Claude Sonnet 5 vs GPT-5: Which Model Should Developers Choose?

Claude Sonnet 5 (Non-reasoning, High Effort) is the stronger default for coding-focused applications, while GPT-5 (high) remains attractive for lower blended cost and math-heavy workloads. The Artificial Analysis data gives Claude Sonnet 5 a coding index of 66.4 versus 37.8 for GPT-5, and an intelligence index of 41.7 versus 34.7. Artificial Analysis supplies the comparison data.

The choice is not simply a contest between a newer model and an older model. Claude Sonnet 5 is listed by Anthropic as claude-sonnet-5, and the official documentation describes it as a combination of speed and intelligence. Anthropic’s model overview does not define a separate API model called “Non-reasoning, High Effort.” GPT-5 high is likewise a reasoning configuration, not a separate gpt-5-high API model. OpenAI’s developer documentation describes high as a reasoning_effort setting.

Developers should therefore compare the actual API identifiers, request settings, workload shape, and operational risk. The benchmark advantage points toward Claude Sonnet 5 for software work, while the price and math evidence leave GPT-5 relevant for narrower use cases.

Claude Sonnet 5 (Non-reasoning, High Effort) offers the clearest coding advantage, while GPT-5 (high) offers the lower blended price and the only reported math score.

Decision area	Better-supported choice	Why it matters
Coding	Claude Sonnet 5 (Non-reasoning, High Effort)	Its coding index is 66.4 versus 37.8 for GPT-5.
General intelligence	Claude Sonnet 5 (Non-reasoning, High Effort)	Its intelligence index is 41.7 versus 34.7.
Math evidence	GPT-5 (high)	GPT-5 has a math index of 94.3; no Claude Sonnet 5 math value is provided.
Blended cost	GPT-5 (high)	GPT-5 costs $3.4375 versus $4 per 1M blended tokens.
Input cost	GPT-5 (high)	GPT-5 input pricing is $1.25 versus $2 per 1M tokens.
Output cost	Tie	Both models are listed at $10 per 1M output tokens.
Reported output speed	Claude Sonnet 5 (Non-reasoning, High Effort)	Claude Sonnet 5 records 64.222 median output tokens per second; GPT-5 has no value in the dataset.
Initial latency	Tie	Both models record 0.3 seconds.

The evidence supports Claude Sonnet 5 as the safer starting point for coding agents, code review, refactoring, and developer tools. The coding gap is large enough to affect task completion, not merely leaderboard position. However, the data does not establish a universal winner for math, long-form reasoning, audio, video, or real-world production reliability.

The official capability profiles also differ. Claude Sonnet 5 accepts text and image input and produces text, with multilingual and vision capabilities. Anthropic’s model overview also documents Adaptive thinking but not Extended thinking through thinking.type: "enabled". GPT-5 accepts text and image input but does not support audio or video input and output. GPT-5 model documentation records that limitation.

Claude Sonnet 5 (Non-reasoning, High Effort) is the stronger evidence-backed choice for coding tasks, but the available measurements do not prove that every software workflow will improve by the same amount.

A coding index of 66.4 versus 37.8 suggests a meaningful difference in tasks that require code generation, repository changes, debugging, and implementation judgment. It does not tell you whether the winning model produces fewer review cycles, fewer regressions, or better architectural decisions in your repository. Those outcomes depend on prompt design, tool access, test coverage, context selection, and the model’s ability to follow local conventions.

GPT-5’s official benchmark material presents strong coding and agentic-task positioning. OpenAI’s developer announcement reports SWE-bench Verified at 74.9%, Aider polyglot at 88%, and τ²-bench telecom at 96.7%. The same source notes that the SWE-bench result excluded 23 problems from 500 because they could not be passed reliably in OpenAI’s infrastructure, and that the Aider evaluation used high reasoning effort. Those details make the results useful, but not directly interchangeable with the Artificial Analysis coding index.

Claude Sonnet 5 has a measured median output speed of 64.222 tokens per second, while GPT-5 has no corresponding value in the supplied data. Both models show 0.3 seconds of latency. That means Claude has the stronger available responsiveness evidence, but a full speed ranking remains unsupported.

The practical test should focus on accepted patches, failed tool calls, test repair, and reviewer effort. The supplied research contains no controlled community consensus for Claude Sonnet 5. A Reddit report describes GPT-5 as useful for small bug fixes but less complete for full applications and potentially risky in complex existing codebases. The Reddit report is anecdotal, so treat it as a test hypothesis rather than a measured conclusion.

