explicit evidence around code or mathematical reasoning. Its supplied coding score is 37.8, and its supplied math score is 94.3. Those results do not establish production accuracy for a specific repository, test suite, or domain. They do indicate that GPT-5 has a documented evaluation footprint that Claude lacks in this dataset.

Claude’s measurable median output speed is 56.53 tokens per second. GPT-5 has no speed value in the supplied snapshot, so declaring a speed winner would exceed the evidence. Both models show 0.3 seconds of latency. That tie says little about total completion time, because generation duration, reasoning work, streaming behavior, retries, and tool calls can dominate the user experience.

The configuration difference matters. GPT-5 high is explicitly represented as high reasoning effort in OpenAI’s material. Anthropic does not clearly document the tested Claude configuration as a standalone API model or publish its dedicated effort semantics. Developers should therefore reproduce representative tasks with the exact production parameters before treating the chart as a final ranking.

Community evidence adds a caution for GPT-5 rather than a verdict. A Reddit user reported that GPT-5 was useful for small bug fixes but could produce simplified application interfaces or incorrect changes in complex existing codebases. The account was not a controlled benchmark. See Tried GPT-5 Here Are My First Impressions. No equally reliable community evidence was found for this exact Claude configuration.

GPT-5 (high) is the cheaper choice for the supplied token mix, but workload shape and cache behavior can change the operational advantage.

The supplied blended price is $3.4375 per 1M tokens for GPT-5 and $6 for Claude Sonnet 4.6. GPT-5 also has the lower input price, at $1.25 versus $3, and the lower output price, at $10 versus $15. This makes GPT-5 the safer starting point for applications with substantial prompt volume, generated code, or repeated batch processing.

The price gap matters most when model calls are frequent or outputs are large. A cheaper input rate can reduce the cost of repository context, documentation, and long instructions. A cheaper output rate can reduce the cost of generated patches, explanations, structured results, and agent traces. The supplied data does not provide traffic proportions beyond the blended comparison, so a team should validate the result against its own input-to-output distribution.

Claude has documented prompt caching prices of $3.75 per MTok for a 5-minute write, $6 per MTok for a 1-hour write, and $0.30 per MTok for cache hits and refreshes. Those prices create a meaningful exception for workloads that repeatedly reuse large stable prompts. The supplied GPT-5 material lists cached input at $0.125 per 1M tokens, but the two pricing descriptions use different units and cache terminology. Developers should confirm billing semantics before comparing cache-heavy workloads directly.

Tokenizer behavior is another hidden cost variable. Anthropic says Claude Sonnet 4.6 uses its previous-generation tokenizer, while later Claude models use a newer tokenizer that usually produces about 30% more tokens for the same text. That statement does not make Claude cheaper or more expensive by itself. It means token counts from another Claude generation should not be copied into a Sonnet 4.6 budget.

Regional inference can also affect Claude pricing. Anthropic documents a 1.1 multiplier for inference_geo: "us", with global as the default standard price. The chart’s $6 Claude blended value should therefore be treated as a baseline rather than a universal invoice estimate. Pricing details are documented in Claude pricing and GPT-5 model documentation.

GPT-5 (high) is the best first choice for most new developer workflows, while Claude Sonnet 4.6 deserves a focused trial when streaming speed or Anthropic-specific integration matters.

Choose GPT-5 high when the product depends on code generation, mathematical reasoning, structured tool use, or cost control. OpenAI documents function calling, structured outputs, streaming, and custom tools with grammar-constrained output. GPT-5 also has supplied coding and math evidence, which gives engineering teams more concrete material for an initial risk assessment. These capabilities are described in GPT-5 for developers and the GPT-5 model documentation.

Choose Claude Sonnet 4.6 when measured output speed is a priority and the team already operates comfortably within Anthropic’s API and pricing model. Its 56.53 median output tokens per second is the only supplied speed measurement, so it may be attractive for applications where users watch long responses arrive. That advantage remains provisional because GPT-5 speed data is missing.

For an existing-codebase agent, neither model should receive unrestricted write access based on this evidence alone. GPT-5 has a reported community risk of hallucinated or incorrect modifications in complex repositories. Claude has no reliable community evidence for this exact low-effort configuration, which means its risk is unknown rather than lower. Use tests, patch review, constrained tools, and rollback controls for either model.

The largest selection risk is lifecycle ambiguity. OpenAI marks the fixed GPT-5 snapshot as deprecated, although the gpt-5 alias remains documented. Anthropic still lists Sonnet 4.6 for pricing, but the supplied research did not confirm a stable API alias, model ID, or platform availability. Teams that value predictable deployment should verify the exact endpoint, model identifier, retirement policy, and regional availability before launch.

A sensible decision is to start with GPT-5 high for the general path, then run Claude Sonnet 4.6 against the same production-shaped prompts. Reverse that choice only if Claude’s measured speed or task-specific quality produces a clear user benefit that offsets its higher supplied token cost.

GPT-5 (high) is the safer default for evidence-backed developer selection, but the final choice depends on workload-specific testing and lifecycle requirements.