Summary — Context and Chronology

OpenAI has commercially launched GPT-5.4 across its API and Codex surfaces, offering two tiers meant to serve everyday subscribers and heavier professional users. The company framed the release as a step from single-turn responses toward sustained, multi-step workflows that can operate across desktop applications, browsers, and large document collections. Paid subscribers are routed to the model more frequently; free-tier access is scoped and may be auto-routed only for select queries.

At a technical level, GPT-5.4 couples greatly expanded context capacity with native computer-control primitives: the model can issue UI events, drive browser automation, and integrate with desktop apps to run end-to-end sequences. OpenAI has exposed a firm ceiling up to 1,000,000 tokens with a default compaction floor around 272,000 tokens; requests that exceed the compaction threshold are billed under a separate, higher-cost regime (a 2× multiplier for affected inputs).

OpenAI presented benchmarks showing meaningful wins on agentic and web-navigation tasks: a reported 17 percentage-point gain on a persistent web-browse test and sizeable single-task navigation improvements, with some desktop navigation scores moving from the high 40s into the mid 70s on an internal suite. The company also emphasized an indexed tool-discovery mechanic (tool-search) that retrieves tool definitions on demand, which OpenAI says cut token consumption by roughly 47% on a 250-task sample.

For developers, Codex gained latency and workflow controls including a /fast mode (up to 1.5× faster on supported workloads) and an experimental Playwright integration to enable visible, interactive debugging. OpenAI is bundling embedded spreadsheet connectors for Excel and Google Sheets and shipping finance-oriented skills and partner connectors to market-data vendors—moves explicitly aimed at automating repeated professional tasks such as modeling, presentations, and multi-document research.

Complementing GPT-5.4, OpenAI’s recent Responses API updates matter operationally: server-side compaction preserves a small active state from long histories, hosted shell containers provide managed runtimes with persistent filesystems and network reachability, and Skills manifests package procedural abilities as versioned artifacts. Those platform pieces reduce bespoke engineering for resumable agents and make it practical to run sessions that span millions of tokens and hundreds of tool calls without immediate degradation—while also increasing coupling to OpenAI’s hosted execution and governance surface.

Not all advancements are unique to OpenAI. Anthropic’s Opus 4.6 also reported a one‑million‑token context window and introduced durable Task and agent primitives that persist multi-step plans as DAGs for resumability and audit—an approach that emphasizes portable, auditable artifacts. These simultaneous vendor moves demonstrate convergence on long-context, resumable agents, but they diverge in execution: OpenAI pushes a tightly integrated hosted stack (compaction, shells, Skills), while competitors stress on‑disk task primitives, adaptive refusal/safety behavior, and model-agnostic connector patterns.

Another adjacent product thread is OpenAI’s low-latency Codex variants: the Codex-Spark preview routes highly interactive workloads to Cerebras wafer-scale engines for subsecond responsiveness and ships alongside a native macOS Codex client that supports background agents and parallel orchestration. That hardware pairing promises latency gains for iterative developer workflows but raises supplier concentration and supply‑chain exposure as demand scales.

Commercially, OpenAI priced GPT-5.4 and a higher-end Pro tier at materially higher per-token rates than many peers and layered the long-context surcharge to create distinct cost regimes. The company contends that efficiency gains (tool-search, compaction) will offset sticker prices for common workflows, but the billing architecture will push builders to rethink prompt design, staging, and when to run large end‑to‑end agent passes versus iterative compacted calls.

Practically, organizations should expect stronger automation potential for bounded finance and analytics tasks but uneven generalization in open or adversarial settings; the showcased benchmarks favor tasks amenable to tool access and orchestration. Enterprises adopting end-to-end agents must invest in sandboxing, immutable logs, secrets controls, and human-in-the-loop gates because persistent runtimes and desktop control broaden the audit surface and elevate compliance risk.