GPT-5.5 Instant

GPT-5.5 Instant is the new default ChatGPT model as of May 5, 2026, replacing GPT-5.3 Instant across all tiers - Free, Plus, Pro, Team, and Enterprise. The "Instant" name follows the same pattern OpenAI has used since GPT-5.2: it's the fast conversational version of the current generation, tuned for everyday chat rather than extended reasoning chains. It runs on the same underlying architecture as GPT-5.5 (released April 23), but with reasoning effort set to minimal by default.

The release is two months after GPT-5.3 Instant, which drew criticism for safety regressions - its graphic violence and sexual content compliance scores fell notably versus its predecessor. GPT-5.5 Instant doesn't come with the same caveats. OpenAI's system card for this update doesn't flag similar regressions, and the model's improvements focus on factual accuracy rather than behavioral loosening.

One distinction worth noting upfront: unlike GPT-5.3 Instant, which had its own API pricing tier at $1.75/$14 per million tokens, GPT-5.5 Instant has no separate endpoint. In the API, you access this behavior through gpt-5.5 with reasoning_effort: "minimal" - and you pay the same $5/$30 rate as the full model.

Key Specifications

Benchmark Performance

OpenAI published Instant-specific metrics comparing GPT-5.5 Instant against GPT-5.3 Instant - the predecessor rather than a cross-model leaderboard. For standard reasoning benchmarks, the model shares its architecture with GPT-5.5, so the full model's scores apply where reasoning effort is the controlling factor.

The AIME improvement is the most striking number. A 15.8-point jump on a math competition benchmark suggests the underlying GPT-5.5 architecture delivers truly stronger reasoning - not just a tuning adjustment. The MMMU-Pro gains (+6.8 points) point to improved multimodal understanding across text-image tasks.

The hallucination figures need the standard caveat: these come from OpenAI's internal evaluations on high-stakes domains (medicine, law, finance) and haven't been independently replicated yet. The 52.5% figure is based on the same evaluation corpus used for previous Instant releases, which makes comparisons internally consistent, but doesn't tell you how the model performs on other hallucination benchmarks. See the hallucination benchmarks leaderboard for third-party tracking.

For broader positioning: GPT-5.5 (full model, same architecture) scores 93.6% on GPQA Diamond, 58.6% on SWE-Bench Pro, and 82.7% on Terminal-Bench 2.0. These numbers reflect high-reasoning-effort runs and aren't specific to the Instant configuration, but they establish the ceiling of the architecture.

OpenAI's internal factuality benchmarks for GPT-5.5 Instant versus GPT-5.3 Instant across medicine, law, and finance prompts. Source: decrypt.co

Key Capabilities

Hallucination reduction and factual accuracy. The headline claim is 52.5% fewer hallucinated claims on high-stakes prompts when the model uses web search, and 37.3% fewer on conversations users had previously flagged for factual errors. These improvements concentrate in medicine, law, and finance - domains where the previous Instant generation was already a weak point. The without-web improvement is narrower but real: the model's baseline factual recall is better than GPT-5.3 Instant independently of retrieval.

Personalization from memory and Gmail. GPT-5.5 Instant introduces "memory sources" - a transparency layer showing users which context ChatGPT used to personalize a response, including saved memories, past chats, and connected Gmail data. Plus and Pro subscribers get Gmail integration first; Free, Go, Business, and Enterprise users are in a later rollout wave. Users can inspect, delete, or correct memory sources, and temporary chats remain fully available for sessions that shouldn't touch memory at all.

Response conciseness. Responses are 30.2% shorter by word count and 29.2% shorter by line count compared to GPT-5.3 Instant. The formatting shift is notable: fewer unsolicited bullet points, fewer trailing emoji, less throat-clearing before the actual answer. This isn't just stylistic - shorter responses at the same accuracy level directly reduces output token cost in API use cases.

Multimodal input. Image understanding is available across all tiers. The MMMU-Pro improvement (+6.8 points) suggests real gains in complex visual reasoning tasks, not just basic captioning. Audio and video input aren't supported in the API; those remain on separate model tracks.

ChatGPT's default model updated to GPT-5.5 Instant on May 5, 2026, visible across Free, Plus, and Pro tiers. Source: decrypt.co

Pricing and Availability

GPT-5.5 Instant is available to all ChatGPT users as of May 5, 2026 - no tier gate. The API situation is different from previous Instant generations: there's no separate gpt-5.5-instant model ID. Developers who want the low-latency conversational configuration call gpt-5.5 with reasoning_effort: "minimal" and pay the standard $5/$30 per million tokens.

