#01DeepSeek
DeepSeek V3.2
Open-weight model from DeepSeek. Among the cheapest capable LLMs available.
Monthly cost
$3.500
Lowest in this comparison
- Per call
- $0.000350
- Per 1,000 calls
- $0.3500
- Annual estimate
- $42.00
ToolsCachingCode
Vol. 01 — Issue 01Updated Apr 2026
For builders, buyers, and operatorsCalculators, comparisons,
and clarity for the
AI economy.
Independent pricing intelligence for GPT-5.4, Claude Opus 4.7, Gemini 3 Pro, DeepSeek and 100+ other models. Count tokens, estimate API costs, and pick the right model for your workload — without the marketing noise.
Pricing tracked from
OpenAIAnthropicGoogleDeepSeekxAIMistralMeta
Fig. 1 — Cross-model pricing trajectories, Q1 2026
01 — Token Calculator
Paste in any prompt. Get the exact cost across every major model.
Token estimates use a content-aware heuristic (~4 chars / token for English, ~3.2 for code, 1.5 tokens / CJK character). For exact counts, run your text through the official tokenizer of your chosen provider.
Your prompt
Tokens
79
Words
45
Chars
315
Cost assumptions
2.0× input
≈ 158 output tokens estimated
Assumes 70% of input is cacheable system prompt
GPT-5.4
OpenAI · OpenAI's flagship multimodal model with deep reasoning and 1M context.
$2.5 / 1M in$15 / 1M out$0.25 / 1M cached
$0.002568
per call
$2.568/mo
GPT-5 mini
OpenAI · Cost-efficient GPT-5 variant for high-volume production workloads.
$0.25 / 1M in$2 / 1M out$0.025 / 1M cached
$0.000336
per call
$0.3358/mo
Claude Opus 4.7
Anthropic · Anthropic's most capable model. Best-in-class for complex reasoning and coding.
$5 / 1M in$25 / 1M out$0.5 / 1M cached
$0.004345
per call
$4.345/mo
Claude Sonnet 4.6
Anthropic · Balanced flagship model. Strong general-purpose performance at moderate cost.
$3 / 1M in$15 / 1M out$0.3 / 1M cached
$0.002607
per call
$2.607/mo
Gemini 3 Pro
Google · Google's flagship with native multimodality and 1M context.
$2 / 1M in$12 / 1M out$0.2 / 1M cached
$0.002054
per call
$2.054/mo
Best value
DeepSeek V3.2
DeepSeek · Open-weight model from DeepSeek. Among the cheapest capable LLMs available.
$0.14 / 1M in$0.28 / 1M out$0.028 / 1M cached
$0.000055
per call
$0.0553/mo
Estimates based on listed per-token pricing. Actual costs may vary based on prompt caching usage, batch API discounts, and enterprise pricing agreements. Reasoning models may bill thinking tokens separately.
AdvertisementSponsored
02 — Model Comparison
Every major model, side by side. Sortable, filterable.
Pricing, context windows, and capabilities for 20 models from 9 providers — verified against official provider pricing pages as of April 2026.
In Showdown
GPT-5.4Claude Sonnet 4.6Gemini 3 ProDeepSeek V3.2
4/4Provider
Tier
Showing 20 of 20 models
| Model | Provider | Context | Input / 1M↑ | Cached / 1M | Output / 1M | Capabilities | Showdown |
|---|---|---|---|---|---|---|---|
GPT-5 nano | OpenAI | 400K | $0.050 | $0.005 | $0.400 | ToolsCaching | |
Qwen3 235Bopen weights | Alibaba | 262K | $0.071 | — | $0.100 | ToolsCodeReasoning | |
Mistral Small 3.2open weights | Mistral | 131K | $0.075 | — | $0.200 | ToolsCode | |
Llama 4 Scoutopen weights | Meta | 328K | $0.080 | — | $0.300 | VisionToolsCode | |
Gemini 2.5 Flash-Lite | Google | 1M | $0.100 | $0.010 | $0.400 | VisionToolsCaching | |
DeepSeek V3.2open weights | DeepSeek | 164K | $0.140 | $0.028 | $0.280 | ToolsCachingCode | |
GPT-5 mini | OpenAI | 400K | $0.250 | $0.025 | $2.00 | VisionToolsCachingCode | |
Llama 4 Maverickopen weights | Meta | 128K | $0.270 | — | $0.850 | VisionToolsCode | |
Grok 4 mini | xAI | 128K | $0.300 | $0.075 | $1.50 | ToolsCaching | |
Gemini 3 Flash | Google | 1M | $0.500 | $0.050 | $3.00 | VisionToolsCachingAudio | |
DeepSeek R1.5open weights | DeepSeek | 128K | $0.550 | $0.140 | $2.19 | ReasoningCode | |
