OpenAI · GPT-5 · Small
GPT-5 mini pricing and limits
The cost-efficient GPT-5 variant. It keeps the reasoning behaviour and 400K context of GPT-5 at a fifth of the price, which makes it the default workhorse for most production agent steps.
| Line | Tokens | Rate / 1M | Cost |
|---|---|---|---|
| Input, uncached | 3,000 | $0.25 | $0.00075 |
| Input, cache read | 6,000 | $0.025 | $0.00015 |
| Output | 900 | $2.00 | $0.00180 |
- Input / 1M
- $0.25
- Output / 1M
- $2.00
- Cached / 1M
- $0.025
- Cache write / 1M
- n/a
- Blended 3:1
- $0.69
- Context
- 400K
What drives GPT-5 mini cost
The same reasoning token mechanics apply as on GPT-5: output tokens include hidden reasoning, so a low effort setting matters even on the small model. Because the per-token price is low, teams often stop watching it, and a chatty agent at high effort can quietly cost more than a well-tuned GPT-5 call.
Automatic prompt caching bills repeated prefixes at a tenth of the input price. In high-volume pipelines where the instructions dwarf the user payload, cached input can be the majority of tokens, so tracking cache hit rate per tenant tells you whether a deploy broke your prefix.
Cost of a typical call
Take a support agent turn with 9,000 input tokens, 6,000 of them served from cache, and 900 output tokens. At GPT-5 mini list prices ($0.25 input, $0.025 cached input, $2.00 output per 1M) one call costs $0.00270, against $0.00405 with no cache hits. At 600,000 calls a month that is $1,620, or $2,430 uncached.
Blended at a 3:1 input to output ratio, GPT-5 mini costs $0.69 per 1M tokens. Routing the same call to the fallback in the policy below, GPT-5 nano, costs $0.00054, which is what a tenant pays you in COGS once they cross the hard cap and PALM SOUQ downgrades instead of blocking.
Cost comparison
GPT-5 mini against three alternatives
The same support agent turn priced on GPT-4.1 mini, Claude Haiku 4.5, Gemini 2.5 Flash. Move the sliders to match your traffic; models without a cache-read rate bill every input token in full.
Benchmark quality per task before switching. A cheaper model that needs more steps or longer outputs can cost more per finished task than its rate suggests, which is why PALM SOUQ reports cost per feature, not only per token.
GPT-5 mini this guide
$0.00269 per call
$1,347 / mo
baseline
GPT-4.1 mini
$0.00323 per call
$1,616 / mo
+20% vs GPT-5 mini
Claude Haiku 4.5
$0.00807 per call
$4,037 / mo
+200% vs GPT-5 mini
Gemini 2.5 Flash
$0.00359 per call
$1,797 / mo
+33% vs GPT-5 mini
List prices from the PALM SOUQ LLM Price Index, computed per call as uncached input + cache reads + output. Reasoning tokens count as output on reasoning models.
Rate limits by usage tier
Limits are per organization and per model, raised automatically by usage tier. Smaller models usually get higher token-per-minute ceilings than the flagship at the same tier, but bulk jobs can still saturate them; spread heavy tenants with per-tenant rolling windows so one customer cannot starve the rest.
Exact limits vary by account tier and change over time, so we do not reprint them here. Check the limits page for your own account in the OpenAI documentation at platform.openai.com/docs. PALM SOUQ enforces your own per-tenant limits in front of the provider's, so a single customer cannot consume the shared allowance that every other tenant depends on.
Circuit-breaker best practices
Use token velocity as the primary breaker. Cheap models are where runaway loops go unnoticed longest, because each call looks harmless on a dashboard; a 50k tokens per second ceiling over a 10 second window flags the burst immediately.
Set a per-tenant daily dollar cap with a soft alert at 80%. On a model this cheap, a hard cap rarely blocks real users but reliably stops background jobs that never terminate.
When a breaker trips, the gateway answers in under 50 ms with HTTP 429 and X-PalmSouq-Reason: circuit_breaker_triggered, before the request reaches OpenAI. Your agent framework sees a normal rate-limit response and stops, and the audit log records which heuristic fired and for which tenant.
# palmsouq.policy.yaml
policy: gpt-5-mini-pro
match:
model: gpt-5-mini
tier: pro # from X-PalmSouq-Tier
budgets:
monthly_usd: 150 # per tenant
soft_cap_pct: 80 # webhook to your app
on_hard_cap: fallback # or: block (HTTP 429)
fallback:
- gpt-5-nano
breakers:
tool_thrashing: { identical_calls: 3, window_steps: 8 }
token_velocity: { max_tokens_per_s: 50000, window_s: 10 }
semantic_echo: { repeats: 3 }
max_steps: 20
max_reasoning_tokens: 6000 # per request
cache:
meter_reads: true- budgetsPer-tenant monthly dollar cap with an 80% soft-cap webhook, keyed by X-PalmSouq-Tenant.
- fallbackAt the hard cap, requests reroute to GPT-5 nano ($0.14 blended per 1M) instead of failing.
- breakersAll four loop heuristics on, evaluated in the request path with under 5 ms added latency.
Where it fits
Default choice for customer-facing agents, RAG answers and structured extraction at scale. Escalate to GPT-5 only on steps where an evaluator shows a measurable quality gap.
Provider documentation: platform.openai.com/docs
GPT-5 mini pricing questions
How much does GPT-5 mini cost per million tokens?
At list price GPT-5 mini costs $0.25 per 1M input tokens and $2.00 per 1M output tokens, with cached input at $0.025 per 1M. Prices were last reviewed in the PALM SOUQ LLM Price Index and are for the standard tier and shortest context band.
Does GPT-5 mini support prompt caching?
Yes. Cached input is billed at $0.025 per 1M tokens against $0.25 for uncached input. PALM SOUQ records cached tokens per tenant from the provider's usage data, so cache savings show up in each customer's COGS.
What are the GPT-5 mini rate limits?
Limits are per organization and per model, raised automatically by usage tier. Smaller models usually get higher token-per-minute ceilings than the flagship at the same tier, but bulk jobs can still saturate them; spread heavy tenants with per-tenant rolling windows so one customer cannot starve the rest. For your account's exact numbers, see the OpenAI documentation at platform.openai.com/docs.
How do I stop runaway GPT-5 mini agent loops?
Use token velocity as the primary breaker. Cheap models are where runaway loops go unnoticed longest, because each call looks harmless on a dashboard; a 50k tokens per second ceiling over a 10 second window flags the burst immediately. The sample policy on this page enables all four PALM SOUQ heuristics with a max step cap of 20 and falls back to GPT-5 nano at the hard cap.
Is there a batch discount for GPT-5 mini?
50% off input and output through the Batch API (24-hour completion window). Batch traffic suits evaluations, backfills and nightly enrichment jobs that do not need a real-time answer.
Govern GPT-5 mini spend per customer.
Point your OpenAI client at gateway.palmsouqltd.com, pass X-PalmSouq-Tenant, and every GPT-5 mini token is attributed, capped and protected by circuit breakers.
14-day free production trial · 100k requests included · no card required