| There is no best model | There are only best fits for a workload, a budget, and a latency target |
| Route, do not upgrade | A tier spread of 5x to 25x means routing beats picking a single "best" model |
| Cache the stable part | Cached input runs around 90% cheaper. Worth the effort every time |
| Batch anything offline | Around 50% off, where offered. Free money for non-urgent jobs |
| Verify before you budget | Vendor pages move faster than aggregators. Trust the source, not the summary |
| This table shows | Prices, context windows, and typical fit. Not a quality ranking |
| This table does not show | Benchmarks. Leaderboards move faster than prices, and rankings depend on the eval |
| Never budget from this page | Prices change weekly. Confirm the official pricing page on contract day |
| Model | Tier | Input | Output | Cached In | Context | Best for |
|---|---|---|---|---|---|---|
| Anthropic (Claude) | ||||||
| Claude Opus 4.8 | Frontier | 5.00 | 25.00 | 0.50 | 1M | Complex reasoning, agentic coding, long-horizon autonomy |
| Claude Sonnet 5 * | Balanced | 2.00 | 10.00 | 0.20 | 1M | Default production workhorse: coding, RAG, analysis |
| Claude Haiku 4.5 | Budget | 1.00 | 5.00 | 0.10 | 200K | Classification, routing, extraction, moderation |
| Claude Fable 5 | Premium | 10.00 | 50.00 | 1.00 | 1M | Where fewer failed loops justify the token rate |
| OpenAI (GPT) | ||||||
| GPT-5.5 | Frontier | 5.00 | 30.00 | 0.50 | 1M | Hardest coding and agent tasks |
| GPT-5.4 | Balanced | 2.50 | 15.00 | 0.25 | 1.05M | Lower-cost frontier work. Strong cost-quality balance |
| GPT-5.4 mini | Budget | 0.75 | 4.50 | 0.075 | 400K | Routed production traffic below the frontier |
| GPT-5.4 nano | Ultra | 0.20 | 1.25 | 0.02 | 400K | High-volume routing, extraction, classification |
| Google (Gemini) | ||||||
| Gemini 3.1 Pro | Frontier | 2.00 | 12.00 | 0.20 | 1M | Top-tier reasoning, multimodal, long-context |
| Gemini 3.5 Flash | Balanced | 1.50 | 9.00 | 0.15 | 1M | Speed-meets-frontier with native grounding |
| Gemini 3 Flash | Budget | 0.50 | 3.00 | 0.05 | 1M | Cost-effective general work |
| Gemini 2.5 Flash-Lite | Ultra | 0.10 | 0.40 | 0.01 | 1M | Cheapest Google tier: classification, routing at scale |
| DeepSeek | ||||||
| DeepSeek V3 | Ultra | 0.27 | 1.10 | 0.027 | 128K | Lowest-cost flagship-class option for text |
| Lever | Typical saving | How |
|---|---|---|
| Trim the context | Free | Delete stale history, tighten system prompt, drop irrelevant docs. Attention dilutes anyway |
| Prompt caching | Up to ~90% off repeats | Put stable content (system, examples, docs) at the front, byte-identical every call |
| Batch processing | ~50% off | Overnight jobs, evals, backfills. Latency for money |
| Tier routing | 5x to 25x cheaper | Classify the query, route easy traffic to a budget model, keep the frontier for the hard ones |
| Cap output tokens | Variable | Output usually costs 4x to 6x input. Bound it in the request |
| Move up a tier | Last resort | Only after the above. A bigger model is a lever, not an answer |
| Workload | Reach for |
|---|---|
| Classification, routing, extraction | Ultra tier (Haiku 4.5, GPT-5.4 nano, Flash-Lite) |
| RAG synthesis, general chat | Balanced tier (Sonnet 5, GPT-5.4, Gemini 3.5 Flash) |
| Agentic coding, long tasks | Frontier (Opus 4.8, GPT-5.5, Gemini 3.1 Pro) |
| Multimodal, images, video | Gemini 3.1 Pro or Opus 4.8 |
| Long-context Q&A (500K+) | Any 1M-window model; watch Gemini Pro's price step at 200K |
| High-volume, low latency | Budget or Ultra tier; DeepSeek V3 if privacy allows |
| Regulated or safety-critical | Frontier + evals + guardrails. Never the ultra tier |
| Sonnet 5 price step | Reverts to $3 / $15 on 1 Sep 2026. Hard-code a reminder |
| Claude Fable 5 | Access history: suspended 12 Jun 2026, restored 1 Jul 2026. Confirm before adoption |
| OpenAI GPT-5.6 family | Limited preview in July 2026. May go GA and displace GPT-5.5 |
| Gemini long-context | Pro tiers step up ~2x above 200K tokens |
| DeepSeek | Figures vary widely across aggregators. Trust the official page only |
| Do |
|---|
| Track token spend per feature, not per model |
| Cache the front of your prompt aggressively |
| Route by query complexity, not by mood |
| Evaluate models on your workload, not on public benchmarks |
| Set hard cost caps and rate limits at deploy |
| Re-check pricing before every quarterly budget |
| Do not |
|---|
| Pick a single model for the whole product |
| Trust aggregator pricing pages. They lag |
| Skip caching because "the code is more complex" |
| Assume last month's benchmark still holds |
| Confuse a bigger context with better performance |
| Ship without a per-feature spend alert |
| QR-09 Cost Calculator | Enter your workload, see monthly cost across every model above. Reads the same pricing file as this page |
| QR-15 Routing Card | Which tier for which job. Small/fast vs reasoning-heavy, when to enable extended thinking |
| Business Case & ROI (BCR-05, BCR-06) | Both articles link back here rather than embedding prices. Never trust an inline price older than a month |