GPT-5.5
GPT-5.5 was OpenAI's top model from late April to early July 2026, but it is now two generations old and poor value in the API. OpenAI released GPT-5.5 Thinking and Pro in ChatGPT on 23 April 2026 and in the API the next day. GPT-5.5 Instant became ChatGPT's everyday default on 5 May 2026. Its codename inside OpenAI was "Spud".
It improved on GPT-5.4 in coding, maths and long-document work. OpenAI reported 82.7% on Terminal-Bench 2.0 and 74.0% on a 1-million-token retrieval test, up from 36.6%. But it also doubled the API price to $5 per million input tokens and $30 per million output tokens. GPT-5.6 replaced it on 9 July 2026, and GPT-6 Sol now does similar work for $2 / $10.
GPT-5.5 was a solid step up from GPT-5.4, but there is little reason to choose it today.
What it got right, on OpenAI's own figures:
- Terminal work: 82.7% on Terminal-Bench 2.0, up from 75.1% for GPT-5.4.
- Long documents: 74.0% on the MRCR v2 1M-token retrieval test, double GPT-5.4's 36.6%.
- Hard maths: 35.4% on FrontierMath Tier 4, up from 27.1%.
It was not a clean sweep. Claude Opus 4.7 scored higher on SWE-Bench Pro (64.3% vs 58.6%), and Gemini 3.1 Pro edged it on BrowseComp. The Decoder also reported that real API bills ran about 20% higher than GPT-5.4, despite OpenAI's claim that it used fewer tokens.
Who should use it: teams with production prompts tuned to gpt-5.5-2026-04-23 who have not yet re-tested on a newer model.
Who should not: anyone starting fresh. GPT-6 Sol costs $2 / $10 against GPT-5.5's $5 / $30, is newer, and beats it on OpenAI's agent benchmarks. For the hardest work, GPT-6 Astra is the current flagship.
Score breakdown
Best for
- Existing apps already tuned to the GPT-5.5 snapshot
- Long-document question answering where prompts are already validated
- Comparing old and new OpenAI models during a migration
What we like
- 82.7% on Terminal-Bench 2.0, a big jump over GPT-5.4's 75.1% (OpenAI)
- Much better at finding facts deep in 1M-token prompts (74.0% vs 36.6% on MRCR v2)
- Instant, Thinking and Pro versions covered quick chat through deep research
- Stable dated snapshot with no deprecation notice as of September 2026
Watch out for
- Double GPT-5.4's API price and 2.5x GPT-6 Sol's input price
- Trailed Claude Opus 4.7 on SWE-Bench Pro (58.6% vs 64.3%)
- Independent tests found real costs about 20% above GPT-5.4
- Replaced in ChatGPT by GPT-5.6 within 11 weeks
Specs at a glance
| Developer | OpenAI |
|---|---|
| Codename | Spud |
| Release dates | Thinking and Pro: 23 April 2026 (ChatGPT), 24 April 2026 (API); Instant: 5 May 2026; GPT-5.5-Cyber preview: 7 May 2026 |
| Variants | GPT-5.5 Instant, GPT-5.5 Thinking, GPT-5.5 Pro, GPT-5.5-Cyber (vetted security teams only) |
| API snapshot | gpt-5.5-2026-04-23 |
| Context window | 1,050,000 tokens |
| Max output | 128,000 tokens |
| Knowledge cutoff | 1 December 2025 |
| Input / output | Text and image in; text out |
| Reasoning effort | none, low, medium (default), high, xhigh |
| API pricing | $5 input, $0.50 cached input, $30 output per 1M tokens |
| Long-prompt surcharge | Over 272K input tokens: 2x input and 1.5x output for the session |
| GPT-5.5 Pro API pricing | $30 input, $180 output per 1M tokens |
| Endpoints | Chat Completions, Responses, Batch |
| Predecessor / successor | GPT-5.4 / GPT-5.6 |
Benchmarks
Benchmarks are standard tests. Vendor-run results are marked as such; independent results are preferred where they exist.
| Benchmark | Score | Source | Note |
|---|---|---|---|
| Terminal-Bench 2.0 | 82.7% | OpenAI | GPT-5.4: 75.1%; Claude Opus 4.7: 69.4% |
| GDPval (wins or ties vs experts) | 84.9% | OpenAI | GPT-5.4: 83.0% |
| OSWorld-Verified | 78.7% | OpenAI | GPT-5.4: 75.0%; Claude Opus 4.7: 78.0% |
| SWE-Bench Pro | 58.6% | OpenAI | Claude Opus 4.7: 64.3% |
| BrowseComp | 84.4% | OpenAI | Gemini 3.1 Pro: 85.9% |
| MRCR v2 (1M tokens) | 74.0% | OpenAI | GPT-5.4: 36.6% |
| FrontierMath Tier 4 | 35.4% | OpenAI | GPT-5.4: 27.1% |
Pricing
| Plan / tier | Price | Notes |
|---|---|---|
| API input | $5 per 1M tokens | 2x above 272K input tokens |
| API cached input | $0.50 per 1M tokens | |
| API output | $30 per 1M tokens | 1.5x above 272K input tokens |
| GPT-5.5 Pro (API) | $30 in / $180 out per 1M tokens | |
| Batch | 50% of standard |
GPT-5.5 vs GPT-5.4 vs rivals (April 2026)
| Benchmark | GPT-5.4 | GPT-5.5 | Claude Opus 4.7 | Gemini 3.1 Pro |
|---|---|---|---|---|
| Terminal-Bench 2.0 | 75.1% | 82.7% | 69.4% | 68.5% |
| GDPval | 83.0% | 84.9% | 80.3% | 67.3% |
| OSWorld-Verified | 75.0% | 78.7% | 78.0% | n/a |
| SWE-Bench Pro | n/a | 58.6% | 64.3% | n/a |
| BrowseComp | 82.7% | 84.4% | 79.3% | 85.9% |
| FrontierMath Tier 4 | 27.1% | 35.4% | 22.9% | 16.7% |
Figures are from OpenAI's launch materials as reported by The Decoder. Vendor-run numbers use the vendor's own settings, so treat small gaps with caution.
