✅ Verified against primary sources · Updated continuously · No ads
Token pricing verified against official OpenAI, Anthropic, Google Gemini and other provider pages.
A practical hub for cost decisions by engineers and product teams.
Engineer-focused LLM cost data, verified against official primary sources
OpenAI (GPT-5.6 Sol/Terra/Luna, GPT-5.5/5.4), Anthropic Claude (Opus 5/4.8, Fable 5, Sonnet 5, Haiku 4.5), Google Gemini (3.7/3.6/3.5 Flash, 3.1 Flash-Lite), DeepSeek V4.1 Flash and Z.ai GLM — in USD per 1M tokens, verified against each provider's official page.
See pricing comparison →Enter input and output token counts to instantly compare costs across major LLM APIs in USD (plus JPY reference). No network calls — everything runs in your browser.
Open the calculator →From DeepSeek R1 (Jan 2025) to GPT-5.6, Claude Opus 5, Gemini 3.7 Flash and GLM 5.3 Flash — major model releases verified against primary sources and organized chronologically, each with an official link.
See the timeline →"What exactly is cheaper about a cache hit?" "What is this benchmark number measuring?" — the terms that show up in pricing tables and announcements, one per page: the gap between cache hits, cache misses and cache writes, the conditions behind a benchmark score, cache TTLs and context-window boundaries. Everything checked against the vendors' own documentation. Entry 1: how to read GPT-6 Astra's benchmarks. Entry 2: prompt caching. Entry 3: what an agent harness is. Entry 4: the 1M context window. Entry 5: benchmarks. Entry 6: benchmark operators. Entry 7: zero data retention (ZDR). Entry 8: the four data-protection requirements (not stored, not seen, not learned from, not sent abroad). #9 covers zero operator access. Entry 10: training restrictions (no-training). Entry 11: data residency — storage and processing are two promises.
Read the glossary →Released 2026-09-22. Sol is $2/$10 and Luna $0.10/$0.50 — 50% below the GPT-5.6 versions. 1,050,000 context, 128K output, 90% off cached input. Six official benchmarks, each with the operator named. GPT-5.6 Sol and Luna are now legacy.
Sol / Luna deep dive →Released 2026-09-22. $4 input / $20 output / $0.20 cache read (down from Opus 5's $5/$25). 1M context, 128K output. All nine official benchmarks with the harness owner noted, the basis for the 40%-cheaper claim, and the change that thinking can no longer be turned off.
Opus 5.5 deep dive →Released 2026-09-28 at $2 input / $10 output, the same as the previous-generation Sonnet 5, and $0.10 cache reads — half of Sonnet 5's $0.20 (the docs call it "0.05x on Claude Opus 5.5 and Claude Sonnet 5.5"). 1M context, 128K max output, adaptive thinking on by default with a default effort of high and a between_tools setting that turns up-front thinking off. No official benchmarks are published, so we publish none: only verified specs and the model ID on each of the five platforms.
Released 2026-10-07. At $0.10 input / $0.50 output it is one tenth of Haiku 4.5 up to 100,000 tokens; above that a second tier applies ($0.50 / $2.50). Cache reads $0.01, 1M context, and the new tokenizer that counts the same text as roughly 30% more tokens.
Haiku 5.5 deep dive →$2 input / $10 output matches GPT-6 Sol, while cached input halves to $0.10. Prompts above 272,000 input tokens are repriced at 2x input and 1.5x output for the whole request; Fast is 2x, Ultrafast 6x. 1.05M context, reasoning effort low to max (no none).
A new frontier for coding, knowledge work and long-horizon agents. $10/$50 with $0.25 cache reads — up to ~45% cheaper in practice. 1M context, 128K output. Terminal-Bench-Science more than doubles Fable 5 (52.6%). Sister model Mythos 5.1 is trusted-access only.
Fable 5.1 details →A 6B-active MoE that beats Qwen3.7-Plus, previewing the Qwen4 architecture. Input $0.15 / output $0.47, 1M context, 64K output, image/video. Specs, benchmarks and comparisons from primary sources.
Qwen3.8-Flash details →Z.ai (Zhipu)'s first natively multimodal GLM-5 model. 320B MoE (18B active) beating GLM-5.2 at ~1/10 the price. Input $0.15 / output $0.50. Specs, benchmarks and comparisons.
GLM 5.3 Flash details →Dropped to $0.20/$1.20. Cheaper than Gemini Flash-Lite with GPT-5.5-beating performance. Competitor comparisons, use cases and a quick selection guide.
Luna details →Cut to $4/$20 (at least through 2026-11-21). 1.05M context, 128K output, top-tier reasoning. Comparison vs Claude Opus 5 and the Cloudflare AI Gateway 50% off.
Sol details →$0.25/$1.50, multimodal. Built on the Gemini 3 Pro architecture with 2.5 Flash-level quality. Deep comparison vs GPT-5.6 Luna and thinking-level control.
Flash-Lite details →Released 2026-09-21. Beats Grok 4.6 on all seven official benchmarks. 500K context, $4/$12 above 200K, Fast at 2x. Includes the third-party gap.
Grok 4.7 details →$2/$6 standard, fast variant 2×, 500K context. AA Intelligence Index 61, tying GPT-5.6 Sol. Pricing, performance and comparisons from primary sources.
Grok 4.6 details →$0.75/$3.75 launch (thru 12/31) → standard $1.50/$7.50. 1M context, 64K output. FrontierCode 43.6% · DeepSWE 65.3%.
3.7 Flash details →DeepSeek V4.1 Flash (released 2026-09-10): pricing from $0.15/$0.60, 1M context, native image input, VS Code GitHub Copilot Chat setup, the V4 Pro continuation (the 9/14 routing notice was withdrawn) and an Azure AI Foundry deployment walkthrough.
Practical guide →Image, audio, video, and PDF/file in/out support and pricing in one table. Gemini 3.8 Flash (image/audio/video/PDF input), GPT-6 Astra (image input only, $10/$50), Claude Fable 5.1, Qwen3.8-Flash, GLM 5.3 Flash, Grok 4.6, DeepSeek V4.1 Flash (image input supported). Also covers models that bill 2x for audio, and how live streaming, image & speech generation split into dedicated models.
View the Multimodal guide →Microsoft's in-house family served on Azure AI Foundry (Microsoft Foundry). MAI-Image-2.6/2.6-Flash (image generation), MAI-Transcribe-2 (speech recognition, $0.10/hr limited), MAI-Thinking-1 (reasoning) and more. Pricing, usage and a release timeline, verified against primary sources.
See the MAI series →DeepSeek, Qwen (Alibaba), GLM (Zhipu), Kimi (Moonshot) and MiniMax at a glance. Indicative pricing, open-weight status and an update log. Only what can be verified against primary sources.
China AI overview →Data quality policy