benchrPricing breakdown
| Tier | Rate / 1M tokens |
|---|---|
| Standard input | $5.00 |
| Standard output | $30.00 |
| Cached input | $0.50 |
| Batch input | $2.50 |
| Batch output | $15.00 |
| Context window | 1,050,000 tokens |
| Max output | 128,000 tokens |
The 400K vs 1M context question
GPT-5 caps at 400,000 tokens — useful for most tasks, but a hard ceiling for large repository analysis, extended agent sessions, or processing full legal or technical documents without truncation. GPT-5.5 extends this to 1,050,000 tokens, matching Anthropic and DeepSeek. If you're regularly approaching 300K tokens per call, the context ceiling is a practical constraint, not a theoretical one.
The practical test: check your 90th-percentile context length in production. If it's under 300K, GPT-5's limit isn't your problem and the upgrade doesn't help you.
Computer use and agentic reasoning
GPT-5.5 is OpenAI's primary model for autonomous agent tasks. ARC-AGI-2 score: 22% versus GPT-5's 16% — a 6-point improvement on the benchmark designed to test genuine reasoning flexibility, not pattern matching. For multi-step agents where the model must take actions, recover from failures, and maintain state over long sessions, GPT-5.5's stronger reasoning ceiling translates into fewer retries and shorter task completion loops.
The caching economics
Cached input costs $0.50/1M — 90% off the standard rate. For agents with large, repeated system prompts, the a 90% cache hit rate drops GPT-5.5's effective input cost from $5.00 to approximately $0.95 per million — roughly the same as full-rate GPT-5. Output at $30/1M doesn't cache, so output-heavy workloads still pay the full GPT-5.5 premium.
For the narrative version of this rate card — including how the long-context reprice hits whole sessions — read the GPT-5.5 pricing explainer.
Batch API: the 50% route
GPT-5.5 via the OpenAI Batch API: $2.50/1M input, $15/1M output. Batch jobs run asynchronously with up to 24-hour turnaround. For evals, document classification, synthetic data generation, or any non-real-time processing pipeline, the batch rate makes GPT-5.5 cost-competitive with real-time GPT-5. This is the clearest case for GPT-5.5 in cost-sensitive environments.
Cost scenarios
At 10M input + 3M output per month, standard pricing: $50 + $90 = $140/month. GPT-5 at the same volume: $12.50 + $30 = $42.50/month — GPT-5.5 is 3.3× more expensive. With the batch discount: $25 + $45 = $70/month. With 90% cache hit on input, standard pricing: ~$9.50 input + $90 output = $99.50/month. The batch route is the most cost-efficient path to GPT-5.5 capability for teams that can tolerate async execution.
Use-case fit
Best for: Tasks that exceed GPT-5's 400K context limit; autonomous agent workloads where the 9-point SWE-bench improvement reduces failure rates; computer use pipelines; batch processing where the 50% batch discount justifies the upgrade.
Skip if: Your context needs fit within 400K tokens and reasoning depth isn't a bottleneck — GPT-5 at $1.25 handles the same tasks at one-quarter the cost. Also skip for pure volume workloads; GPT-5 or DeepSeek V4-Pro are more rational.
Decision checklist
Before routing to GPT-5.5: measure your actual 90th-percentile context length in your current production traffic — if it's under 300K, the context upgrade doesn't help you. Run an A/B on a representative sample of your hardest tasks to measure whether the SWE-bench improvement is visible in your specific failure modes.
Consider the batch API first: if your workload tolerates 24-hour turnaround, the batch pricing brings GPT-5.5 to $2.50/$15 — a different cost conversation than standard pricing.