GPT-5.5 Guide: The Flagship That Now Costs More Than Its Successor

GPT-5.5 is in an odd position. When OpenAI released it on 24 April 2026 it was the frontier option, priced at double its predecessor. By October it costs more per token than GPT-5.6 Sol, the model that replaced it as flagship, and two further generations have shipped since. That does not make it useless, but it does mean the default reason to call it has gone.

This guide covers what actually changed between GPT-5.4 and GPT-5.5, the one default that quietly changes your bill, and the narrow set of cases where staying on GPT-5.5 is the right call.

What changed from GPT-5.4

OpenAI's model page describes GPT-5.5 as "a new class of intelligence for coding and professional work". The catalogue description on OpenRouter is more specific: stronger reasoning, higher reliability and better token efficiency on hard tasks, building on GPT-5.4. OpenAI has not published a parameter count or architecture details for either.

What is verifiable from the two model pages:

• Knowledge cutoff moved from 31 August 2025 to 1 December 2025.
• Context stayed the same: 1,050,000 tokens, 128,000 maximum output.
• Modalities are unchanged: text and image in, text out.
• Price doubled, from $2.50 to $5 per million input tokens and from $15 to $30 per million output. Cached input went from $0.25 to $0.50.
• Endpoints are Chat Completions, Responses and Batch, the same set as GPT-5.4.

The Responses tool list for GPT-5.5 is broad: function calling, web and file search, tool search, image generation, code interpreter, hosted shell, apply patch, skills, computer use and MCP. If you came from Chat Completions, most of those are only reachable through Responses.

The default effort moved, and that is your bill

This is the change most people miss. GPT-5.4 defaults to none reasoning effort. GPT-5.5 defaults to medium. Both accept none, low, medium, high and xhigh, but a request that omits the field behaves very differently on each.

Swap the model string in an existing integration and leave everything else alone, and every call now produces reasoning tokens that bill as output — at twice the old output rate. A team that budgeted the upgrade as "double the price" can see considerably more than double, because the token count rose as well. The cost of model upgrades covers this pattern in general; GPT-5.5 is a clean example of it.

Latency moves with it. Artificial Analysis measured a time to first token of around 35 seconds for GPT-5.5 at xhigh on OpenAI's API at the time of writing, against a median of under four seconds across the models it tracks. Set effort explicitly on every request rather than inheriting a default.

The price argument has flipped

On 9 July 2026 OpenAI released GPT-5.6 Sol, and on 21 August cut its price to $4 input and $20 output, describing the change as 20% lower input and 33% lower output pricing. GPT-5.5 stayed at $5 and $30. OpenAI's own Sol page lists the three side by side: Sol at $4 and $20, GPT-5.5 at $5 and $30, GPT-5.4 at $2.50 and $15.

Two caveats keep this from being a simple "move to Sol". OpenAI describes Sol's rate as promotional pricing available at least through 21 November 2026, so it may rise. And GPT-6 Sol, released on 22 September, lists at $2 and $10, below both. If you are leaving GPT-5.5 for cost reasons, the GPT-5.6 Sol guide and the GPT-6 Sol guide are the two comparisons to read.

Why you might still pin it

There are concrete reasons to keep GPT-5.5, and none of them is "it is the smartest":

• A fixed snapshot. gpt-5.5-2026-04-23 is the only snapshot listed. Prompts, graders and regression suites tuned against it keep working.
• Reasoning does not carry across families. OpenAI's reasoning guide says GPT-5.6 models can reuse each other's persisted reasoning, but reasoning does not carry between GPT-5.6 and GPT-5.5. Long-running stateful sessions should not switch model mid-flight.
• Prompt caching changed. Migrating off GPT-5.5 or earlier means replacing prompt_cache_retention with prompt_cache_options.ttl, and cache writes on GPT-5.6 bill at 1.25 times the uncached input rate. That is a code change, not a string swap.

If your workload depends on cached prefixes, model the new cache-write charge before assuming the newer model is cheaper. Prompt caching economics has the break-even arithmetic.

Batch and long-context arithmetic

OpenRouter lists a :batch variant of GPT-5.5 at $2.50 input, $15 output and $0.25 cached at the time of writing — half the standard rate. It is the same model on an asynchronous queue: you submit work and collect the results later instead of waiting on the connection. For evaluation runs, backfills and document processing, that halves the bill with no quality trade.

Watch the long-context rule. Prompts over 272K input tokens bill at twice the input rate and one and a half times the output rate for the entire session, and OpenAI says this applies to standard, batch and flex processing alike. A 1M window is available, but filling it doubles the input price. Batch API savings covers when the latency trade pays.

Stay or move: a decision rule

1. If you are starting new work, do not start on GPT-5.5. Evaluate GPT-5.6 Sol and GPT-6 Sol first.
2. If you have a validated GPT-5.5 integration, set reasoning effort explicitly, then run your own regression set against the replacement before switching.
3. If you need more depth than GPT-5.5 gives at xhigh, the sibling is GPT-5.5 Pro — at six times the per-token price, it is for a small number of hard calls, not a default.