OpenAI reduced standard o3 prices by 80% to $2 per million input tokens and $8 per million output tokens. The o3-pro model costs $20 per million input tokens and $80 per million output tokens. This premium model targets complex domains like math, science, and business analysis. The o3-mini model is a cost-effective version for lighter tasks. Users on the Plus tier get limited access to o3 for $20 per month. The Pro tier provides unlimited usage for $200 per month. OpenAI also introduced three Pro plans: Pro 100 at $100, Pro 200 at $200, and Pro 500 at $500. The Pro 500 plan provides the highest included usage of the three plans. The Pro 500 plan also remains the only Pro plan with Ultrafast access included. In Ultrafast mode, these models use the included plan allowance at eight times the Standard rate and draw from credit balances at six times the Standard rate. For users on Pro 100 or Pro 200, buying credits does not unlock Ultrafast. The cost for the o4-mini model is $1.10 per million input tokens and $4.40 per million output tokens. Claude 4.0 offers flexible reasoning at $3 per million input tokens and $15 per million output tokens. Unlike the dedicated o3-pro, Claude 4.0 adapts to task complexity dynamically. You should evaluate these tiers based on your specific reasoning needs.
| Model | Input (per 1M) | Output (per 1M) |
|---|---|---|
| o3 | $2.00 | $8.00 |
| o3-pro | $20.00 | $80.00 |
| o4-mini | $1.10 | $4.40 |
Reasoning benchmark gains
The o3 model demonstrates significant reasoning gains compared to its predecessor. On the 2024 American Invitational Mathematics Exam, o3 scored 96.7%, missing just one question. It achieved 87.7% accuracy on GPQA Diamond, which consists of PhD-level science questions. For the ARC AGI benchmark, o3 reached 88% on high compute settings, surpassing the 85% threshold for human-level performance. This test requires the model to predict the next geometric pattern in a grid without prior examples. The test pushes the model to think beyond pre-learned answers because each test is unique. In one instance, a yellow square with red dots gets a red border, and a yellow square with blue dots gets a blue border. The model must then apply this logic to a test input with a yellow square and magenta dots. In competitive coding, o3 reached a Codeforces rating of 2727, placing it in the top 0.8% of human coders. This performance marks a 46% improvement over o1 in software engineering tasks. The model also achieved 71.7% accuracy on Bench Verified, which simulates real-world software engineering tasks. OpenAI uses deliberative alignment to teach models the rules of safety and ethics in plain language. This method helps the model reason through tough situations instead of just reacting. While standard models respond in 15 to 30 seconds, the o3-pro model can take anywhere from 2 to 10 minutes to process complex problems to ensure it reaches the correct conclusion on the first try.
Usage limits and subscription changes
Changes to the Pro 200 plan affect many existing subscribers. Those with active subscriptions before September 29, 2026, keep their previous usage allowance through October 29, 2026. After that, they move to a lower allowance at the same $200 monthly price. The Pro 200 usage in Codex and ChatGPT Work dropped from 20 times to 10 times the Plus allowance. The Pro 200 plan is available for new subscriptions, though these users receive a lower allowance than those who were grandfathered in. The monthly price for Pro 200 remains at $200 per month. The o3-pro model justifies its high cost for specialized technical work. Because o-series models bill internal reasoning tokens at output rates, effective costs run 3x to 10x the base rate depending on task complexity. Reasoning tokens bill as output whether or not they appear in the response. This means the cost of a response depends on the amount of internal deliberation. Some users find that the two-minute waits for o3-pro frustrate their workflow. The Pro 100 and Pro 200 tiers provide different usage levels, while Pro 500 offers the highest. Will the higher costs of reasoning tokens eventually force a shift back to standard models?