This pricing represents a significant change from GPT-5.3 Instant, which had its own API tier at $1.75 input / $14.00 output per million tokens. That gap matters for high-volume applications. If you were using GPT-5.3 Instant for cost efficiency in chat applications, switching to GPT-5.5 Instant means a roughly 2.9x increase in input costs and a 2.1x increase in output costs. For latency-critical use cases where GPT-5.5's deeper reasoning modes aren't needed, this is worth tracking.

GPT-5.3 Instant is deprecated but will remain available to paid ChatGPT users for approximately three months before retirement. API users can pin to explicit model snapshots to preserve GPT-5.3 Instant behavior while the window is open.

For cost comparison: Claude Opus 4.7 runs at higher per-token rates for reasoning workloads. At the conversational, low-reasoning-effort tier, GPT-5.5 Instant's $5/$30 pricing sits in the mid-range - cheaper than top-tier reasoning models but no longer the budget option the Instant line used to represent.

The new "memory sources" panel shows users which context GPT-5.5 Instant used to personalize a response - and allows deletions. Source: decrypt.co

Strengths and Weaknesses

Strengths

• Sizable hallucination reduction (52.5% on high-stakes prompts) with no documented safety regressions
• AIME 2025 score of 81.2 - a 15.8-point jump from GPT-5.3 Instant, showing real reasoning gains
• Concise responses by default: 30.2% fewer words, fewer unnecessary emoji and bullet points
• Gmail and memory personalization with full user controls (inspect, delete, or opt out via temporary chats)
• Available across all ChatGPT tiers right away, including Free tier

Weaknesses

• API pricing ($5/$30/M tokens) is 2.9x more expensive than GPT-5.3 Instant was for input tokens - no budget tier for Instant-class conversational work
• Hallucination benchmark numbers are OpenAI internal evaluations, not independently confirmed
• Gmail personalization limited to Plus and Pro initially; Free/Go/Business/Enterprise timeline not specified
• Context window in ChatGPT web UI is unspecified - the 1M token limit is API-only
• Knowledge cutoff remains December 2025 despite the May 2026 release

FAQ

What replaced GPT-5.3 Instant?

GPT-5.5 Instant, released May 5, 2026, is the new default ChatGPT model. GPT-5.3 Instant remains accessible to paid users for roughly 3 months before retirement.

Is GPT-5.5 Instant cheaper than GPT-5.3 Instant?

No. GPT-5.5 Instant is priced at $5/$30 per million tokens, the same as GPT-5.5. GPT-5.3 Instant was $1.75/$14 per million tokens, making GPT-5.5 Instant about 2.9x more expensive for input tokens.

What API model name do I use for GPT-5.5 Instant?

Use gpt-5.5 with reasoning_effort: "minimal" in the API. The chat-latest alias also resolves to this model. There's no separate gpt-5.5-instant model ID.

Does GPT-5.5 Instant have safety regressions like GPT-5.3 Instant did?

No documented regressions have been published for GPT-5.5 Instant. GPT-5.3 Instant had measurable drops in sexual content and graphic violence compliance; this update does not show similar issues in OpenAI's system card.

When will Free-tier users get Gmail personalization?

OpenAI hasn't given a specific timeline. The initial rollout covers Plus and Pro users on the web, with Free, Go, Business, and Enterprise described as "coming later."

Related Coverage

• GPT-5.5 - The full API model that shares GPT-5.5 Instant's architecture
• GPT-5.3 Instant - The predecessor model this replaces
• OpenAI Launches GPT-5.5 for Agents and Work - News coverage of the full GPT-5.5 launch
• GPT-5.5 Brings Mythos-Like Hacking to the Masses - Coverage of GPT-5.5's cybersecurity capabilities
• Chatbot Arena ELO Rankings - Human preference rankings across models
• Cost Efficiency Leaderboard - Per-token pricing comparison across all major models
• Hallucination Benchmarks Leaderboard - Third-party hallucination tracking

Sources

• GPT-5.5 Instant: smarter, clearer, and more personalized - OpenAI
• OpenAI releases GPT-5.5 Instant, new default ChatGPT model - TechCrunch
• GPT-5.5 Instant makes ChatGPT smarter, concise and reliable - SiliconAngle
• OpenAI rolls out GPT-5.5 Instant as new default ChatGPT model - TechBriefly
• OpenAI upgraded ChatGPT default model - Decrypt
• GPT-5.5 Model API Documentation - OpenAI
• OpenAI GPT-5.5 Instant - eWeek