GPT-5 and GPT-5-mini API Current Limit Increase: Multi-fold TPM for large-scale inference and batch processing
This increase covers multiple levels of usage tiers: Tier 1 of GPT-5 has been raised from 30K to 500K TPM (batch processing limit is 1.5M), Tier 2 has been raised to 1M (batch processing is 3M), Tier 3 has been raised to 2M, and Tier 4 has been raised to 4M. GPT-5-mini's Tier 1 is raised to 500K (batch processing 5M). For AI workloads that require high concurrency and long context, this is an immediate throughput boost.
1. List of changes
1. GPT-5 (Standard Model)
Tier 1: 30K → 500K TPM (batch 1.5M)
Tier 2: 450K → 1M (batch 3M)
Tier 3: 800K → 2M
Tier 4: 2M → 4M
2, GPT-5-mini (lightweight model)
Tier 1: 200K → 500K TPM (batch 5M)
2. What does this mean for engineering
1. Concurrency and long contexts are more stable
High TPM directly alleviates throughput bottlenecks in contexts above 32K, and batch evaluation, long-form article generation, and multi-tool agents can reduce queues and throttling fallbacks.
2. Improved batch processing cost performance
Ahigher batch queue allows small requests to be merged, reducing the handshake and network overhead of each call, and is suitable for log summary and multi-prompt parallelism.
3. Cost and current throttling governance are more controllable
, and moreeffective tokens can be carried under the same budget. With rate limiting and de-escalation policies, peaks can be flattened to batch channels.
3. Quick landing list
1. Routing and quotas
(1) Route long context and evaluation tasks to GPT-5; Use GPT-5-mini for light interaction and monitoring.
(2) Set TPM thresholds for each project and environment to avoid "overcrowding" for a single tenant.
(3) Enable exponential backoff of failed retries to prevent instantaneous congestion.
2. Batch processing and caching
(1) Merge similar requests and control the batch size in the optimal range of the model.
(2) Enable hint and retrieval result caching to reduce duplicate token consumption.
(3) Convection output retention timeout and breakpoint continuation.
3. Measurement and regression
(1) Track the acceptance rate, revocation rate, and unit token cost.
(2) Perform stress test baselines for 8K, 32K, and 128K contexts.
(3) Reserve the old quota fallback path to prevent policy switching jitter.
Frequently Asked Questions (Q&A)
Q: How can I confirm my organization's current GPT-5 and GPT-5-mini limits and tiers?
A: View the usage tier and model quota of your organization on the Quotas page of the platform, and check the actual TPM and batch quota with the billing and usage reports.
Q: How do TPM counting rules relate to max_tokens?
A: TPM is calculated based on the input token and the set maximum output, whichever is larger, and it is recommended to keep the maximum output close to the real demand to avoid "inflated" occupancy.
Q: Can batch processing replace concurrent requests across the board?
A: Suitable for similar tasks that can tolerate delays; Interactive conversations and tool calls are still dominated by low-latency single requests, supplemented by batch processing.
Q: Is this limit increase effective for the long term?
A: The official announcement is a "limit increase", and the specific long-term strategy is subject to the platform documentation and follow-up announcements, and it is recommended to retain the limit revert and multi-model bottom.
