GPT-4 era

If you are already on a GPT-4-era tier

This page is for two readers: someone with a working production deployment pinned to a GPT-4-era tier who needs to know when migration is worth deferring, and someone running gpt-oss as the self-host option (the only category without a GPT-5 successor).

If you are mid-migration, the tier replacements are:

• GPT-5 Mini at the 5.4-thinking tier (Quality Score 84.2, $0.75 input / $4.5 output per million) is cheaper than most GPT-4-era tiers and quality-comparable to or above them at the chat workhorse workload.
• GPT-5 Mini at the 5.0 effort tier (QS 77.3, $0.25 input / $2 output) is the cheapest competent OpenAI-side option that still carries current-generation behaviour.

When staying on a GPT-4-era tier is defensible

• Pinned evals or fine-tunes. If your production eval was qualified on GPT-4o or GPT-4.1 and the result is critical, the migration cost includes re-running the eval on GPT-5 mini before switching. Plan that work, do not skip it.
• You are using o3 specifically. The o-series reasoning approach was the experiment GPT-5 unified. o3 at QS 81.6 ($2 / $8 per million) is a uniquely cheap reasoning option in our index; if your workload was tuned for o-series output behaviour, the migration to GPT-5 thinking modes is not a drop-in.
• You need open-weights deployment. gpt-oss (120b and 20b) is OpenAI's first open-weights line. Hosted-pricing rows in our index list both at unusually low rates ($0.037 input / $0.17 output per million for 120b on hosted routes), and the 20b variant is the realistic single-GPU self-host candidate in the broader OpenAI catalog. There is no GPT-5 open-weights option, so this is a category, not a tier comparison.
• Cheapest tier that still works. GPT-4.1 nano at $0.1 / $0.4 is the cheapest priced OpenAI tier in our index. For repetitive low-stakes turns where the score gap to GPT-5 nano ($0.05 / $0.4) does not move the unit economics, staying pinned to a working 4.1-nano deployment is defensible.

Where the data is weak

This page covers a wide catalog and the coverage is uneven across it.

• The o-series and chat models are not directly comparable. o1, o3, and o4 mini have benchmark coverage on the reasoning-flavoured evals (GPQA Diamond, AIME) and lighter coverage on the chat-flavoured ones. The chat 4.1/4.5/4o lines are the inverse. Cross-reading a single Quality Score across both is a category error; the per-variant rows on this page show the split.
• GPT-4.5 has the thinnest data of any tier here. Some fields (context window, list pricing) are unset in our index. If your decision depends on 4.5 specifically, cross-check against OpenAI's own docs.
• gpt-oss benchmark scores are listed across both thinking and non-thinking modes. The score gap is large (72.3 vs 59.8 for 120b); read the Mode column in the variant table before quoting any oss number.
• Pricing on this page is the published list price. OpenAI volume and Azure routing change unit economics; list price is a calibration anchor only.

When to look outside this era

• GPT-5 family (/en/ai/llm/gpt-5) is the natural successor for every tier on this page except gpt-oss. If the migration question is still open, that surface is the comparison to read.
• Open-weights at the same workload tier: Qwen3 and DeepSeek V4 both ship open-weights variants at chat workhorse quality with hosted-API pricing competitive with gpt-oss. If gpt-oss is the reason to stay in this era, those two families are the cross-family comparison worth doing.

Sources worth reading

• OpenAI API pricing: vendor price list (current generation; GPT-4-era tiers listed alongside live ones)
• OpenAI deprecations: which previous-generation models still ship and which have a sunset date
• gpt-oss on Hugging Face: model cards and weights for the open-weights line