LLM Comparison Table
The 11 current frontier tiers: input, cached and output pricing, context and max output, reasoning mode, tool calling, structured output, image input — and the cost per thousand requests for your own token mix. Every column sorts. Checked 2026-08-15, synced weekly.
| Model | Input $/M | Cached $/M | Output $/M | Context | Max output | Capabilities | Released | Per 1k requests |
|---|---|---|---|---|---|---|---|---|
| Claude Opus 5 Anthropic · 2026-05 | $5 | $0.5 | $25 | 1000k | 128k | RTSV | 2026-07-24 | — |
| Claude Sonnet 5 Anthropic · 2026-01-31 | $2 | $0.2 | $10 | 1000k | 128k | RTSV | 2026-06-29 | — |
| Claude Haiku 4.5 Anthropic · 2025-02-28 | $1 | $0.1 | $5 | 200k | 64k | RTSV | 2025-10-15 | — |
| DeepSeek V4 Pro DeepSeek | $0.435 | $0.003625 | $0.87 | 1000k | 384k | RTSV | 2026-08-12 | — |
| DeepSeek V4 Flash DeepSeek · 2025-05 | $0.14 | $0.0028 | $0.28 | 1000k | 384k | RTSVO | 2026-07-31 | — |
| Gemini 3.7 Flash Google · 2026-03 | $0.75 | $0.075 | $3.75 | 1049k | 66k | RTSV | 2026-08-13 | — |
| Gemini 3.5 Flash Lite Google · 2026-03 | $0.3 | $0.03 | $2.5 | 1049k | 66k | RTSV | 2026-07-21 | — |
| GPT-5.6 Sol OpenAI · 2026-02-16 | $5 | $0.5 | $30 | 1050k | 128k | RTSV | 2026-07-09 | — |
| GPT-5.6 Terra OpenAI · 2026-02-16 | $2 | $0.2 | $12 | 1050k | 128k | RTSV | 2026-07-09 | — |
| GPT-5.3 Codex OpenAI · 2025-08-31 | $1.75 | $0.175 | $14 | 400k | 128k | RTSV | 2026-02-05 | — |
| GPT-5.6 Luna OpenAI · 2026-02-16 | $0.2 | $0.02 | $1.2 | 1050k | 128k | RTSV | 2026-07-09 | — |
Click a header to sort; "per 1k requests" recomputes from the token mix on the left.
R Reasoning mode · T Tool calling · S Structured output · V Image input · O Open weights · 2026-05 Knowledge cutoff
List prices, synced weekly from models.dev and cross-checked against LiteLLM. Discounts, batch tiers and long-context rates are not in the table.
Why a price pair does not settle anything
Whether "$5 / $25" beats "$2 / $12" depends on the shape of your traffic. RAG workloads are input-heavy — thousands of tokens of context for a few hundred tokens of answer — so the input rate dominates. Code generation and long-form writing invert that, and output is priced 5–6× input almost everywhere. So the rightmost column is not a copied price: it blends both rates at the token mix you set, per thousand requests. Change the ratio and the ranking changes with it.
The capability marks are limited to the four that change an architecture decision: R reasoning mode, T tool calling, S structured output (schema-constrained), and V image input. A model without tool calling cannot drive an agent; one without structured output means writing your own JSON repair layer, which is a separate engineering bill.
What separates this from most comparison tables is how it is maintained: it is not typed in. A weekly script syncs from models.dev and cross-checks LiteLLM's price file; a move over 60%, or a disagreement over 25% between the two sources, makes it refuse to write and ask for a human. New models arrive by family and tier, and superseded tiers drop out. So this is this week's data, not somebody's snapshot from six months ago.
Not in the table: batch and low-priority discounts (typically half), long-context tiers (roughly double above each vendor's threshold), cache-write rates, and whatever you negotiated. To model those, use the cost calculator, where every price is editable.