Model cost guides
What each LLM really costs per call
Price breakdowns, cache economics, rate-limit structure and a circuit-breaker policy for every major model, computed from one price table so the numbers always agree.
- 1Claude Opus 4.50 to 70% of budget$10.00baseline
- 2Claude Sonnet 4.570 to 90%$6.0040% cheaper
- 3Claude Haiku 4.590 to 100%$2.0080% cheaper
- Model guides
- 28
- Providers and hosts
- 9
- With cache-read pricing
- 22 of 28
- Blended range / 1M
- $0.058 to $30.00
Browse by provider
Pick a model, see the economics
Filter by tier or by whether the provider discounts cached input. Each guide prices a realistic workload and compares it against three alternatives.
28 guides
OpenAI
8 guides
- GPT-5OpenAI's flagship reasoning model for agents and coding. Input is cheap relative to its capability, but every hidden reasoning token is billed at the output rate, so the effort setting decides your bill more than prompt length does.in $1.25out $10.00cache $0.13400K
- GPT-5 miniThe cost-efficient GPT-5 variant. It keeps the reasoning behaviour and 400K context of GPT-5 at a fifth of the price, which makes it the default workhorse for most production agent steps.in $0.25out $2.00cache $0.025400K
- GPT-4.1OpenAI's non-reasoning flagship with a 1M token context window. No hidden thinking tokens, so cost is predictable from prompt and completion length, and long documents do not carry a context surcharge.in $2.00out $8.00cache $0.501M
- GPT-4.1 miniA small, fast GPT-4.1 variant that keeps the 1M context window. A common choice for high-volume extraction and classification where reasoning tokens would be wasted spend.in $0.40out $1.60cache $0.101M
- GPT-4oOpenAI's multimodal model for text, image and audio input. Still widely deployed in production, though newer models now beat it on price for text-only work.in $2.50out $10.00cache $1.25128K
- GPT-4o miniThe low-cost GPT-4o variant, still a popular default for chat features and lightweight agents. Cheap per call, which is exactly why loops on it go unnoticed.in $0.15out $0.60cache $0.075128K
- o3OpenAI's high-effort reasoning model. The list price looks moderate, but reasoning tokens are billed as output and can outnumber visible tokens many times over, so cost per answer varies widely.in $2.00out $8.00cache $0.50200K
- o4-miniA smaller, faster OpenAI reasoning model at roughly half the o3 rate. Popular for coding and math agents where reasoning helps but o3 cost is hard to justify.in $1.10out $4.40cache $0.28200K
Anthropic
6 guides
- Claude Opus 4.5Anthropic's most capable model at a much lower rate than earlier Opus versions. Strong for long-horizon agents and coding, with prompt caching that rewards stable prefixes and penalizes churn.in $5.00out $25.00cache $0.50200K
- Claude Opus 4.1The previous Opus generation at a much higher rate than Opus 4.5. Still pinned in some production systems; worth re-pricing because the same workload often costs far less on the newer model.in $15.00out $75.00cache $1.50200K
- Claude Sonnet 4.5Anthropic's default model for agents and coding, and the model behind most of the loop incidents we analyze. The output rate is five times input, and explicit cache control decides whether long agent prompts are affordable.in $3.00out $15.00cache $0.30200K
- Claude Sonnet 4The previous Sonnet generation at the same list price as Sonnet 4.5. Common in production systems pinned to a model snapshot; caching behaviour and cost drivers are the same.in $3.00out $15.00cache $0.30200K
- Claude Haiku 4.5Anthropic's fast small model with near-frontier coding ability. A natural fallback target for Sonnet and a strong default for sub-agents.in $1.00out $5.00cache $0.10200K
- Claude 3.5 HaikuThe previous small Claude model, slightly cheaper than Haiku 4.5. Still used for high-volume classification and as a final fallback tier.in $0.80out $4.00cache $0.080200K
3 guides
- Gemini 2.5 ProGoogle's flagship thinking model with a 1M token context window. Pricing is tiered by prompt length, and thinking tokens are billed as output, so long-context agents need explicit guardrails.in $1.25out $10.00cache $0.311M
- Gemini 2.5 FlashGoogle's price-performance thinking model with a 1M context window and a configurable thinking budget. One of the cheapest ways to run long-context work.in $0.30out $2.50cache $0.0751M
