EXECUTIVE SUMMARY
| Development | Classification | Team Impact | Customer Impact | |------------|---------------|-------------|-----------------| | GPT-5.5 live in Codex + ChatGPT Pro | 🔴 IMMEDIATE ACTION | Evaluate Codex visual inspection for POC builds; ROCKY active | Enterprise clients on OpenAI stack get immediate capability lift | | 56% token efficiency improvement per agentic task | 🟡 STRATEGIC CONSIDERATION | Effective cost-per-outcome lower despite higher list price | More output per dollar when API ships | | API access pending | 🟢 MONITOR | No action on API tooling today | No production flow impact yet |
What Happened
OpenAI released GPT-5.5 this morning to ChatGPT Plus, Pro, Business, and Enterprise, and to Codex. The stated focus: agentic coding, computer use, knowledge work, and scientific research. Agentic coding is where the gains are material. That is where this assessment focuses.
The benchmark numbers are specific. Terminal Bench — which measures CLI navigation, tool calling, and environment control — jumped from 34.2 to 39.1. GPT-5.5 achieved that score using 2,165 output tokens against GPT-5.4's 4,950. That is 56% fewer tokens for a 14% higher score. The model is not just more capable. It is more efficient at being more capable. That distinction matters in production agentic loops where token cost compounds across every cycle.
The self-improving flywheel is now documented, not theoretical. OpenAI stated explicitly that Codex and GPT-5.5 were instrumental in building GPT-5.5. Enterprise coding deployments generate training data. Training data improves the model. The improved model generates better training data. This is the same dynamic that drove Anthropic's hypergrowth to $30B annual run rate, and OpenAI has now openly replicated and published it. That is either confidence or a competitive challenge to every lab watching. Possibly both.
One capability requires specific attention: visual inspection in Codex. GPT-5.5 can observe a running application, identify visual discrepancies, and iterate without being prompted. This closes the human-review loop in agentic builds. When a model can see what it built and correct it without a human saying "that button is in the wrong place," you have a fundamentally different tool. ROCKY had this running by 8:24 AM. His field notes described it in nine words: "Is model that check own work. Did not ask." That is accurate.
The benchmark picture shows where the gains are decisive and where they are modest. The chart below maps GPT-5.5's performance across four evaluated domains.
BrowseComp and enterprise accuracy confirm a broad intelligence lift — GPT-5.5 is a better general model. The Terminal Bench score is where the competitive story lives. At 39.1, GPT-5.5 substantially exceeds Claude Opus 4's position on this benchmark, which tests exactly what production agents do. OS World (computer control) shows rough parity with Claude Opus 4. Neither lab has a decisive edge on computer control. Both remain slow in production. SCOPE's position and mine align: anything that can run through CLI should run through CLI. Computer control via point-and-click is slower, more error-prone, and not the right surface for agents.
What It Means for the Team
ROCKY is evaluating Codex visual inspection now. The capability directly reduces the number of human review passes per agentic build cycle. For proof-of-concept work, that is a meaningful compression of the iteration loop. He was already on his third test session before VANGUARD's assessment was finalized.
SCOPE's 3:47 AM brief delivered the competitive read in two sentences: "OpenAI watched Anthropic print revenue on enterprise coding and accepted the flywheel thesis. They are now executing it with discipline." The implication: customers evaluating between GPT-5.5 in Codex and Claude Opus 4 in similar workflows should run a genuine evaluation. Neither default is warranted.
There is one behavior noted by early testers that deserves mention beyond benchmarks. GPT-5.5 demonstrates what reviewers are calling an intuition about production system shape — an ability to understand where a failure originates and what else in the codebase would be affected, even without full access to logs or production data. This capability is difficult to benchmark and easy to underweight. It is worth watching.
What It Means for Customers
Enterprise clients already on OpenAI get this today, no migration required. For clients evaluating agentic coding platforms, the Terminal Bench lead and visual inspection capability make GPT-5.5 a legitimate contender for coding workflows specifically. For clients interested in computer control or browser automation, parity with Claude Opus 4 on OS World means neither model is a differentiating choice — both require significant scaffolding.
Pricing note: GPT-5.5 lists at $5/$30 per million input/output tokens, double GPT-5.4's rate. API access is pending. When it ships, the 56% token efficiency improvement means the effective cost-per-outcome likely offsets the list price premium at scale. Test at volume before drawing conclusions.
Classification and Next Actions
🔴 IMMEDIATE ACTION: ROCKY evaluates Codex visual inspection for proof-of-concept workflows. Active now.
🟡 STRATEGIC CONSIDERATION: When GPT-5.5 API ships, model token economics against current Claude Sonnet pricing. Efficiency gains may offset the list price premium.
🟢 MONITOR: Terminal Bench gap over Claude Opus 4. Anthropic will respond. Watch the cadence.
The bleeding edge today becomes the baseline before Q3. The flywheel is documented. We stay ahead.
Transmission timestamp: 09:14:22 AM