OpenAIProprietary

GPT-5.5 Instant

Name: GPT-5.5 Instant
Author: OpenAI

A high-speed foundation model developed by OpenAI. It balances low latency with high throughput and is the predecessor to GPT-5.6 Luna.

Share:X Hatena LINE

Parameters

Undisclosed

Context Window

License

Proprietary

Release Date

2026-04-01

Japanese Language Capability

✅High-Quality JP

Multilingual model with strong Japanese language processing capabilities.

API Pricing

API pricing for this model is not yet available

Strengths

Weaknesses

Use Cases

Deep Analysis

Artificial Analysis Intelligence Index

#77 of 186 models; above average (median: 30)

GPQA Diamond

84.6%

#84 across all tracked models

AIME 2025

81.2%

Up from 65.4% on GPT-5.3 Instant

MMMU-Pro (Multimodal Reasoning)

76%

Up from 69.2% on GPT-5.3 Instant

Input Price

$5.00/1M tokens

Cache hit: $0.50/1M (90% discount)

Output Price

$30.00/1M tokens

Blended 3:1 ratio: $11.25/1M

Hallucination Reduction

52.5% fewer

vs GPT-5.3 Instant on high-stakes prompts

Context Window

400K tokens

~600 pages; 128K in ChatGPT UI

Strengths

・52.5% fewer hallucinations on high-stakes medical, legal, and financial prompts vs predecessor
・Significantly improved math and multimodal reasoning (AIME +24.2%, MMMU-Pro +9.8%)
・Deep personalization via Memory Sources with full user control over cited context

Weaknesses

・Premium pricing ($5/$30 per 1M tokens) — significantly above market average ($1.71/$8.70)
・Still outperformed by Claude Opus 4.7 on long-form prose writing and detailed reasoning tasks
・Citation behavior shifted heavily toward Reddit (3× most-cited) with brand site citations halved to 6%

Competitor Comparison

Model	Price
GPT-5.3 Instant (predecessor)	$5/$30 per 1M
Claude Opus 4.7 (Anthropic)	Higher (exact pricing not disclosed in sources)
Gemini 3.1 Pro (Google)	N/A (competitive on multimodal tasks)

Overview

GPT-5.5 Instant, released May 5, 2026, is OpenAI's latest default model for ChatGPT, replacing GPT-5.3 Instant across all user tiers including the free plan. Positioned as the everyday workhorse of the GPT-5.5 family, it prioritizes low-latency, high-factuality conversational responses over the extended reasoning of OpenAI's dedicated thinking models. The model is available via the API as the `chat-latest` alias and is integrated into Microsoft 365 Copilot (since May 7, 2026) and GitHub Copilot. The headline improvements center on factual reliability: OpenAI reports 52.5% fewer hallucinated claims on high-stakes prompts and 37.3% fewer inaccurate claims on conversations users previously flagged for errors. These claims are supported by meaningful benchmark jumps — AIME 2025 went from 65.4% to 81.2%, MMMU-Pro from 69.2% to 76%, and HealthBench Professional from 32.9 to 38.4. The model also introduced real-time self-correction mid-response, where it can catch and fix its own reasoning errors before completing an answer. Beyond raw intelligence, GPT-5.5 Instant's most distinctive innovation is its Memory Sources personalization system. The model draws on past conversations, uploaded files, and connected Gmail (for Plus/Pro users) to provide contextually tailored responses, then transparently shows which sources were referenced so users can edit or remove them. Combined with a 30.2% reduction in word count and fewer unnecessary formatting elements, the model represents OpenAI's push toward making AI feel less like a tool and more like a persistent, trusted assistant. However, at $5/$30 per million tokens via API, it remains premium-priced, and independent tests suggest its prose quality still trails Anthropic's Claude Opus 4.7.

