GPT-5 in 2026: Should You Keep the August 2025 Model?

Plenty of production systems still send traffic to gpt-5. They were wired up in late 2025, they work, and nobody has had a reason to touch them. Fourteen months and several model generations later, that is worth a deliberate decision rather than inertia.

This guide covers what GPT-5 actually is in the API, how it compares on paper with what OpenAI sells now, and a checklist for deciding whether to stay or move.

One name, two different things

OpenAI launched GPT-5 on 7 August 2025. In ChatGPT it was presented as a system rather than a single model: a fast model for most questions, a deeper reasoning model for harder ones, and a real-time router choosing between them based on the conversation, its complexity, tool needs and what the user asked for.

That router is a ChatGPT feature. In the API, gpt-5 is a reasoning model you call directly, alongside gpt-5-mini and gpt-5-nano, which OpenAI released the same day. There is no automatic switching; you choose the model and the reasoning effort yourself. This matters when reading old commentary, because much of the launch criticism was about the router, including OpenAI attributing one bad day to an outage in the automatic switcher. API users never had that layer.

The spec sheet, fourteen months on

From OpenAI's model page, which now labels GPT-5 a previous model and recommends GPT-6 Astra:

• Snapshot: gpt-5-2025-08-07 is the default.
• Context: 400,000 tokens, of which up to 272,000 can be input and 128,000 output.
• Knowledge cutoff: 30 September 2024.
• Reasoning effort: minimal, low, medium and high.
• Modalities: text and image in, text out.
• Price: $1.25 per million input tokens, $0.125 cached, $10 output.
• Endpoints: Chat Completions, Responses and Batch; not Realtime or fine-tuning.

OpenRouter lists the same $1.25 and $10, plus an openai/gpt-5:batch endpoint at half price, $0.625 input and $5 output at the time of writing. That variant is the same model through an asynchronous queue: results arrive later, but cost half as much.

The knowledge cutoff is the line that has aged worst. Anything published after September 2024, including every library release, API change and model launch since, is outside what GPT-5 knows unless you put it in the prompt.

What GPT-5 introduced that stuck

The GPT-5 release added a few API features that later models kept, so code written for it is not a dead end:

• A minimal reasoning effort for fast responses on simple tasks. Later families replaced it: GPT-5.6 offers none instead, and GPT-6.1 Sol supports neither, so this is the parameter most likely to break on migration.
• A verbosity control for steering answer length separately from reasoning depth.
• Custom tools, a tool call type that accepts freeform text rather than JSON arguments, which suits passing code or patches to a tool.

OpenAI's current migration notes suggest starting with low effort when an existing request uses minimal, then comparing results on representative tasks.

Reasons to stay pinned

There are honest reasons to keep GPT-5 for now:

• Validated behaviour. If you have an eval suite, prompt library and output parsers tuned to GPT-5, a migration is a project, not a config change. What a model upgrade costs breaks down the hidden work.
• The price is not bad. At $1.25 input and $10 output, GPT-5 is cheaper per token than GPT-5.6 Terra at $2 and $12. For workloads that fit in 272K input tokens and do not need recent knowledge, that gap is real.
• Batch jobs. At half price on batch, a stable nightly pipeline has little to gain from moving unless quality is the problem.

Reasons to move

The case for migrating is stronger for anything new or anything that is struggling:

• Cheaper and newer exists. GPT-5.6 Luna is listed at $0.20 input and $1.20 output, with a 1,050,000-token window and a February 2026 cutoff. For high-volume, simpler tasks it costs a fraction of GPT-5. Whether it is good enough is something to test, not assume.
• Longer context. GPT-5.6 models accept up to 922,000 input tokens against GPT-5's 272,000.
• Newer tools. GPT-5.6 models on the Responses API support hosted shell, apply patch, computer use and tool search, which GPT-5 does not list.
• Deprecation risk. A model labelled previous is on the path to retirement. Moving on your own schedule is cheaper than moving on OpenAI's.

If GPT-5 is too weak rather than too old, the sibling options are GPT-5 Pro for more reasoning on the same generation, or a newer tier. If it is more than you need, GPT-5 Mini is the cheaper member of the same family.

A migration checklist

1. Freeze a set of twenty to fifty real requests and their accepted outputs from GPT-5.
2. Run them against the candidate replacement at a similar effort level, mapping minimal to low or none.
3. Compare pass rate, total tokens and latency, not just per-token price. Newer reasoning models can spend more or fewer tokens on the same task.
4. Check parsers, tool schemas and anything that depends on answer length or format.
5. Shift a small percentage of live traffic and watch error and retry rates before cutting over. A/B testing two models covers how to avoid reading noise as signal.

The short version: keep GPT-5 where it is stable, cheap and not knowledge-sensitive. Move anything new, anything near the 272K input limit, and anything where a 2024 cutoff shows.