Short answer: GPT-5 was designed for coding, reasoning, and agentic tasks. But OpenAI now labels both the gpt-5 alias and gpt-5-2025-08-07 snapshot deprecated, describes GPT-5 as a previous model, and recommends GPT-6 Astra. For a new production integration, start by evaluating the currently recommended model rather than assuming GPT-5’s launch-era reputation makes it the best current choice. (developers.openai.com)
If you have a specific reason to use GPT-5—such as maintaining an existing integration or comparing model behavior—test it against your own coding tasks. Keep model quality, coding-agent experience, and cost separate in that evaluation.
What GPT-5 was built to do
OpenAI’s August 2025 developer announcement presented GPT-5 as a model for coding and agentic tasks, including fixing bugs, editing code, and answering questions about complex codebases. Those are launch-era product claims, not a current head-to-head ranking or a guarantee of results on your repository. (developers.openai.com)
The practical answer to “Is GPT-5 good for coding?” depends on the task and setup. Developers in an August 2025 discussion reported both successes and problems with instruction-following, code quality, patch application, and latency. These are anecdotes, not controlled comparisons; they suggest what to test, not a universal verdict.
What to compare in a coding trial
Use tasks that resemble the work you actually want to delegate. For example, select a small bug fix, a change spanning several files, and a code-review or debugging task. Run the same tasks with the same repository context and tools for each model you compare.
Record observable results, not just whether the model’s explanation sounds convincing:
- Correctness: Did the patch satisfy the acceptance criteria? Did the relevant tests pass?
- Instruction adherence: Did it respect constraints such as files not to change, required patterns, or requested design choices?
- Patch quality: Was the change minimal and understandable, or did it add redundant logic or unrelated edits?
- Time and review: How long did the run take, and how much developer time did it take to verify and correct the result?
- Context handling: For a multi-step task, did it retain the important requirements as the work progressed?
A hypothetical scoring example: if Model A gets a task passing in 20 minutes but needs 25 minutes of review, while Model B takes 30 minutes and needs 5 minutes of review, Model B may be the better fit if total engineering time is your priority. These numbers are illustrative, not GPT-5 test results.
For long tasks, consider splitting work into reviewable units and checking each patch before proceeding. Participants in the developer discussion described shorter tasks, fresh sessions, and summaries as ways they managed context. These are user-reported workflow ideas, not a validated GPT-5-specific protocol.
Check the production API status before adopting GPT-5
OpenAI’s current model page marks both gpt-5 and gpt-5-2025-08-07 as deprecated. It lists a 400,000-token context window, a 128,000-token maximum output, function calling and structured outputs, and support for the Responses and Chat Completions endpoints. These specifications describe the documented model; they do not establish that it will remain available. OpenAI’s deprecation guidance says deprecated models are in the process of being retired and that the model’s shutdown_date can be checked through the List models or Retrieve a model endpoint. The GPT-5 page does not state a shutdown date, so the public documentation reviewed here does not settle how long a particular account can continue using it. (developers.openai.com)
The GPT-5 page lists text and image input, text output, and reasoning-effort values of minimal, low, medium, and high. Confirm any requirements for your integration against the current model documentation; do not assume features or access in ChatGPT, Codex, or an IDE integration automatically apply to API use. (developers.openai.com)
Decision rule: If you are maintaining an existing GPT-5 integration, check its current model status and plan a migration evaluation. If you are choosing a model for a new production system, benchmark the currently recommended model alongside any legacy option you need to compare. OpenAI’s GPT-5 page recommends GPT-6 Astra. (developers.openai.com)
GPT-5 API pricing: standard versus Batch
The current GPT-5 model page displays $1.25 per 1 million input tokens, $0.125 per 1 million cached input tokens, and $10 per 1 million output tokens, explicitly labeled Batch API pricing. GPT-5 does not appear in the current general pricing page’s displayed model tables, so the official public pricing information reviewed here does not establish a standard-processing rate for GPT-5. Do not use the Batch rates as an interactive coding quote; check applicable pricing and availability for your exact model and processing mode before budgeting. (developers.openai.com)
For a Batch-only illustration, assume one request has 20,000 input tokens, no cached input, and 5,000 total billable output tokens. The 5,000 output-token assumption includes any reasoning tokens, not just visible code or text; OpenAI documents reasoning tokens as part of output usage. At the listed Batch rates:
- Input: 20,000 ÷ 1,000,000 × $1.25 = $0.025
- Output: 5,000 ÷ 1,000,000 × $10 = $0.05
- Total: $0.075 for that one request, before any other applicable charges
This arithmetic is not a standard-processing estimate. For a coding task involving multiple model calls, add the billable input and output usage for every call; in multi-turn Responses API chains, prior input tokens are billed again. Actual usage can also differ from visible text because reasoning tokens count toward output. (developers.openai.com)
Bottom line
GPT-5 was positioned for coding, but for a decision made in November 2026, OpenAI’s current documentation marks it deprecated and recommends GPT-6 Astra. The documentation reviewed here gives GPT-5’s Batch rates, but does not provide a standard-processing rate in the displayed pricing tables; the Batch example above cannot budget interactive coding. Choose based on current model availability, the API capabilities your application needs, and results on representative code tasks. Track test success, instruction adherence, latency, review effort, and cost together. Launch-era reports and benchmark claims alone cannot tell you whether a model fits your production workflow. (developers.openai.com)
Sources
Observability and usage | OpenAI API
Migrate to the Responses API | OpenAI API