TokenAtlasThe ranked list
1. Gemini 1.5 Flash — $0.075 in / $0.30 out. 2. Claude Haiku — $0.25 in / $1.25 out. 3. GPT-4o-mini — $0.15 in / $0.60 out. 4. Mistral Small — $0.20 in / $0.60 out. 5. GPT-3.5-turbo — $0.50 in / $1.50 out.
Cost ≠ value
Flash beats Haiku on raw price but Haiku has stronger structured output. GPT-4o-mini wins on tool calling. Always benchmark against your task.
When to pick the cheapest
Classification, routing, simple summarization, embeddings prep, and high-volume bulk transforms. Reserve frontier models for reasoning-heavy steps.
A bulk workload, priced across the tiers
At 50M input tokens and 50M output tokens a month (a large classification/summarization job): GPT-4o mini = (50M x $0.15) + (50M x $0.6) = $7.50 + $30 = $37.50. DeepSeek V3 = (50M x $0.27) + (50M x $1.10) = $13.50 + $55 = $68.50. Llama 3.1 70B = (50M x $0.59) + (50M x $0.79) = $29.50 + $39.50 = $69. Claude 3.5 Haiku = (50M x $0.8) + (50M x $4) = $40 + $200 = $240. For reference, GPT-4o at this volume = $125 + $500 = $625 — GPT-4o mini is roughly 16.7x cheaper than GPT-4o here.
Output price is where the cheap tier splits apart
Input rates across these models are all within a narrow band ($0.15-$0.80/1M), but output rates span $0.60 to $4/1M — a 6.7x range. On an output-heavy job, Claude 3.5 Haiku's advantage over frontier models shrinks fastest, while GPT-4o mini's low output rate keeps its cost advantage largest. Check your input:output ratio before assuming any 'cheap' model behaves the same way on your workload.
How much slack there is before the discount disappears
At $37.50 vs GPT-4o's $625 for the same token volume, GPT-4o mini would need roughly 16.7x more calls (retries, re-runs, longer follow-ups) before its total cost caught up to GPT-4o's — a large margin, but not infinite. For models closer to the middle of the range, like Claude 3.5 Haiku at $240 versus GPT-4o's $625 (2.6x cheaper), far fewer retries would erase the savings. Model this against your own expected retry rate rather than assuming the sticker-price gap survives production.
Running this on your own numbers
Rank models by cost using your actual monthly input and output volumes in TokenAtlas, not the generic per-1M rate alone — a model with a lower headline price can still cost more on your specific ratio if its output rate is high relative to the rest of the field, as Claude 3.5 Haiku shows above.
Frequently asked questions
- Are these prices current?
- Updated within 24h of provider price changes.
- Will the cheapest model do my job?
- Run a 100-sample benchmark first — TokenAtlas Pro includes a side-by-side eval runner.
- Is Flash really frontier-grade?
- On benchmarks, Flash beats GPT-3.5 and approaches Sonnet for many tasks.