OpenAI Launches GPT-5.2 With Three Tiers — Matches Human Experts on 70% of Business Tasks

OpenAI has shipped GPT-5.2, its latest frontier model, structured across three performance tiers: Instant (fast, cost-efficient), Thinking (deeper chain-of-thought reasoning), and Pro (research-grade, maximum capability). The release is rolling out now to ChatGPT paid subscribers with no change to existing subscription pricing.

Benchmark Results

The headline number from OpenAI’s internal GDPval benchmark — which measures performance across 44 real business tasks — puts GPT-5.2 at 70.9% human-expert-level match, up sharply from GPT-5.1’s 38.8%. That near-doubling in a single generation is the most aggressive capability jump OpenAI has claimed on this metric.

Additional benchmark highlights reported by OpenAI:

• SWE-Bench (software engineering): improved score versus GPT-5.1, though OpenAI did not disclose the exact figure at launch
• ARC-AGI (abstract reasoning and generalization): notable improvement in problem-solving
• Code debugging, long-context handling, and multi-step project execution all show measurable gains per OpenAI’s internal evaluations

Capabilities

GPT-5.2 adds or refines: spreadsheet generation, presentation building, production-level code writing, image perception, long-context document understanding, and multi-step project orchestration. OpenAI’s announcement specifically highlights that the model can “more reliably debug production code, implement feature requests, and refactor large codebases” — language clearly aimed at the developer market.

The model also demonstrates improved image understanding across all three tiers, not just the Pro variant.

API Pricing

For developers accessing via API:

• Input: $1.75 per million tokens
• Output: $14 per million tokens
• Cached inputs: 90% discount applied automatically

This is higher per-token than GPT-5.1, but OpenAI argues superior token efficiency means achieving equivalent output quality costs less in aggregate. Independent benchmarkers will verify this claim over the coming weeks.

Competitive Context

The release comes after CEO Sam Altman’s December “code red” memo warning the company to accelerate in response to Google’s Gemini 3 series. Altman later walked back the alarm, saying Gemini’s gains were “less significant than feared.” Regardless, the speed of GPT-5.2’s arrival — and the aggressive benchmark framing — signals OpenAI is treating the race for frontier model supremacy as an existential priority.

Gemini 3.1 Pro currently leads 13 of 16 major tracked benchmarks and ties GPT-5.4 Pro on the Artificial Analysis Intelligence Index. GPT-5.2 does not appear to close that gap on all fronts, but the GDPval score — measuring practical business utility rather than pure reasoning — is a deliberate narrative play targeting enterprise customers.

What This Means for Developers

The Instant tier will be the default for most API integrations — lower latency, lower cost. Thinking tier is the relevant one for agentic pipelines that require multi-step reasoning. Pro tier targets research applications and complex code generation workloads where cost is secondary.

OpenAI’s emphasis on code quality improvements is notable: the company is clearly targeting Cursor, Windsurf, and other AI coding tools that currently route significant token volume to competitors. Whether GPT-5.2 wins back coding workloads depends on how it performs on real codebases in the hands of developers — not OpenAI’s own evals.