GPT-5.5 Instant Arrives: Smarter, Clearer, More Personalized AI

OpenAI quietly switched the default ChatGPT model to GPT‑5.5 Instant , fully replacing the previous GPT‑5.3 Instant.

Three Key Changes

1. Smarter – reduced hallucinations

Internal high‑risk evaluation (medical, legal, finance) shows hallucination assertions down 52.5% versus GPT‑5.3.

On user‑flagged “fact‑problem” dialogs, error rate drops 37.3% .

Typical example: a user uploads an algebra sketch. GPT‑5.3 declares the problem unsolvable, while GPT‑5.5 identifies the mistaken expression (should be x²‑3x‑6), back‑traces the error, and returns the correct root (3+√33)/2. The author notes this self‑correction as a core quality of the default model.

2. Clearer – less verbosity

OpenAI’s wording: “reduce verbosity and overformatting”.

Example: user asks how to tell a coworker to stop chatting. GPT‑5.3 replies with four paragraphs, a checklist, and emojis; GPT‑5.5 gives five tiered suggestions in a single sentence, using 30.2% fewer words and 29.2% fewer lines .

Long‑time users note the reduction of emojis and enthusiastic tone, resulting in a more professional feel.

3. More Personalized – visible memory sources

Instant now automatically incorporates past chats, uploaded files, and connected Gmail into responses.

Example: a user asks for a new tea place. GPT‑5.3 gives a generic recommendation for San Francisco; GPT‑5.5 knows the user frequents Asha Tea House, prefers Taiwanese high‑mountain tea, and directly suggests the next venue.

In each personalized reply you can see which context (saved memory, past dialogue) was used; you can delete, edit, or start a temporary chat with no memory.

The author rates this transparency highly, warning that loss of visibility could lead to loss of control.

Usability and Rollout

Rollout began today, fully replacing the default model.

API name: chat-latest.

Paid users can keep GPT‑5.3 for three months via manual model selection.

Personalization (past chats/files/Gmail) initially for Plus/Pro web, later for Free/Go/Business/Enterprise.

Memory sources will be opened gradually across all tiers.

Benchmark Highlights (selected)

Terminal‑Bench 2.0 (command‑line agent): 82.7% (GPT‑5.5) vs 75.1% (GPT‑5.4).

GDPval (44 professions): 84.9% vs 83.0% (GPT‑5.4).

OSWorld‑Verified (real‑computer ops): 78.7% vs 75.0%.

FrontierMath Tier 4 (advanced math): 35.4% vs 27.1%.

ARC‑AGI‑2 (Verified): 85.0% vs 73.3%.

τ²‑bench Telecom (customer‑service flow): 98.0% vs 92.8%.

CyberGym (cybersecurity): 81.8% vs 79.0%.

Graphwalks BFS 1 mil f1 (long context): 45.4% vs 9.4%.

The main SOTA focus is agent programming, long context, computer operation, cybersecurity, and advanced mathematics , often achieving results with fewer tokens.

Agent Programming Strength

Terminal‑Bench 2.0: 82.7% – multi‑step planning and tool collaboration resembling a real engineer.

SWE‑Bench Pro: 58.6% – end‑to‑end resolution of real GitHub issues in a single attempt.

Internal Expert‑SWE (median 20‑hour human task) surpasses GPT‑5.4.

Internal Usage at OpenAI

85% of employees use Codex weekly.

Finance team processed 24,771 K‑1 tax forms (71,637 pages), finishing two weeks early.

PR team built an automatic Slack agent to filter low‑risk speaking requests.

Market expansion team saved 5‑10 hours per person per week via weekly report automation.

Pricing Strategy

GPT‑5.5: $5 / M input, $30 / M output, 1 M context.

GPT‑5.5 Pro: $30 / M input, $180 / M output, 1 M context.

Codex subscription included in Plus/Pro/Business/Enterprise/Edu/Go with 400 K context.

Additional Options

Batch / Flex at half price for non‑urgent tasks.

Priority mode at 2.5× price for faster response.

Codex fast mode: 1.5× speed, 2.5× cost (same pricing model as Priority).

Author’s Perspective

Release timeline: April 23 – main GPT‑5.5 (Pro/paid priority); May 5 – Instant upgrade (default model, full rollout).

Underlying message: first deliver flagship capabilities to paying users, then cascade the benefits to all users.

For ordinary users the two most noticeable changes are:

More accurate, less fluff – fewer “ChatGPT‑style” quirks, more professional tone.

Visible personalization – you can see, edit, or disable the memory sources used.

Potential drawbacks:

The model feels “colder” with fewer emojis and less enthusiasm.

Stricter security filtering in cybersecurity scenarios may initially limit some professional users.

Conclusion

The keyword for GPT‑5.5 is not “bigger” but more accurate, more efficient, more capable at work . The Instant upgrade spreads these gains to everyone, even free users.