Claude Haiku 4.5 | Anthropic | 200K | $1.00 | $0.100 | $5.00 | VisionToolsCaching | |
Gemini 2.5 Pro | Google | 2M | $1.25 | $0.125 | $10.00 | VisionReasoningToolsCachingAudioCode | |
Gemini 3 Pro | Google | 1M | $2.00 | $0.200 | $12.00 | VisionReasoningToolsCachingAudioCode | |
Mistral Large 2.1 | Mistral | 128K | $2.00 | — | $6.00 | VisionToolsCode | |
GPT-5.4 | OpenAI | 1M | $2.50 | $0.250 | $15.00 | VisionReasoningToolsCachingCode | |
Command R+ 08-2025 | Cohere | 128K | $2.50 | — | $10.00 | ToolsCode | |
Claude Sonnet 4.6 | Anthropic | 200K | $3.00 | $0.300 | $15.00 | VisionReasoningToolsCachingCode | |
Grok 4 | xAI | 256K | $3.00 | $0.750 | $15.00 | VisionReasoningToolsCachingCode | |
Claude Opus 4.7 | Anthropic | 200K | $5.00 | $0.500 | $25.00 | VisionReasoningToolsCachingCode |
Prices in USD per 1 million tokens. Click column headers to sort. Open-source models typically have multiple hosting providers — listed price reflects a representative provider (Together.ai, Fireworks, etc.). Verify with the provider before purchase.
03 — Cost Showdown
Pit any models head-to-head against your workload.
Choose two to four models, dial in your real input and output token sizes, and see exactly what each one will cost per month and per year. Share the result with a single link.
Models in this showdown
4 / 4GPT-5.4OpenAIClaude Sonnet 4.6AnthropicGemini 3 ProGoogleDeepSeek V3.2DeepSeek
1,500/ call
500/ call
10,000
#02Google
Gemini 3 Pro
Google's flagship with native multimodality and 1M context.
Monthly cost
$90.00
25.7× the cheapest
- Per call
- $0.009000
- Per 1,000 calls
- $9.000
- Annual estimate
- $1,080.00
VisionReasoningToolsCachingAudioCode
#03OpenAI
GPT-5.4
OpenAI's flagship multimodal model with deep reasoning and 1M context.
Monthly cost
$112.50
32.1× the cheapest
- Per call
- $0.0113
- Per 1,000 calls
- $11.25
- Annual estimate
- $1,350.00
VisionReasoningToolsCachingCode
#04Anthropic
Claude Sonnet 4.6
Balanced flagship model. Strong general-purpose performance at moderate cost.
Monthly cost
$120.00
34.3× the cheapest
- Per call
- $0.0120
- Per 1,000 calls
- $12.00
- Annual estimate
- $1,440.00
VisionReasoningToolsCachingCode
"DeepSeek V3.2 is 34.3× cheaper than Claude Sonnet 4.6 for this workload."
04 — Workload Scenarios
What does your actual workload cost?
Pre-built scenarios with realistic token volumes for common use cases. Hover over a model to see per-call breakdown.
Monthly cost — sorted cheapest first
50,000 calls × (1,200 in / 350 out)
01DeepSeek V3.2
$9.268/mo
02Gemini 2.5 Flash-Lite
1.1× more$9.760/mo
03GPT-5 mini
4.5× more$41.90/mo
04Claude Haiku 4.5
12.4× more$115.10/mo
05Gemini 3 Pro
28.6× more$265.20/mo
06GPT-5.4
35.8× more$331.50/mo
07Claude Sonnet 4.6
37.3× more$345.30/mo
08Claude Opus 4.7
62.1× more$575.50/mo
"DeepSeek V3.2 is 62× cheaper than Claude Opus 4.7 for this workload."
Saving roughly $566.23 per month — though capability and quality differ. Test with your actual prompts before committing.
AdvertisementSponsored
05 — On the HorizonPricing pending
Models we're
tracking next.
The calculator only lists models whose API pricing has been officially published. Below are releases we're watching.
The Token Meter does not publish unverified prices. When a provider discloses official numbers, the entry moves out of the Watchlist and into the live calculator — usually within 24 to 48 hours.
DeepSeek
DeepSeek V4
Expected to continue DeepSeek's aggressive price-to-capability trajectory. We will add it to the calculator the day official pricing is published.
OpenAI
GPT-5.3 Codex
Code-specialized variant in OpenAI's GPT-5 family. We will add it to the calculator the day it appears on the official pricing page.
