From Assistant to Partner: How GPT-5.5 Quietly Reshapes How We Work With Machines

*English translation and structured summary of a Chinese forum post from zhichai.net.*

Introduction: When "Chatting" Is No Longer the Goal

The author compares previous AI models to a diligent intern who keeps running back with questions, while GPT-5.5 is likened to an experienced project manager on a construction site: it plans, executes, uses tools, detects and corrects errors, and finishes multi-step tasks without constant supervision. OpenAI describes it as "a new class of intelligence built for real work." After a rapid string of updates (GPT-5.3, 5.3 Instant, 5.4 Thinking, 5.4 Pro) that drew muted reactions, GPT-5.5 is presented as a directional shift: from question-answering machine to trustworthy collaborator.

Core Change: From Frequent Escalation to Full Delegation

Earlier models stalled mid-task asking "A or B?" or requesting more context. GPT-5.5 resolves ambiguity on its own, calls tools to verify, advances to the next step, and self-checks outputs — like handing the kitchen keys to a competent chef rather than phoning in takeout.

Coding: System-Level Judgment and Efficiency

• Terminal-Bench 2.0 (complex command-line workflows): 82.7%, vs 75.1% for GPT-5.4, and roughly 13–14 points above Claude Opus 4.7 and Gemini 3.1 Pro.
• Expert-SWE (long tasks that take senior engineers ~20 hours): 73.1%, up 4.6 points from 68.5%, with fewer tokens consumed and no loss of speed.
• Dan Shipper (Every) fed a broken post-launch bug state to GPT-5.5; the 5.4 model missed it, but 5.5 correctly concluded that part of the system needed a rewrite — evidence of "system sense," not just cleverness.
• MagicPath CEO Pietro Schirano reported GPT-5.5 merged a branch with hundreds of frontend changes into main in 20 minutes, resolving all conflicts in one pass. NVIDIA engineers described losing internal preview access as "like losing an arm."

Knowledge Work Across 44 Professions

• GDPval (44 professions: lawyers, financial analysts, sales, marketing, engineers): 84.9%, 4.6 points above Claude and 17.6 above Gemini.
• Tau2-bench (complex customer-service workflows, no prompt tuning): 98.0% vs 92.8% for GPT-5.4.
• OfficeQA Pro: 54.1% vs 18.1% for Gemini 3.1 Pro — a nearly 36-point gap.
• OpenAI internal usage: over 85% of employees use Codex with GPT-5.5 weekly; the finance team reviewed 24,771 K-1 tax forms (71,637 pages) two weeks earlier than the prior year; a GTM colleague saves 5–10 hours weekly on business reports.

Scientific Research: From Assistant to Participant

• FrontierMath Tier 4: 35.4% (Gemini 3.1 Pro: 16.7%, Claude Opus 4.7: 22.9%).
• Immunology professor Derya Unutmaz (Jackson Laboratory) used GPT-5.5 Pro to analyze a dataset of 62 samples and nearly 28,000 genes' expression data, producing a detailed research report in days rather than months.
• An internal version of GPT-5.5, with custom tools, produced a new proof related to a Ramsey number, later verified in the Lean proof assistant.

Long Context, Reasoning, and Tool Use

• Long context (512K–1M): 74% vs 36.6% for GPT-5.4.
• ARC-AGI-2: 85.0%, up from 73.3%.
• MCP Atlas (tool calling): Claude Opus 4.7 leads, 79.1% vs 75.3% — competition remains real.
• GPT-5.5 Pro uses Parallel Test Time Compute (running multiple reasoning paths and synthesizing); on a biochemistry implicit-knowledge evaluation it scored 81.67%, above the expert baseline, while the standard version fell below it.
• Co-designed with NVIDIA GB200/GB300 NVL72 systems; Codex analyzed production traffic and wrote a custom load-balancing algorithm improving token generation speed by over 20%.

Safety Evaluation

Bio/chemical and cyber capabilities are rated "High" under OpenAI's Preparedness Framework, triggering the strictest safeguards: ~200 early partners testing pre-release, external evaluations by SecureBio and US CAISI, a bio bug bounty, and Apollo Research alignment assessments finding no significant deception. Capture the Flags: 88.1% (vs 83.7%); CyberGym: 81.8% vs 79.0%.

Pricing and the Pro Version

The standard GPT-5.5 suits most users (available with Plus subscription); Pro adds compute for deeper reasoning on high-stakes tasks (research, legal, medical). They share one base model with different inference depth.

Conclusion: The Lights Finally Stay On

Echoing the history of electrification, the author argues GPT-5.5 marks the pivot from "assisting thinking" to "doing the work for you." It is not perfect — Claude still leads on MCP Atlas and real-world deployment challenges remain — but the direction is clear: AI as a trusted work partner rather than a chat tool.

References

1. OpenAI official GPT-5.5 release documentation and positioning statements. 2. Public benchmark reports: Terminal-Bench 2.0, Expert-SWE. 3. Developer case studies shared by Dan Shipper, Pietro Schirano, and others. 4. OpenAI internal employee usage statistics; GDPval and Tau2-bench data. 5. FrontierMath, ARC-AGI-2, and safety evaluation reports (Apollo Research, etc.).