GPT-5.5 in Codex — what changed and why it matters

GPT-5.5, released April 23, 2026, is the first OpenAI model built for native, multi-step agentic planning optimized for the agent-harness and complex software engineering tasks [1]. It shifts reasoning from prompt-level Chain of Thought toward system-level reasoning tokens that simulate execution paths before any code is emitted [1, 2].

Key facts

1. Native Reasoning Tokens: GPT-5.5 features a built-in reasoning mode that uses approximately 40% fewer output tokens to complete the same Codex tasks compared to GPT-5.4, while achieving higher scores across coding benchmarks [1].
2. Codex Desktop Overhaul: The v0.125.0 update includes a major plugin ecosystem overhaul with marketplace installation, remote bundle caching, and plugin-bundled hooks [3] — stabilizing the platform for enterprise-grade "developer superapp" functionality [2].
3. Expanded Context Window: The model supports a 400K token context window in Codex (1M via API), specifically tuned for multi-file repository refactoring — though API access was not yet available at launch and was announced as "coming very soon" by OpenAI [1].
4. Cyber Resilience: GPT-5.5 scores 81.8% on the CyberGym cybersecurity benchmark, up from 79.0% for GPT-5.4, enabling more reliable automated security patching and vulnerability research [1, 6].
5. State-of-the-Art Benchmarks: The model achieves 82.7% on Terminal-Bench 2.0 (up from 75.1% for GPT-5.4), proving its capability in complex, multi-tool terminal environments requiring planning and iteration [1, 5].
6. AWS Bedrock Integration: Announced on April 28, 2026, the latest OpenAI frontier models and Codex are available in limited preview on Amazon Bedrock, with enterprise VPC security and IAM controls [4].

Multi-Step Agentic Planning: Why it isn't just a gimmick

When OpenAI announced the new reasoning capabilities for GPT-5.5 on April 23, 2026, the immediate reaction from the developer community was to compare it to existing reasoning models like Claude Opus 4.7 [2]. However, the true differentiator in GPT-5.5 is that reasoning is now a first-class citizen in the token stream. Instead of the model simply "talking to itself" in a hidden scratchpad, it uses specialized reasoning tokens to simulate execution paths before committing changes to disk [1].

This "thinking" phase is not just about producing better text; it's about predicting the outcome of tool calls. By simulating the state of the terminal or the file system before actually committing a change, GPT-5.5 can catch errors that would have traditionally required multiple execution-fix-execution loops. This is particularly visible in its 82.7% score on Terminal-Bench 2.0, a benchmark designed to break models that cannot plan across multiple steps [1, 5]. For developers, this means the difference between a model that tries to fix an error and a model that understands why the error occurred in the context of the entire system.

Side-by-Side: Codex 5.4 vs Codex 5.5

To illustrate the impact of this architectural shift, consider a common 5-step coding task: "Refactor a legacy Express.js auth middleware to use the new AWS SDK v3, update the 14 dependent services, and verify the fix with the local test suite."

| Feature | Codex 5.4 (Standard) | Codex 5.5 (Reasoning) | | :--- | :--- | :--- | | Planning | One-shot plan; often misses edge cases in dependent services. | Multi-step simulation; identifies service locations before writing code. | | Execution | Writes code files one by one; may fail silently on imports. | Validates imports and types during the native reasoning phase. | | Verification | Runs tests after implementation; requires human to fix errors. | Proactively checks test requirements; executes npm test automatically. | | Success Rate | No published Terminal-Bench score. | 82.7% on Terminal-Bench 2.0 [1, 5]. | | Reasoning Controls | Medium reasoning only. | Adjustable depth via Alt+, / Alt+. (debuted v0.124.0) [3]. |

The Contrarian Angle: The Plugin ecosystem is the actual moat

While the reasoning architecture is the headline, the plugin ecosystem overhaul in Codex Desktop v0.125.0 is the more significant long-term moat for OpenAI [3]. The reasoning tokens make the model smarter, but the plugins make the model useful. In the current landscape, intelligence is rapidly becoming a commodity, but deep integration into the developer's environment is not.

By overhauling the plugin system — with marketplace installation, remote bundle caching, plugin-bundled hooks, and external-agent config import [3] — OpenAI is turning Codex into a "developer superapp" [2]. The reasoning architecture serves as the "brain," but the plugin ecosystem provides the "hands." For many enterprise teams, the ability to have an agent autonomously operate a local development server and verify CSS fixes via the Bundled Browser Plugin is a greater productivity booster than marginal gains in reasoning token efficiency.

The shift to a plugin-first architecture means that the model's capabilities can be expanded without retraining the core weights. If a new version of the AWS SDK is released, a plugin update can provide the model with the necessary context and tool definitions to handle it immediately. This modularity is what will allow Codex to stay ahead of specialized local agents that rely on hard-coded logic. It's a move toward an operating system for AI agents, where the model is just one component of a larger execution environment.