The price jump
GPT-5.5 was the first GPT-5.x model to raise prices sharply:
| Model | Input / output per 1M tokens |
|---|---|
| GPT-5.4 | $2.50 / $15 |
| GPT-5.5 | $5 / $30 |
| GPT-5.6 Sol (Sept 2026) | $4 / $20 |
| GPT-6 Sol | $2 / $10 |
| GPT-6 Astra | $10 / $50 |
OpenAI argued that GPT-5.5 needed fewer tokens per task, so total cost would rise less than the rate card suggests. The Decoder's reporting found real bills still came in roughly 20% higher than GPT-5.4.
GPT-5.5 in ChatGPT
- 23 April 2026: GPT-5.5 Thinking and Pro reach paid plans (Plus, Pro, Business, Enterprise).
- 5 May 2026: GPT-5.5 Instant replaces GPT-5.3 Instant as the default for everyone, including free users.
- 28 May 2026: OpenAI updates Instant to write more naturally, with fewer bullet-heavy replies.
- 9 July 2026: GPT-5.6 Luna, Terra and Sol arrive and take over.
We could not find an official retirement date for GPT-5.5 in ChatGPT as of 23 September 2026. The API model has no deprecation notice.
Quirks worth knowing
Wikipedia records that GPT-5.x models, starting with GPT-5.1, had a habit of mentioning goblins and gremlins. OpenAI traced it to reward signals used to train the "Nerdy" personality and reduced it with data filtering. It is harmless, but a reminder that personality settings can leak into answers in odd ways.
OpenAI also offered GPT-5.5-Cyber from 7 May 2026 as a limited preview for vetted security teams. It is not a general-purpose product.
- If you still run GPT-5.5, test the same prompts on GPT-6 Sol first. At $2 / $10 it could cut your bill by more than half.
- Keep prompts under 272K tokens when you can. Above that line, the whole session is billed at 2x input and 1.5x output.
- Set reasoning effort to
noneorlowfor simple tasks. GPT-5.5's default ismedium, and every reasoning token is billed at $30 per million. - Re-use a fixed system prompt at the start of each request so prompt caching kicks in. Cached input costs $0.50 instead of $5.
- Pin
gpt-5.5-2026-04-23rather than a moving alias so behaviour does not change under you.
Jargon explained
- Terminal-Bench
- A test where the AI must finish real tasks by typing commands into a computer terminal.
- MRCR
- A long-context test that hides several similar facts in a huge prompt and checks whether the model can find the right one.
- Cached input
- Prompt text the API has seen recently. OpenAI charges much less when it can reuse it.
- FrontierMath
- A set of very hard, unpublished maths problems written by research mathematicians. Tier 4 is the hardest level.
- Snapshot
- A fixed, dated version of a model that will not change, so your app behaves the same over time.
Alternatives to consider
GPT-6 Sol
OpenAI's current mid-tier model at $2 / $10, well under half GPT-5.5's price.
GPT-5.6
Direct successor with Luna, Terra and Sol sizes.
GPT-6 Astra
OpenAI's flagship for the hardest reasoning and agent tasks.
Claude Fable 5
Anthropic's rival, strong on SWE-Bench Pro coding.
Gemini 3.5 Pro
Google's announced long-context flagship; still unreleased (partner testing only) as of September 2026.
Frequently asked questions
When was GPT-5.5 released?
GPT-5.5 Thinking and Pro launched in ChatGPT on 23 April 2026, and in the API on 24 April. GPT-5.5 Instant followed on 5 May 2026.
How much does GPT-5.5 cost in the API?
$5 per million input tokens, $0.50 cached and $30 per million output tokens. Prompts over 272K tokens cost 2x input and 1.5x output. GPT-5.5 Pro costs $30 / $180.
What is GPT-5.5's context window?
1,050,000 tokens, with up to 128,000 output tokens. Its knowledge cutoff is 1 December 2025.
Is GPT-5.5 better than GPT-5.4?
Yes on most of OpenAI's tests, for example 82.7% vs 75.1% on Terminal-Bench 2.0. It also costs twice as much per token.
Should I still use GPT-5.5?
Only if your app is tuned to it. For new work, GPT-6 Sol is newer and much cheaper at $2 / $10.
What replaced GPT-5.5?
GPT-5.6 on 9 July 2026, then the GPT-6 family in September 2026.
Sources
Every fact on this page comes from public information. Vendor figures are labelled as vendor claims.
- GPT-5.5 (Wikipedia)
- GPT-5.5 model page (OpenAI API docs)
- OpenAI unveils GPT-5.5, claims a "new class of intelligence" at double the API price (The Decoder)
- OpenAI upgrades GPT-5.5, as it plans to retire legacy ChatGPT models (BleepingComputer)
- Model Release Notes (OpenAI Help Center)
- GPT-5.5 on OpenRouter (OpenRouter)
- GPT-6 Sol model page (OpenAI API docs)