Benchmarks & Performance

GPT-5.5 Instant shows substantial improvements over its predecessor GPT-5.3 Instant across multiple evaluation dimensions: | Benchmark | GPT-5.3 Instant | GPT-5.5 Instant | Change | |---|---|---|---| | AIME 2025 (Math) | 65.4% | 81.2% | +24.2% | | MMMU-Pro (Multimodal Reasoning) | 69.2% | 76.0% | +9.8% | | HealthBench | 49.6 | 51.4 | +3.6% | | HealthBench Professional | 32.9 | 38.4 | +16.7% | | Hallucinations (High-Stakes) | Baseline | −52.5% | Major reduction | | Inaccurate Claims (User-Flagged) | Baseline | −37.3% | Major reduction | Additional benchmark data from independent evaluations: | Metric | Score | Context | |---|---|---| | Artificial Analysis Intelligence Index | 34 (est.) | #77 of 186; median of comparable models: 30 | | GPQA Diamond | 84.6% | #84 across tracked models | | SciCode | 50.3% | #33 across tracked models | | Humanity's Last Exam (HLE) | 20.3% | #94 across tracked models | | Terminal-Bench v2.1 (Hard) | 42.4% | Agentic coding and terminal use | | IFBench | 71.5% | Instruction following | | τ²-bench (Tau²-bench) | 49.4% | Agentic tool use | The model scores above average on the Artificial Analysis Intelligence Index (34 vs median 30) but is classified as particularly expensive relative to comparable-performing models. Its SciCode score (50.3%, #33) indicates competitive coding capability, while its GPQA Diamond score (84.6%) reflects solid scientific reasoning. The HLE score (20.3%) shows room for improvement on extremely challenging reasoning tasks.

Detailed Comparison

**GPT-5.5 Instant vs GPT-5.3 Instant (Predecessor):** The direct predecessor was available for three months post-launch before retirement. GPT-5.5 Instant improves on virtually every benchmark — most dramatically in math (+24.2% on AIME 2025) and hallucination reduction (52.5% fewer on high-stakes prompts). Responses are 30.2% shorter and 29.2% fewer lines, with less verbose formatting and fewer gratuitous emojis. The personalization system is significantly more capable, surfacing 10 distinct user patterns vs 7 in testing. Pricing is identical ($5/$30 per 1M tokens). **GPT-5.5 Instant vs Claude Opus 4.7 (Anthropic):** Tom's Guide ran a 7-category head-to-head and Claude won all seven categories. Where Claude excels: long-form writing quality, paragraph-style prose, and detailed reasoning with explanations. Where GPT-5.5 Instant is competitive: token efficiency (72% fewer output tokens on equivalent tasks per MindStudio), agentic coding, speed, and personalization/memory features. Claude Opus 4.7 offers superior factual accuracy in independent testing. For developers, GPT-5.5 Instant's lower token count translates directly to cost savings on high-volume API workloads. **GPT-5.5 Instant vs Gemini 3.1 Pro (Google):** Gemini 3.1 Pro outperforms on multimodal/image analysis tasks and is described as competitive on overall image understanding. GPT-5.5 Instant edges ahead on memory/personalization capabilities and token efficiency. Both models are available as defaults in their respective consumer products. Pricing comparison is incomplete based on available sources. **Key Specification Comparison:** | Feature | GPT-5.5 Instant | GPT-5.5 (Full) | |---|---|---| | Context Window | 400K (API) / 128K (ChatGPT) | 1.05M+ (922K input, 128K output) | | Max Output | 128K tokens | 128K tokens | | Input Price | $5.00/1M | $5.00/1M | | Output Price | $30.00/1M | $30.00/1M | | GPT-5.5 Pro | N/A | $30 input / $180 output per 1M | | Use Case | Everyday chat, quick Q&A | Agentic coding, multi-step tasks | | API Alias | `chat-latest` | `gpt-5.5` |

Community Feedback

The developer and user community response has been mixed but largely positive on specific dimensions. Independent reviewer Sara Oliveira (The New Stack) tested three of OpenAI's claims and found that the accuracy claim fully held up — GPT-5.5 Instant correctly handled a tricky Claude context window question that GPT-5.2 hallucinated. However, the conciseness claim was only partially validated: in direct testing, GPT-5.2 was actually more concise on three test prompts, though GPT-5.5 was more conversational and natural in tone. PrimeAICenter's 3-day testing confirmed meaningful improvements in medical question handling (correctly explaining ibuprofen/aspirin interaction nuances) and financial queries. The self-correction mid-response capability was highlighted as a genuine surprise — the model catching and correcting its own math errors in real-time. For API builders, Pickuma flagged a significant concern: the silent default swap from GPT-5.3 to GPT-5.5 on the `chat-latest` alias means teams without robust evaluation harnesses may experience unexplained behavior changes. The recommendation is to pin versioned model IDs in production and use `chat-latest` only for staging. Response format drift (more bulleted lists by default) was noted as potentially breaking downstream parsers. Writesonic research revealed an unexpected citation behavior shift: brand websites are cited only 6% of the time (down from 13.4%), while Reddit became the most-cited domain by a 3× margin. This has significant implications for SEO and content marketing professionals. Additionally, reports suggest 16% of Instant prompts are being silently routed to the Thinking tier even with auto-switch disabled. The personalization/Memory Sources feature received cautious praise. Users appreciate the transparency of seeing which past chats shaped a response, but concerns exist about the model pulling outdated context and giving misdirected answers. The Gmail integration, initially rolling out to Plus/Pro web users, has raised privacy questions that OpenAI has partially addressed with opt-in controls and the ability to delete cited context.