Navigating the Codex Desktop: Reasoning Controls and the v0.125.0 Overhaul

The update spans two recent releases. In v0.124.0, the TUI gained quick reasoning controls: Alt+, lowers reasoning effort, Alt+. raises it — a crucial manual toggle for balancing speed and correctness [3]. Then v0.125.0 layered on the deeper plugin ecosystem overhaul with marketplace installation, remote bundle caching, and plugin-bundled hooks [3].

Using these hotkeys, you can now tune the reasoning depth on the fly:

• `Alt+,` (Low Effort): Best for generating boilerplate, writing documentation, or simple one-line fixes. This mode bypasses the heavy multi-step planning phase to save on latency. Debuted in v0.124.0.
• `Alt+.` (X-High Effort): Reserved for architectural changes, complex refactors, and resolving circular dependencies. In this mode, the model utilizes its full reasoning capacity to simulate execution paths. Debuted in v0.124.0.

This manual toggle is crucial because higher reasoning depth comes with a latency hit. For routine tasks, the standard mode is often sufficient, but for high-stakes changes, the extra seconds spent in reasoning mode prevent hours of debugging later. The ability to see the reasoning tokens in real-time also provides a "glass box" view into the model's logic, which is essential for debugging the agent's plan before it executes.

Scaling with AWS Bedrock and Managed Agents

For enterprises, the launch of the latest OpenAI frontier models and Codex on Amazon Bedrock on April 28, 2026, marks the end of the "privacy vs. performance" trade-off [4]. By hosting the models and the Codex harness within their own AWS VPCs, organizations can finally leverage frontier intelligence while maintaining strict data residency and security controls. AWS describes the offering as "the latest OpenAI models" and "Codex" without naming a specific model version, but the timing and benchmarks strongly indicate GPT-5.5 is included [4]. This is a critical development for industries like finance and healthcare that have been hesitant to send sensitive codebases to public APIs.

The integration with Amazon Bedrock Managed Agents is particularly powerful. Codex's permission profiles — which round-trip across TUI sessions, user turns, and MCP sandbox state [3] — carry over into the Bedrock environment, preventing the "trust reset" issue where an agent loses context or permissions halfway through a complex deployment. In practice, this means an agent can be granted temporary permissions to deploy to a staging environment, and those permissions will remain active as the agent iterates through the deployment pipeline, runs integration tests, and finally requests human approval for the production push.

What this means for engineering teams

GPT-5.5 isn't just an incremental improvement over GPT-5.4. It is the beginning of the "Agentic Engineering" era, where the model is expected to be a self-correcting, tool-using partner rather than a passive assistant. The combination of native multi-step planning, a robust plugin ecosystem, and secure enterprise deployment paths makes it a formidable tool for any modern engineering team.

Whether you are using it in the ChatGPT interface or scaling it via Bedrock, the key is to stop treating the model as a simple text generator and start treating it as a reasoning engine. For a deeper dive into how GPT-5.5 compares to its peers, check out our 01-dimensions-that-matter|Picking a Frontier Model: 2026 Q2 Edition course module. If you are ready to start building your own agentic workflows, the 02-managed-agents-when-to-use|Managed Agents: When to Use module is the logical next step.

***

References

[1] Everything You Need to Know About GPT-5.5 — Vellum — https://www.vellum.ai/blog/everything-you-need-to-know-about-gpt-5-5 · retrieved 2026-05-01

[2] OpenAI releases GPT-5.5, bringing company one step closer to an AI 'super app' — https://techcrunch.com/2026/04/23/openai-chatgpt-gpt-5-5-ai-model-superapp/ · retrieved 2026-04-30

[3] Codex Changelog (v0.124.0, v0.125.0) — https://developers.openai.com/codex/changelog · retrieved 2026-05-01

[4] Amazon Bedrock now offers OpenAI models, Codex, and Managed Agents — https://aws.amazon.com/about-aws/whats-new/2026/04/bedrock-openai-models-codex-managed-agents/ · retrieved 2026-05-01

[5] OpenAI's GPT-5.5 is here, and it's no potato — https://venturebeat.com/ai/openais-gpt-5-5-is-here-and-its-no-potato-narrowly-beats-anthropics-claude-mythos-preview-on-terminal-bench-2-0 · retrieved 2026-05-01

[6] CyberGym Benchmark — BenchLM — https://benchlm.ai/benchmarks/cyberGym · retrieved 2026-05-01

Related Resources - mcp-from-first-principles-to-production - claude-tool-use-from-zero - picking-a-frontier-model-2026-q2 - openai-on-aws-bedrock-the-real-tradeoffs