- Gemini 2.5 Flash-LiteThe lowest-cost Gemini 2.5 model, built for throughput. Thinking is off by default, which keeps per-call cost close to visible tokens.in $0.10out $0.40cache $0.0251M
DeepSeek
2 guides
- DeepSeek V3.2 (chat)DeepSeek's general chat model in non-thinking mode. Very low list price and automatic context caching make it one of the cheapest capable models in the index.in $0.28out $0.42cache $0.028128K
- DeepSeek V3.2 (reasoner)DeepSeek V3.2 in thinking mode. Same list price as the chat mode, but the chain of thought is billed as output, so output tokens per request are much higher.in $0.28out $0.42cache $0.028128K
Mistral
3 guides
- Mistral Large 2Mistral's large general-purpose model, available on la Plateforme and major clouds. Output is only three times input, and there is no discounted cache-read rate in the list price.in $2.00out $6.00cache n/a128K
- Mistral Medium 3Mistral's mid-sized model priced well below frontier models while targeting strong performance on coding and multimodal tasks.in $0.40out $2.00cache n/a128K
- CodestralMistral's coding model with a 256K context window, built for code completion and fill-in-the-middle as well as chat. Low output price suits high-frequency IDE traffic.in $0.30out $0.90cache n/a256K
Amazon
2 guides
- Amazon Nova ProAmazon's multimodal mid-tier model on Bedrock. Priced well under frontier models, with Bedrock prompt caching and batch inference for further savings.in $0.80out $3.20cache $0.20300K
- Amazon Nova LiteA very low-cost multimodal model on Bedrock, with Nova Micro below it for text only. Both suit high-volume workloads inside AWS.in $0.060out $0.24cache $0.015300K
Meta (Groq)
2 guides
- Llama 3.3 70BMeta's open-weight 70B model served on Groq's LPU hardware. Very fast output and output priced close to input; the same model is available from several hosts at different rates.in $0.59out $0.79cache n/a128K
- Llama 4 MaverickMeta's Llama 4 Maverick, a mixture-of-experts model with image input, served on Groq. Cheaper than Llama 3.3 70B on Groq at list price.in $0.20out $0.60cache n/a128K
xAI
1 guide
Cohere
1 guide
Inside every guide
Five things we check for each model
- 01
Price breakdown
Input, output, cached input, cache write and batch availability, per 1M tokens.
- 02
Cost drivers
Reasoning tokens, cache write premiums, long-context bands: whatever moves this model's bill.
- 03
Typical call
A realistic workload priced line by line, with and without cache hits.
- 04
Rate limits
How the provider structures limits by tier, and where to read your account's numbers.
- 05
Breaker policy
A PALM SOUQ policy file with budgets, a fallback and all four loop heuristics.
About the model guides
Where do the prices in these guides come from?
Every number is read from the PALM SOUQ LLM Price Index, which lists provider list prices for the standard tier and shortest context band. The last full review was in June 2026. Each guide computes its examples from those rows, so one correction updates every page.
Why do the guides not list exact rate limits?
Provider rate limits depend on your account's usage tier and change often. Each guide describes how the provider structures limits and names the documentation domain where you can read your own account's numbers. PALM SOUQ enforces your per-tenant limits in front of whatever the provider allows.
What does blended 3:1 mean?
A blended price weights three input tokens for every output token, a common shape for chat and agent traffic: (3 x input price + output price) / 4. It is a quick way to rank models; real cost depends on your own ratio, caching and reasoning tokens.
Can I use the sample policies as they are?
They are working starting points. Budgets and step caps should come from your own traffic: PALM SOUQ records steps per session and spend per tenant, and most teams set caps a margin above the 99th percentile they observe in the first week.
Don't wait for a surprise $10,000 OpenAI invoice.
Route your agent traffic through PALM SOUQ in 2 minutes. Change one base URL, pass a tenant header, and every token is attributed and governed.
14-day free production trial · 100k requests included · no card required