GPT-5.1 Guide: The Release That Made Reasoning Optional

The most consequential change in GPT-5.1 was not a benchmark. It was a default. GPT-5 ran with medium reasoning effort unless you said otherwise; GPT-5.1 shipped with reasoning effort set to none. Code that upgraded the model name and changed nothing else got a faster, cheaper, and on hard prompts noticeably shallower model.

That one setting explains most of what people noticed about the release, and it is still the first thing to check if you are running GPT-5.1 today.

The default changed: none instead of medium

OpenAI's API changelog for 13 November 2025 is explicit: GPT-5.1 defaults to a none reasoning setting rather than the medium default used in GPT-5. The model page lists four levels: none, low, medium and high.

With none, the model answers without spending hidden reasoning tokens. That cuts time to first token and output cost, and it is the right choice for extraction, classification, rewriting and short conversational turns. It is the wrong choice for multi-step maths, debugging, or planning a change across several files, where the reasoning is the work.

The practical consequence: if you migrated from GPT-5 and saw quality drop on hard tasks, set the effort explicitly before concluding the model is worse. Reasoning models explained covers when thinking longer helps, and when it only adds latency.

Instant and Thinking: two behaviours, one family

In ChatGPT, GPT-5.1 arrived as two variants: Instant for everyday, low-latency replies, and Thinking for harder questions. OpenAI pitched both as adapting how much they think to the difficulty of the question, alongside new tone and personality controls in the ChatGPT app.

On the API that split surfaced as two model ids. gpt-5.1 is the configurable model described here. gpt-5.1-chat-latest is the snapshot used in ChatGPT, with a smaller 128,000-token context window and a 16,384-token output cap. If you want ChatGPT's conversational behaviour, the chat alias is closer; if you want control over effort, tools and long outputs, use gpt-5.1.

Specs and price

From OpenAI's model page:

• Context window — 400,000 tokens, with up to 128,000 output tokens
• Modalities — text and image in, text out
• Knowledge cutoff — 30 September 2024
• Snapshot — gpt-5.1-2025-11-13
• Endpoints — Chat Completions, Responses and Batch
• Price — $1.25 per million input tokens, $0.125 cached input, $10 per million output

That price is identical to GPT-5, so the upgrade was free on paper. In practice the bill depended on effort: the same request at high can produce many times the output tokens it does at none, all billed at the output rate. The catalogue we sync from OpenRouter listed the same $1.25 and $10 rates at the time of writing.

The knowledge cutoff is the weak point. September 2024 was already more than a year old at release, so any question about recent libraries or APIs needs retrieval or web search rather than trust in the model's memory.

Batch and the :batch listing

Unlike the Codex models of the same generation, GPT-5.1 supports OpenAI's Batch endpoint. OpenRouter also lists a gpt-5.1:batch variant: the same model on a discounted asynchronous endpoint, shown at $0.625 input and $5 output per million tokens at the time of writing, half the standard rate.

Batch is the right place for evaluation runs, bulk classification, backfills and nightly reports, anything where an answer within hours is fine. It is the wrong place for anything a user is waiting on. Batch API savings math works through when the discount outweighs the extra plumbing.

GPT-5.1 or GPT-5.2

GPT-5.2 followed a month later with a newer knowledge cutoff, an extra xhigh effort level and compaction-based context management, at $1.75 input and $14 output. That is 40% more on both lines.

GPT-5.1 still makes sense where reasoning is mostly off and the work is short: routing, extraction, summaries, chat turns. There the extra capability of 5.2 rarely shows, and the price gap does. Where tasks are long, agentic or depend on knowledge from 2025, 5.2 is the better model of the two.

Deprecation timeline and planning

OpenAI added gpt-5.1 to its deprecations page on 1 October 2026, with removal from the API on 1 April 2027 and gpt-6-sol named as the replacement. Six months is a generous window, but it is easy to waste.

1. Find every place the model id is hard-coded, including scripts and CI jobs, and move it to one config value.
2. Record the reasoning effort each call uses today. A replacement may have a different default, exactly as 5.1 did.
3. Build a small regression set from real prompts and accepted outputs, and run it against the replacement well before the deadline.

Model deprecation and migration covers the rest. The lesson GPT-5.1 taught is worth carrying into that migration: when a model upgrade looks worse, check the defaults before you check the model.