Use Cases

**1. Everyday Factual Q&A and Research:** GPT-5.5 Instant's primary use case — and where it excels most. The 52.5% reduction in hallucinations on high-stakes topics (medicine, law, finance) makes it meaningfully more trustworthy for factual lookups. Example: asking about drug interactions now produces nuanced, calibrated answers distinguishing known risks from uncertainties, whereas prior versions gave confidently wrong blanket statements. Choose over alternatives when speed matters more than exhaustive depth, and when factual reliability is critical. **2. Personalized Productivity and Planning:** With Memory Sources enabled, GPT-5.5 Instant can draw on past conversations, uploaded files, and connected Gmail to provide contextually tailored recommendations. Example: asking for restaurant or activity suggestions now factors in stated preferences from previous chats rather than returning generic options. This works best for users who have built up a ChatGPT history. Choose over Claude (which lacks comparable memory infrastructure) when persistent assistant context is valuable; choose over Gemini when cross-referencing email context is needed. **3. Quick Coding Assistance and Debugging:** For non-developers and lightweight coding tasks — modifying existing scripts, debugging errors, understanding unfamiliar code — GPT-5.5 Instant provides tighter, less invasive responses than its predecessor. It's less likely to introduce unnecessary changes while fixing issues. The SciCode score (50.3%) and Terminal-Bench score (42.4%) confirm competent coding ability. Choose over alternatives for quick coding help where latency matters; switch to Claude Opus 4.7 or the full GPT-5.5 for complex agentic coding pipelines and architecture-level work. **4. Image and Chart Analysis:** With MMMU-Pro at 76% (up from 69.2%), GPT-5.5 Instant improved notably on visual reasoning tasks. It's effective for analyzing product photos, handwritten diagrams, charts, and screenshots alongside text queries. Example: uploading a handwritten math problem and getting a corrected solution with self-correction of initial errors. Choose over alternatives for integrated text+image queries in a conversational flow; Gemini 3.1 Pro may still be preferred for pure vision-heavy workloads.

Latest News

- **May 5, 2026:** GPT-5.5 Instant officially launched as ChatGPT's default model across all tiers (Free, Plus, Pro, Go, Business, Enterprise). Replaced GPT-5.3 Instant which remains available for paid users for three months. - **May 7, 2026:** GPT-5.5 Instant integrated into Microsoft 365 Copilot, appearing as 'GPT-5.5 Quick response' in the model selector. - **May 2026:** Available in GitHub Copilot with a 7.5× premium request multiplier at launch (promotional pricing). - **May 2026:** GPT-5.5 Instant available via API as the `chat-latest` alias. Pricing: $5.00 input / $30.00 output per 1M tokens; cached input at $0.50/1M (90% discount). - **May 2026:** Memory Sources feature launched, showing users which past chats, saved memories, or connected files shaped personalized responses. Users can delete or correct cited entries. - **June 9, 2026:** Personalization improvements rolled out to ChatGPT Go & Free tiers (with reduced past-chat context on Free tier). - **July 9, 2026:** GPT-5.6 (successor) announced — GPT-5.5 is now the predecessor model in the lineage. GPT-5.6 became the preferred model in Microsoft 365 Copilot. - **Ongoing:** OpenAI has not announced pricing changes for GPT-5.5 Instant. The three-month window for GPT-5.3 Instant retirement via API is active.

The headline improvements center on factual reliability: OpenAI reports 52.5% fewer hallucinated claims on high-stakes prompts and 37.3% fewer inaccurate claims on conversations users previously flagged for errors. These claims are supported by meaningful benchmark jumps — AIME 2025 went from 65.4% to 81.2%, MMMU-Pro from 69.2% to 76%, and HealthBench Professional from 32.9 to 38.4. The model also introduced real-time self-correction mid-response, where it can catch and fix its own reasoning errors before completing an answer.

Beyond raw intelligence, GPT-5.5 Instant's most distinctive innovation is its Memory Sources personalization system. The model draws on past conversations, uploaded files, and connected Gmail (for Plus/Pro users) to provide contextually tailored responses, then transparently shows which sources were referenced so users can edit or remove them. Combined with a 30.2% reduction in word count and fewer unnecessary formatting elements, the model represents OpenAI's push toward making AI feel less like a tool and more like a persistent, trusted assistant. However, at $5/$30 per million tokens via API, it remains premium-priced, and independent tests suggest its prose quality still trails Anthropic's Claude Opus 4.7.

Sources

Analysis generated: 2026-07-17