GPT-5 (high) is cheaper on the supplied blended and input prices, but Claude Sonnet 5 (Non-reasoning, High Effort) can still be the better economic choice when stronger coding output reduces rework.

GPT-5 costs $3.4375 per 1M blended tokens compared with $4 for Claude Sonnet 5. GPT-5 also has the lower input price at $1.25 versus $2 per 1M tokens. Output pricing is equal at $10 per 1M tokens. The price difference therefore matters most for input-heavy workloads, large repository context, retrieval-heavy prompts, and applications that send substantial instructions on every request.

Raw token pricing does not capture engineering cost. A model that needs more retries, produces less complete patches, or requires more human review can exceed the cheaper model’s API savings. The supplied benchmark gap in coding provides a reason to test whether Claude Sonnet 5 reduces those costs, but it does not provide retry rates, accepted-patch rates, or total task cost. No universal cost winner can be proven without workload data.

Anthropic also warns that its newer tokenizer typically produces about 30% more tokens for the same text, although the exact increase depends on content and workload. Anthropic’s pricing documentation makes that a material budgeting concern for large inputs. The same page lists Claude Sonnet 5 at $2 input and $10 output during the introductory pricing period, followed by standard prices of $3 input and $15 output. Those prices create a time-sensitive procurement issue.

GPT-5’s documented current pricing is $1.25 input, $0.125 cached input, and $10 output per 1M tokens. OpenAI’s GPT-5 model documentation supports the lower input-cost case. Developers should model cache behavior, tokenizer expansion, retries, and human review together.

Claude Sonnet 5 (Non-reasoning, High Effort) should be the default choice for coding products, while GPT-5 (high) should be selected for specific math, price, or existing-platform requirements.

Choose Claude Sonnet 5 when the product’s main value comes from software execution. Its 66.4 coding index is materially above GPT-5’s 37.8 in the supplied comparison. That makes it the stronger candidate for repository agents, pull-request assistance, refactoring, code migration, test generation, and multi-file implementation. The recommendation still requires a task-level evaluation because the data does not measure regression rates or production reliability.

Choose GPT-5 when math is central to the workload. GPT-5 has a math index of 94.3, while no Claude Sonnet 5 math score appears in the data snapshot. That is a genuine evidence boundary, not proof that Claude Sonnet 5 performs poorly at mathematics. It means the supplied material cannot support a math comparison.

Choose GPT-5 when input cost dominates and your prompts are large or frequent. Its blended price is $3.4375 versus $4, and its input price is $1.25 versus $2 per 1M tokens. The savings may disappear if the model creates more repair work, but the price advantage is clear before workload-specific effects.

Choose Claude Sonnet 5 when the available streaming-speed evidence matters. Its median output speed is 64.222 tokens per second, while GPT-5 has no comparable value. Both models have 0.3-second latency, so the data does not establish a latency advantage.

Treat version management as a first-class decision. OpenAI marks the fixed gpt-5-2025-08-07 snapshot as Deprecated and recommends GPT-5.6 in the model documentation. OpenAI’s model documentation still lists gpt-5 as a callable alias. Anthropic’s overview lists claude-sonnet-5 and does not state that a later version has replaced it. Anthropic’s overview also explains that undated identifiers for later generations are fixed snapshots rather than evergreen pointers. Pinning and migration behavior should be tested before production rollout.

Claude Sonnet 5 (Non-reasoning, High Effort) is the better first candidate when your evaluation starts with coding quality and developer productivity.

The comparison has important evidence gaps. The supplied material does not provide a Claude Sonnet 5 math score, a GPT-5 output-speed value, controlled production reliability data, or a broad community consensus for either model. Developers should avoid converting those missing values into assumptions. A small representative evaluation should measure successful task completion, test preservation, tool-call accuracy, review time, retries, and total token use.

API naming also needs care. Claude’s data label describes a configuration, but Anthropic’s official API identifier remains claude-sonnet-5. GPT-5 high describes reasoning_effort=high, not a separate gpt-5-high model. These distinctions affect routing, logging, billing, and migration plans.

The final selection should be reversible. Start with one primary model, keep the prompt and tool contract portable, record task outcomes, and revisit the choice when pricing or model availability changes.