Models & pricing
LLM pricing, side by side
Cost per million tokens and context window for the models engineering teams actually ship on. Sorted by input price. Pick one for the full breakdown, or run your numbers through thecost calculator.
| Model | Tier | Input /1M | Output /1M | Context |
|---|---|---|---|---|
| GPT-4o miniOpenAI | Fast / cheap | $0.15 | $0.60 | 128K |
| Llama 3.3 70BMeta (via hosts) | Open-weight | $0.20 | $0.60 | 128K |
| DeepSeek-V3DeepSeek | Open-weight | $0.27 | $1.10 | 128K |
| Gemini 2.5 FlashGoogle | Fast / cheap | $0.30 | $2.50 | 1M |
| GPT-4.1 miniOpenAI | Fast / cheap | $0.40 | $1.60 | 1M |
| DeepSeek-R1DeepSeek | Open-weight | $0.55 | $2.19 | 128K |
| Claude Haiku 4.5Anthropic | Fast / cheap | $1.00 | $5.00 | 200K |
| OpenAI o4-miniOpenAI | Balanced | $1.10 | $4.40 | 200K |
| Gemini 2.5 ProGoogle | Frontier | $1.25 | $10.00 | 1M |
| Claude Sonnet 5.5Anthropic | Balanced | $2.00 | $10.00 | 1M |
| GPT-4.1OpenAI | Balanced | $2.00 | $8.00 | 1M |
| OpenAI o3OpenAI | Frontier | $2.00 | $8.00 | 200K |
| Mistral LargeMistral | Balanced | $2.00 | $6.00 | 128K |
| GPT-4oOpenAI | Balanced | $2.50 | $10.00 | 128K |
| Claude Sonnet 4.6Anthropic | Balanced | $3.00 | $15.00 | 1M |
| Claude Opus 5.5Anthropic | Frontier | $4.00 | $20.00 | 1M |
| Claude Opus 4.8Anthropic | Frontier | $5.00 | $25.00 | 1M |
| Claude Fable 5.1Anthropic | Frontier | $10.00 | $50.00 | 1M |
| Claude Fable 5Anthropic | Frontier | $10.00 | $50.00 | 1M |
Anthropic figures verified September 25, 2026; others June 2026. Prices change frequently. Always confirm against the provider’s official pricing page before relying on these figures for budgeting.
Choosing between close siblings
Most of the hard calls on this table are between two models from the same family, where the prices are close and only one or two things actually differ. Each of those decisions is written once, here, rather than argued twice from opposite sides on two model pages.
GPT-4.1 or GPT-4o
GPT-4.1 buys a 1M-token window at $2/$8. GPT-4o is the multimodal row at $2.50/$10 in a 128K window. The price gap is small enough to set aside, which leaves the two things that genuinely differ: how long your prompts get, and whether modality is a requirement or a nice-to-have.
GPT-4.1 mini or GPT-4o mini
The same decision a tier down, with the gap widened: $0.40/$1.60 for a 1M window against $0.15/$0.60 for 128K, which is roughly 2.7x the blended cost. If your prompts fit inside 128K, the premium buys capacity you never use. If they do not, nothing else on the OpenAI stack carries that window this cheaply.
Claude Fable 5 or Claude Opus 4.8
Same 1M window, same 128K output ceiling, exactly double the price on both meters. There is no context or throughput argument between them, so the only real question is whether a task genuinely exceeds Opus 4.8. Default to Opus and escalate by exception rather than starting at the top and hoping to trim later.
Claude Sonnet 4.6 or Claude Haiku 4.5
Three times the price ($3/$15 against $1/$5) and five times the context (1M against 200K). Haiku is the volume tier, Sonnet is where the reasoning lives, and the boundary between them is worth setting deliberately with evals — on an Anthropic stack that boundary is where most of the routing saving actually sits.
OpenAI o3 or o4-mini
Same 200K window, same rule that reasoning tokens bill at the output rate, and o4-mini at a little over half the price ($1.10/$4.40 against $2/$8). Because both spend output tokens thinking, the effective gap on a reasoning-heavy route is wider than the headline one. Start on o4-mini and escalate only the routes it fails.
Gemini 2.5 Pro or 2.5 Flash
Both carry a 1M-token window. Flash is roughly a quarter of Pro on both meters ($0.30/$2.50 against $1.25/$10), and Pro’s pricing tiers by prompt size in a way Flash’s headline figures do not signal. Flash is the obvious tier for long-document reading at volume; Pro is the frontier tier of the two.
DeepSeek-V3 or DeepSeek-R1
R1 is the reasoning model at roughly twice the V3 rate ($0.55/$2.19 against $0.27/$1.10). Both are open-weight, both carry a 128K window, and both route through the same offshore hosted API, so the data-residency question is identical either way. Pick on task shape: reasoning-heavy work to R1, everything else to V3.