benchr
OpenAI shipped GPT-5 in August 2025, per OpenAI's launch post. GPT-5 was a big step up from GPT-4o, and nobody disputes that. The real question is whether you should still pay for it once Claude Opus 4.7 and the Gemini 3 family are also on the table. For specific work, yes. This piece walks through which work, drawing on the public benchmark record, OpenAI's own positioning, and the comparison patterns already visible in the open between the three frontier labs.
OpenAI positions GPT-5 across reasoning, multimodal, structured-output, and conversational work. This page does not have a publishable controlled dataset that establishes it as the most flexible, most confident, or most error-prone relative to other frontier models. Those are evaluation questions, not facts to assume.
Where OpenAI positions the model
OpenAI's launch material for GPT-5 emphasizes three things you should weight carefully. The first is reasoning quality on math and structured problem-solving (the company reports strong MATH benchmark performance and improved chain-of-thought behavior). The second is multimodal breadth, covering vision, audio, and text in a single model, with vision quality positioned as best-in-class. The third is "conversational warmth," the company's term for how the model handles open-ended interaction.
Those are provider positioning claims, not independent verification. Use them to select test strata, then run current versions on your actual inputs, tools, safety constraints, and acceptance criteria before deciding where the model belongs.
Coding: test correctness beyond compilation
OpenAI reports about 74.9% for GPT-5 on SWE-bench Verified in the cited launch snapshot, while Anthropic reports a different result for Opus 4.7. Cross-provider figures can use different configurations and do not establish your repository success rate. Community discussion is not an independent reproduction.
Build a repository evaluation that goes beyond compilation: include edge cases, slicing boundaries, concurrency, unfamiliar APIs, regression tests, and an allowed-file list. Give each model the same repository state, prompt, tools, retry budget, and time limit; record test-pass rate, fabricated APIs, unauthorized changes, reviewer minutes, latency, and tokens.
Require human review and automated tests for production changes regardless of model. Compare GPT and Claude under the same protocol instead of treating either as inherently safer. The Opus 4.7 review uses the same evidence boundary.
The compiler isn't the reviewer that matters here. GPT-5's bugs clear the build and wait for a human to notice; Claude tends to flag its own shaky spots first.
Writing, design, and conversational breadth
Open-ended writing and design are relevant GPT-5 evaluation categories because OpenAI positions the model there. Compare anonymized drafts from identical briefs, and score voice, factual fidelity, accessibility, policy compliance, and editing time. This page does not claim a measured first-draft winner.
That translates into a working advantage for any task where stylistic flexibility matters more than correctness or hedging: drafting copy in a brand voice nobody has documented, writing narrative prose with a deliberate atmosphere, producing decks where the words carry as much weight as the data, anything in that family. GPT-5 will give you the most useful first attempt of the three frontier models on that kind of work.
Do not infer language or dialect quality from a broad multilingual label. Create separate blind-review strata for every language and market you serve, including dialectal Arabic, and score meaning, register, grammar, unsupported additions, and editing time with fluent reviewers. The Arabic content guide provides a reproducible protocol.
Niche technical hallucination
Fabricated or outdated API claims are a failure hypothesis to test for every model. Pin library versions, sample common and less-common APIs, and verify every symbol against official documentation or a compiler. Report fabrication and abstention rates by group rather than attributing a comparative frequency without a controlled dataset.
Verify API claims against primary documentation and tests regardless of model tone. For niche frameworks and older APIs, require citations or tool lookup, record uncertainty and abstention, and compare current candidates under the same setup.
What it costs
In the pricing snapshot verified for this article, OpenAI lists GPT-5 at $1.25 per million input tokens and $10 per million output tokens, per its API pricing page. Per-token list price is not total workflow cost. Compare input, output, caching, Batch eligibility, retries, tool calls, latency, failure rates, and review labor; verify the live page before buying capacity. The price-per-use-case guide provides the calculation framework.
| Tier | Input ($/M tokens) | Output ($/M tokens) | Notes |
|---|---|---|---|
| GPT-5 | $1.25 | $10 | Standard frontier tier |
| GPT-5 Mini | $0.25 | $2.00 | Distilled model for volume |
| GPT-5 (Batch API) | $0.625 | $5.00 | 50% off the $1.25/$10 standard rates; asynchronous processing |
For workloads that do not need a synchronous response — overnight document processing or bulk classification, for instance — OpenAI's 50% Batch API discount reduces GPT-5 from the $1.25/$10 standard rates to $0.625/$5 per million input/output tokens. Confirm the current processing window and eligibility in OpenAI's documentation before planning a production batch.
Where GPT-5 earns its keep
A few categories where GPT-5 is the right pick for you.
For visual and design-heavy work such as landing pages, decks, and layouts, include GPT-5 because OpenAI positions it for multimodal tasks. Use identical briefs and reference assets; score requirement coverage, accessibility, visual defects, factual accuracy, and editing time.
For open-ended writing such as voice matching and brand work, run a blind review on the same brief. Score voice, factual discipline, prohibited claims, editing time, and reviewer preference; do not assign a default from an unmeasured first-draft impression.
For multilingual work, test each language and regional register separately with fluent reviewers. Neither a provider label nor this article establishes a safer first pick across non-English languages.
Where to skip it
For production code review, compare candidates on seeded and historical bugs, with tests and human approval required. Score detection, false positives, invented APIs, unsafe edits, and reviewer time; do not exclude or prefer a model from tone alone.
Skip it, too, on reasoning under uncertainty where honesty is part of what you're paying for — legal review, medical questions, financial analysis. On those workloads, a confident answer that turns out to be wrong does more damage than an honest "not sure."
For long-context synthesis, place answer-key facts at several depths and compare citation accuracy, omissions, unsupported statements, and coherence under an equal token budget. Maximum context and provider positioning do not establish the winner.
The verdict that has not moved
GPT-5 is one current candidate among several. OpenAI's documented pricing, positioning, and provider-reported benchmarks justify including it in a comparison; they do not prove a rank, speed advantage, stylistic lead, language breadth, or model-specific failure rate.
Choose one model, multiple models, or a router only after measuring task success, safety, review time, reliability, latency, and total cost. A two-provider setup adds resilience and optionality but also routing, privacy, operations, and billing complexity; it is not a universal recommendation.
If you're forced to pick one for cost reasons, let the work decide. Teams that write more than they code should take GPT-5; teams that code more than they write should take Claude. Most readers sit in the middle, and for them the both-models setup is the answer.