Anthropic
Claude 4.7
Next iteration of Anthropic's Claude 4 family. We will add it to the calculator the day pricing is disclosed in the Anthropic console or pricing page.
Last reviewed: April 2026. If you've seen any of these models go live with official pricing, please let us know — speed is part of our value.
06 — Field Notes
Practical guides for the cost-conscious builder.
Short, opinionated essays on getting more from less. Updated as the model landscape shifts.
Field Note 01
Cost Optimization
6 min read
Five proven ways to reduce your LLM API bill
Prompt caching, smart model routing, output token limits, batch APIs, and prompt compression. Here's what each saves and which order to apply them.
1. Use prompt caching
Caching reusable prefixes (system prompts, few-shot examples, document context) cuts cached input cost to 10% of standard input on most providers. For chatbots with long system prompts, this alone often saves 40-60% of total spend.
2. Route by task complexity
Send classification, simple Q&A, and routing tasks to a small model (GPT-5 nano, Gemini Flash-Lite, Haiku 4.5). Reserve flagship models for tasks that genuinely need reasoning. A two-tier router can cut costs 70-90% with minimal quality impact.
3. Cap output token length
Output tokens are 4-6× more expensive than input on most models. Set max_tokens explicitly. Ask for terse responses. Use structured outputs with schemas to prevent the model from rambling.
4. Use batch APIs for non-urgent work
OpenAI, Anthropic, and Google all offer ~50% discounts on batch endpoints with 24-hour SLAs. Backfills, data labeling, evals, and overnight jobs should never run on real-time pricing.
5. Compress prompts where it doesn't hurt
Strip unnecessary whitespace from JSON. Summarize long context before sending. For RAG, retrieve top-3 chunks instead of top-10. Each removed token compounds across millions of calls.
Field Note 02
Model Selection
4 min read
When does a smaller model actually beat a flagship?
Counterintuitive data on Haiku 4.5, GPT-5 mini, and Gemini Flash. The answer depends more on your eval set than the leaderboard.
Latency-sensitive applications
Real-time chat, voice agents, autocomplete — small models often respond in 200-400ms vs. 2-4s for flagships. Users notice latency more than they notice marginal quality differences.
Well-defined, narrow tasks
Classification, extraction, formatting, summarization with a clear template — these tasks don't benefit from flagship reasoning. Haiku 4.5 and GPT-5 mini frequently match Opus/Sonnet quality at 5× the speed and 1/5 the cost.
High-volume, low-stakes work
Tagging, moderation pre-filters, sentiment analysis at scale. Even a 2% quality drop is acceptable when you're saving $10K/month. Use the saved budget to validate edge cases manually.
When to NOT downgrade
Multi-step reasoning, novel problems, code generation requiring deep context understanding, anything user-facing where quality drift would damage trust. Test with your real evals — leaderboards lie.
Field Note 03
Reasoning Models
5 min read
The hidden cost of reasoning tokens
GPT-5.4, o-series, and DeepSeek R1.5 all bill 'thinking' tokens separately. Here's how to estimate them and when reasoning is worth the premium.
What reasoning tokens are
Reasoning models generate internal chain-of-thought before producing the visible output. Those internal tokens are billed at output rates (often $15-25/1M) but are invisible in the final response, making them easy to under-budget.
Typical multipliers
Light reasoning task: 2-3× the visible output in thinking tokens. Complex math, coding, or multi-step planning: 5-15× the visible output. A response that 'looks like' 500 tokens might actually cost as if it were 5,000.
When reasoning is worth it
Math problems, code generation with subtle logic, multi-hop questions, anything where 'thinking longer' meaningfully improves accuracy. For straightforward extraction or generation, disable reasoning or use a non-reasoning variant.
Set thinking budgets
Most reasoning models now support a max_reasoning_tokens or thinking_budget parameter. Set it. A model thinking 'as long as it needs' is a great way to receive a $400 bill from a single complex query.
07 — Colophon
About this hub
An independent reference for the AI economy.
The Token Meter is a free utility published by DrewIs Intelligence LLC. It exists for one reason: choosing the right AI model has become genuinely confusing, and most existing tools either focus on a single provider or bury the math under marketing copy.
Pricing data is verified against official provider pricing pages and updated regularly. Token estimates use a content-aware heuristic that approximates BPE tokenizers within a few percent — close enough for budgeting, not exact enough for billing reconciliation.
We do not run any AI inference on your text. Everything you type stays in your browser. No accounts, no tracking pixels on your prompts, no telemetry on your calculator inputs.
20+
Models tracked
9
Providers covered
100%
Client-side privacy
$0
Free, no signup
"Built so you can spend more of your budget on actually using AI, and less of it on figuring out what AI to use."